Seed42Lab/AI-paper-crawl · Datasets at AI快站

index int64 0 18.8k	text stringlengths 0 826k	year stringclasses 38 values	No stringlengths 1 4
0	QUESTION ORDERING IN MIXED INITIATIVE PROGRAM SPECIFICATION DIALOGUE Louis Steinberg Department of Computer Science Rutgers University New Brunswick, N. J. 08903 ABSTRACT It would be nice if a computer system could accept a program specification in the form of a mixed initiative dialogue. One capability such a system must have is the ability to ask questions in a coherent order. We will see a number of reasons it is better if such a system produces all the questions it can and has a "dialogue moderator" choose which to ask next, than if the system asks the first question it thinks of. DM [9I, the dialogue moderator of PSI C51, chooses questions by searching a network model of a program, under control of a set of heuristic rules. This technique is simple and flexible. I. Introduction When you need a computer program, it is usually easier to tell a human being what the program should do than to specify the program directly to a computer system (eg a compiler). There are a number of reasons for this, including the knowledge and reasoning ability that a human has. We will concentrate here, however, on another advantage of communicating with humans, their ability to engage in a mixed initiative dialogue, and on one particular capability required for carrying on such a dialogue,*the ability to ask questions in a coherent order. A mixed initiative dialogue is one in which either party may take initiative. From the perspective of the work reported here, to "take initiative" in a dialogue is to alter the structure of the dialogue. This definition is essentially equivalent to that of Bobrow, et al [II, who define taking initiative as establishing or --m---w- ' This work was supported by the Defense Advanced Research Projects Agency at the Department of Defense under contract MDA 903-76-C-0206. The author was also partially supported by an IBM Graduate Fellowship. The work reported here would not have been possible without the environment provided by Cordell Green and the other members of the PSI project. I would also like to thank N. S. Sridharan for his comments on an earlier draft of this paper. violating expectations about what will come next, since it is precisely the structure of a dialogue which gives rise to such expectations. In particular, we will be concerned here with "topic structure", the order and relationships of the topics covered in the dialogue, and with "topic initiative", the ability to affect topic structure. The work described here 191 been done in the context of the PSI program synthesis system [5]. PSI acquires program specifications via mixed initiative, natural language dialogue. II. The General Scheme In order to ask questions, such a system must be able to do two things: it has to decide what aspects of the specification are as yet incomplete, and it has to decide which one of these aspects to ask about next. We will refer to the latter problem, deciding which question to ask next, as the task of "question ordering". A. Order from the Reasoning Process One common way to handle question ordering might be summarized as asking the first question the system thinks of. In this scheme, the system goes through its normal reasoning process, and at some point comes across a fact which it wants to know, but cannot deduce. Whenever this happens, the system stops and asks the user. (See, for example, [11 and 143). Note that the system stops whenever it finds any question to ask. Thus, the system asks each question as it comes up, and the order is determined by the reasoning process. If a system's reasoning process seems natural to the user, then this scheme produces a question order which seems natural, at least to a first approximation. However, there are some problems. The basic problem is that this scheme ties the topic structure of the dialogue to the reasoning procedures of the system. This makes topic structure harder to change, since any change in topic structure requires a change in the reasoning procedure. It can also make it hard to transfer the question ordering methods to another system that uses a different reasoning method. Finally, this method of question ordering assumes that there is a single, sequential reasoning process, and is 61 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. not possible in a system structure such as that of HEARSAY-II [71. B. Order from a Dialog Moderator A better scheme is to have the reasoning process produce as many questions as it can, and to use some other mechanism to select a single one of them to ask next. This scheme largely avoids the problems of the previous one. Its main drawback is that it requires a reasoning process which is able to produce more than one question at a time. An additional advantage of this scheme is that it allows us to implement question ordering in a separate module, with a clearly defined interface to the rest of the system. I have termed such a module a "dialogue moderatortt. Thus, the dialogue moderator is given a list of all the questions currently open, and must choose which one is to be asked next, so as to keep the dialogue well structured. Much recent research (eg [2], C61, [8]> has shown that structure of a dialogue is closely tied to the structure of goals and plans being pursued by the dialogue's participants. One might therefore imagine that the dialogue moderator needs a complete model of goals and plans, both those of the system and those of the user. However, in a program specification dialogue, the goals and plans of both participants are tied very closely to the structure of the program. As will be seen, it has been possible in PSI to use a simple model of the program structure instead of a complex model of goals and plans. (It might be argued that any system which handles natural language will eventually need the full model of goals and plans anyway, so using a simpler model here is no savings in the long run. It should be noted, however, that mixed initiative does not necessarily imply natural language. A useful system might be constructed which handles mixed initiative dialogue in some formal language.) III. J& Specific Method DM is the dialogue moderator of the PSI system. As noted above, DM maintains a simplified model of the program being specified. The program is viewed as a structured set of objects. Each object is either a piece of algorithm or a piece of data structure - the pieces of algorithm correspond roughly to the executable statements of a program, and the pieces of data structure correspond roughly to the variable declarations. A specific loop or a specific input operation might be algorithmic objects, while a set or a 5-tuple might be data structure objects. These objects are structured by two relationships: an object may be a subpart of another (eg an input operation might be a step of a loop, and thus one of its subparts), and an algorithm object may use a data structure object (eg an input operation "usestt the data structure it inputs). DM represents this structure in a standard network form; nodes represent the objects, and arcs represent the relations subpart/superpart and uses/used-by. Each node also has associated with it a list of questions about the object it represents. (A question asks about some attribute of some specific object. The objects, relations, and questions come from other modules of PSI.) In order to choose the next question to ask, DM searches the net, starting at the "present topic". The present topic is the object currently being discussed. Determining which object this is is a difficult and important problem in its own right, involving the syntax of the user's sentences as well as the status of the program specification, and has not been seriously dealt with in this work. Instead, some simple heuristics are used, the main one being to assume that most of the time the user will be talking about the object that the system just asked about. Once the present topic has been chosen, the search proceeds, under control of a set of rules. (The rules are listed in the appendix. See [9] for a discussion of the specific rules.) Each time the search reaches an object, a list of rules is chosen (depending on whether the object is a piece of algorithm or data structure) and these rules are applied in order. Some say to look for a specific kind of question about the current object. Others say to move along some particular kind of arc from the current object, and recursively apply the rules on the object we reach. If no question is found by this recursive application, we come back and continue applying the rules here. If at any point a rule that looks for questions finds one, that question is the one to ask, and the search stops. This scheme of moving through the net and looking for questions, under control of a set of rules, has proven to be simple and flexible. A related technique was used in SCHOLAR [3]. SCHOLAR is a CA1 system which teaches geography by engaging in a mixed initiative dialogue with the student. Both participants may ask and answer questions. SCHOLAR chooses which question to ask by a random (rather than rule directed) walk on a net which encodes its knowledge about geography. As ultimately envisioned, SCHOLAR would teach in a Socratic manner, that is, by asking a carefully designed sequence of questions. However, the structure of goals and plans in such a dialogue is probably very different from the structure of the net as discussed in [31. Because of this, a scheme of moving through this net is unlikely to be useful for producing such a sequence of questions. DM's question ordering behavior has been tested in two ways. First, a log of runs of PSI was surveyed. This log included 42 dialogues which were essentially complete. Each dialogue was checked, both to see if the user complained about the question ordering (there is a comment feature that can be used for such complaints), and also to see if the question order was subjectively acceptable. Except for one instance, later traced to a program bug, DM's behavior was correct. This test was too subjective, however, so a simulated dialogue was recorded, with myself playing the role of PSI and a programmer from outside the PSI group as the user. The inputs DM would have gotten during this dialogue were hand coded and given to DM, and the questions DM chose were compared with those I had chosen. DM had to choose a question at sixteen points, with two to seven questions to choose from. The correct question was chosen at thirteen of these points. An analysis of the errors indicates that they could be removed by some straightforward extensions of the current methodology, particularly by maintaining more history of how the dialogue got to the present topic. IV. Conclusions Thus we see that it is advantageous for a system which engages in mixed initiative dialogue to have the reasoning modules produce all the questions they can at each point in the dialogue, and to have a separate dialogue moderator choose which one to ask next. In such a system, the question ordering mechanism is decoupled from the reasoning process, so that either can be modified without changing the other. A given mechanism for selecting one of the proposed questions can be more easily transferred to a system with very different reasoning mechanism. Also, multiple parallel reasoning processes can be used with this scheme. DM, the dialogue moderator of PSI, represents the program as a simple net of objects and relations. It chooses a question by starting at the node representing the present topic of the dialogue, and searching the net, under control of a set of rules. It is possible to use a simple model of the program, rather than a complex model of goals and plans, because in the program specification task, the participants' goals and plans are so closely tied to the program structure. This general scheme of rule based search is advantageous because it is simple and flexible. These techniques are probably applicable to other settings where the structure of goals and plans can be tied to some simple task related structure. APPENDIX: Question Choice Rules - _I_- (These are slightly simplified versions of the content of the rules. The actual rules consist of LISP code.) Rules for Algorithms Al) Are there questions about the NAME of this object? A2) Look at all objects that are USED-BY this object. A3) Are there questions about this object other than EXIT-TEST, PROMPT, or FORMAT? A4) Are there questions about the PROMPT or FORMAT for this object? A5) Look at all objects that are SUB-PARTS of this object. A61 Are there questions about the EXIT-TEST of this object? A7) Look at all objects that are SUPER-PARTS of this object. Rules for Data Structures Dl) Look at all objects that are SUB-PARTS of this object. D2) Are there questions about the STRUCTURE of this object? D3) Are there OTHER questions about this object? D4) Look at all objects that are SUPER-PARTS of this object. D5) Look at all objects that USE this object. REFERENCES Cl] Bobrow, D., Kaplan, R., Kay, M., Norman, D., Thompson, H., Winograd, T., "GUS, A Frame- Driven Dialogue System." Artificial Intelligence 8 (1977) 155-173. [2] Brown, G., "A Framework for Processing Dialogue", Technical Report 182, MIT Laboratory for Computer Science, June '977. [3] Carbonell, J. R., "AI in CAI: An Artificial Intelligence Approach t.0 Computer-Aided Instruction." IEEE Trans. Man-Machine Svst. J-l- (1970) 190-202. [4] Davis, R., Buchanan, B., Shortliffe, E., Production Rules as a Representation for a Knowledge-Based Consultation Program, Memo AIM- 266, Stanford Artificial Intelligence Laboratory, October, 1975. [5] Green, C., A Summary of the PSI Program Synthesis System, Proceedings of &h& Fifth International Joint Artificial IntelliKZ, Conference on Cambridge Massachusetts, August 1977, 380 - 381. [63 Grosz, B., The Representation and Use of Focus in a System for Understanding Dialogues, Proceedings of &t& Fifth d International Joint Conference on Artificial Intelligence, Cambridge, Massachusetts, August 1977, 67 - 76. [7] Lesser, V., Erman, L., A Retrospective View of the HEARSAY-II Architecture, Proceedin= of the Fifth International Joint Conference on Artificial Intelligence, Cambridge, Massachusetts, August 1977, 380 - 381. 183 Mann, W., Man-Machine Communication Research: Final ReDort, ISI/RR-77-57, USC Information Sciences Institute, February, 1977. [9] Steinberg, L., A Dialogue Moderator for Program Specification Dialogues in t,he psi Svstem, PhD dissertation, Stanford University, in progress. 63	1980	1
1	AUTOMATIC GENERATION OF SEMANTIC ATTACHMENTS IN FOL Luigia Aiello Computer Science Department Stanford University Stanford, California 94305 ABSTRACT Semantic attachment is provided by FOL as a means for associating model values (i.e. LISP code) to symbols of a first order language. This paper presents an algorithm that automatically generates semantic attachments in FOL and discusses the ad- vantages deriving from its use. I INTRODUCTION In FOL (the mechanized reasoning system developed by R. Weyhrauch at the Stanford A.I. Laboratory [4,5,6 1, the knowledge about a given domain of discourse is represented in the form of an L/S structure. F An L/S structure is the FOL counterpart of the logician notion of a theory/model pair. It is a triple <L,S,F> where L is a sorted first order language with equality, S is a simulation structure (i.e. a computable part of a model for a first order theory), and F is a finite set of facts (i.e. axioms and theorems). Semantic attachment is one of the characteriz- ing features of FOL. It allows for the construction of a simulation structure S by attaching a "model value" (i.e. a LISP data structure) to (some of) the constant, function and predicate symbols of a first order language. Note that the intended semantics of a given theory can be specified only partially, i.e. not necessarily all the symbols of the language need to be given an attachment. The FOL evaluator, when evaluating a term (or wff), uses both the semantic and the syntactic information provided within an L/S structure. It uses the semantic attachments by directly invoking the LISP evaluator for computing the value of ground sub-terms of the term (wff). It uses a simplification set, i.e. a user-defined set of rewrite rules to do symbolic evaluations on the term (wff). Semantic information and syntactic informa- tion are repeatedly used - in this order - until no further simplification is possible. The research reported here has been carried out while the author was visiting with the Computer Science Department of Stanford University on leave from IEI of CNR, Pisa, Italy. Author's permanent address: IEI-CNR, via S. Maria 46, I-56100 Pisa, Italy Semantic attachment has been vital in the generation of many FOL proofs, by significantly increasing the efficiency of evaluations. The idea of speeding up a theorem prover by directly invoking the evaluator of the underlying system to compute some functions (predicates) has been used in other proof generating systems. FOL is different from other systems in that it provides the user with the capability of explicitly telling FOL which semantic information he wants to state and use about a given theory. This approach has many advantages, mostly epistemological, that are too long to be discussed here. II AUTOMATIC GENERATION OF SEMANTIC ATTACHMENTS It is common experience among the FOL users that they tend to build L/S structures providing much more syntactic information (by specifying axioms and deriving theorems) than semantic inform- tion (by attaching LISP code to symbols). In recent applications of FOL, L/S structures are big, and (since the information is essentially syntactic)the dimension of the simplification sets is rather large. The unpleasant consequence is that the evaluations tend to be very slow, if feasible at all. This has prompted us to devise and implement an extension of the FOL system, namely, a compiling algorithm from FOL into LISP, which allows for a direct evaluation in LISP of functions and pre- dicates defined in First Order Logic. The compila- tion of systems of function (predicate) definitions from FOL into LISP allows FOL to transform syntac- tic information into semantic information. In other words, the compiling algorithm allows FOL to automatically build parts of a model for a theory, starting from a syntactic description. Semantic attachment has often been criticised as error prone. In fact, the possibility of directly attaching LISP code to symbols of the language allows the FOL user to set up the se- mantic part of an L/S structure in a language different from that of first order logic. This forbids him to use FOL itself to check the relative consistency of the syntactic and semantic part of an L/S structure. From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. The automatic transformation of FOL axioms (or, in general, facts) into semantic attachments, besides the above mentioned advantage of substanti- ally increasing the efficiency of the evaluator, has the advantage of guaranteeing the consistency between the syntactic and semantic specifications of an FOL domain of discourse, or at least, to keep to a minimum the user's freedom of introduc- ing non-detectable inconsistencies. The semantic attachment for a function (predi- cate)symbol can be automatically generated through a compilation if such a symbol appears in the syntactic part of an L/S structure as a definiendum in a system of (possibly mutually recursive) definitions of the following form: Yx 1 . ..x r.fixl,... X r> =6 i@i,Iyi,q⌧19 . l . +I V Yl l - Ys 4Pj (Yl . Y,) = Zj@jr~j,~j,Yl,...,Ysl) Here the f-s are function symbols and the P-s are predicate symbols. The mare terms in4-s, Y-s, F-s and x-s; the t-s are wffs in the$-s, q-s, ‘E-s and y-s. By ? we denote a tuple of constant sym- bols. By 9 (resp.2) we denote a tuple of function (resp. predicate) symbols. @(resp.2) may contain some of f (resp. P), but it is not necessarily limited to them, i.e. other function and predicate symbols besides the definienda can appear in each definiens. The compilation algorithm, when provided with a system of definitions, 'first performs a well- formedness check, then a compilability check. The well-formedness check tests whether or not all the facts to be compiled are definitions, i.e. if they have one of the two following forms (note that here we use the word "definition" in a broader sense than logicians do): fjxl...xr.fi(xl,...,xr) = . . . if Yl l *Ys.(Pj(yl'...,ys)= . . . The compilability check- consists in verifying that a> each definition is a closed wff, i.e. no free variable occurs in it; b) all the individual constants and the function (predicate) symbols appearing in the definiens either are one of the definienda or are attached to a model value (the first case allows for recursion or mutual re- cursion); c) the definiens can contain logical constants, conditionals and logical connectives but no quantifiers. When the FOL evaluator invokes the LISP evaluator, it expects a model value to be returned; it does not know how to handle errors occurring at the LISP level. This, for various reasons too long to be reported here, justifies the three above restrictions. Actually, the second and the third restrictions can be weakened with and appropriate extension of the FOL evaluator and of the compiler (respectively) to cope with the new situation. More details are presented in [l]. To present a simple example of compilation, consider the following facts: yy x.f(x,y) = if P(x) then g(x,x) else f(y,x) - vy x.g(x,y) = x+y If we tell FOL to compile them in an L/S structure where a semantic attachment exists both for the symbol P and for the symbol + (let them be two LISP functions named C-P and PLUS, respectively), it produces the following LISP code: (DE c-f (X y> (COND ((C-P x) (C-g x x)) (T (C-f y xl>>> (DE C-g (x y) (PLUS x Y> ) and attaches it to the function symbols f and g, respectively. III SOUNDNESS OF THE COMPILATION The compiling algorithm is pretty straight- forward, hence, its correctness should not con- stitute a problem. Conversely, a legitimate question is the following: Is the compilation process sound? In other words: Who guarantees that running the FOL evaluator syntactically on a system of definitions gives the same result as running the LISP evaluator on their (compiled) semantic attachments? The answer is that the two evaluations are weakly equivalent, i.e. if both terminate, they produce the same result. This is because the FOL evaluator uses a leftmost outermost strategy of function invocation (which corresponds to call-by- name) while the mechanism used by the LISP evalua- tor is call-by-value. Hence, compiling a function can introduce some nonterminating computations that would not happen if the same function were eval- uated symbolically. This, however, does not constitute a serious problem and it will be overcome in the next version of FOL. In fact, it will be implemented in a purely applicative, call-by-need dialect of LISP (note that, call-by-need is strongly equivalent to call-by-name in purely applicative languages). IV CONCLUSION FOL is an experimental system and, as is often the case with such systems, it evolves through the experience of its designer and users. Particular attention is paid to extend FOL only with new features that either improve its proving power or allow for a more natural interaction between the user and the system (or both) in a uniform way. The addition of the compiling algorithm sketched in the previous sections is in this spirit. This extension of FOL has been very useful in recent applications (see, for instance [2]). Experience has shown that the largest part of the syntactic information in an L/S structure can be compiled. This suggests a further improvement to be done on FOL evaluations. The use of the compiling algorithm leads to L/S structures where (almost) all the function (predicate) symbols of the language have an attachment. Hence, the strategy of the FOL evaluator to use semantic information first (that was the most reasonable one when semantic attachments were very few and symbolic evaluations could be rather long) is in our opinion no longer the best one. In fact, sometimes, properties of functions (stated as axioms or theorems in the syntactic part of the L/S structure) can be used to avoid long computations before invoking the LISP evaluator to compute that function. Finally, a comment on related work. Recently (and independently), Boyer and Moore have added to their theorem prover the possibility of introducing meta-functions, proving them correct and using them to enhance the proving power of their system [3]. This is very much in the spirit of the use of META in FOL and of the compiling algorithm described here. ACKNOWLEDGMENTS The members of the Formal Reasoning Group of the Stanford A.I. Lab are acknowledged for useful discussions. Richard Weyhrauch deserves special thanks for interesting and stimulating conversa- tions about FOL. The financial support of both the Italian National Research Council and ARPA (through Grant No. MDA903-80-C-0102) are acknowledged. REFERENCES [1] Aiello, L., "Evaluating Functions Defined in First Order Logic." Proc of the Logic - - - Programming Workshop, Debrecen, Hungary, 1980. [2] Aiello, L., and Weyhrauch, R. W., "Using Meta- theoretic Reasoning to do Algebra." Proc. of -- the 5th Automated Deduction Conf., Les Arcs, France, 1980. [3] Boyer, R.S., and Moore, J.S., "Metafunctions: Proving them correct and using them efficiently as new proof procedures." C. S. Lab, SRI International, Menlo Park, California, 1979. [4] Weyhrauch, R.W., "FOL: A Proof Checker for First-order Logic." Stanford A.I. Lab, Memo AIM-235.1, 1977. [5] Weyhrauch, R. W., "The Uses of Logic in Artificial Intelligence." Lecture Notes of the Summer School on the Foundations of Artificial Intelligence and Computer Science (FAICS '78), Pisa, Italy, 1978. [6] Weyhrauch, R.W., "Prolegomena to a Mechanized Theory of Formal Reasoning." Stanford A.I. Lab, Memo AIM-315, 1979; Artificial Intelligence Journal, to appear, 1980. 92	1980	10
2	ABSTRACT HCPRVR: AN INTERPRETER FOR LOGIC PROGRAMS Daniel Chester Department of Computer Sciences University of Texas at Austin An overview of a logic program interpreter written in Lisp is presented. The interpreter is a Horn clause-based theorem prover augmented by Lisp functions attached to some predicate names. Its application to natural language processing is discussed. The theory of operation is explained, including the high level organization of the PROVE function and an efficient version of unification. The paper concludes with comments on the overall efficiency of the interpreter. An axiom is either an atomic formula, which can be referred to as a fact, or an expression of the form ( <conclusion> < <premissl> .. <premissN> ) where both the conclusion and the premisses are atomic formulas. The symbol "<" is intended to be a left-pointing arrow. An atomic formula is an arbitrary Lisp ___ - expression beginning with a Lisp atom. That atom is referred to as a relation or predicate name. Some of the other atoms in the expression may be designated as variables by a flag on their property lists. I INTRODUCTION III CALLING LOGIC PROGRAMS - -- HCPRVR, a Horn Clause theorem PRoVeR, is a Lisp program that -interprets a --- simple logical formalism as a programming language. It has been used for over a year now at the University of Texas at Austin to write natural language processing systems. Like Kowalski [II, we find that programming in logic is an efficient way to write programs that are easy to comprehend. Although we now have an interpreter/compiler for the logic programming language Prolog [2], we continue to use HCPRVR because it allows us to remain in a Lisp environment where there is greater flexibility and a more familiar notation. This paper outlines how HCPRVR works to provide logic programming in a Lisp environment. The syntax of logic programs is given, followed by a description of how such programs are invoked. Then attachment of Lisp functions to predicates is explained. Our approach to processing natural language in logic programs is outlined briefly. The operation of HCPRVR is presented by giving details of the PROVE and MATCH functions. The paper closes with some remarks on efficiency. II LOGIC PROGRAM SYNTAX ~___-- ___ A logic program is an ordered list of axioms. -- -e--e This work was supported by NSF Grant MCS 74-2491 -8. There are two ways to call a logic program in HCPRVR. One way is to apply the EXPR function TRY to an atomic formula. The other way is to apply the FEXPR function ? to a list of one or more atomic formulas, i.e., by evaluating an expression of the form ( ? <formulal> ..- <formulaN> > In either case the PROVE function is called to try to find values for the variables in the formulas that makes them into theorems implied by the axioms. If it finds a set of values, it displays the formulas to the interactive user and asks him whether another set of values should be sought. When told not to seek further, it terminates after assigning the formulas, with the variables replaced by their values, to the Lisp atom VAL. IV PREDICATE NAMES AS FUNCTIONS ~- Occasionally it is useful to let a predicate name be a Lisp function that gets called instead of letting HCPRVR prove the formula in the usual way. The predicate name NEQ, for example, tests its two arguments for inequality by means of a Lisp function because it would be impractical to have axioms of the form (NEQ X Y) for every pair of 93 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. constants X and Y such that X does not equal Y. Predicate names that are also functions are FEXPRs and expect that their arguments have been expanded into lists in which all bound variables have been replaced by their values. These predicate names must be marked as functions by having the Lisp property FN set to T, e.g., executing (PUT '<predicate name> 'FN T), so thi: HCPRVR will interpret them as functions. By letting syntactic categories be predicates with three arguments, we can make axioms that pull phrases off of a list of words until we get a sentence that consumes the whole list. In addition, arbitrary tests can be performed on the phrase representations to check whether they can be semantically combined. Usually the phrase representation in the conclusion part of an axiom tells how the component representations are combined, while the premisses tell how the phrase should be factored into the component phrases, what their' representations should be, and what restrictions they have. Thus, the axiom ((S X (U ACTOR V . W) Z) < (NP X V Y) (VP Y (U . W) Z) (NUMBER v ~1) (NUMBER U N2) (EQ N1 N2)) says that an initial segment of word list X is a sentence if first there is a noun phrase ending where word list Y begins, followed by a verb phrase ending where word list Z begins, and both phrases agree in number (singular or plural). Furthermore, the noun phrase representation V is made the actor of the verb U in the verb phrase, and the rest of the verb phrase representation, W, is carried along in the representation for the sentence. After suitable axioms have been stored, the sentence THE CAT IS ON THE MAT can be parsed by typing (? (S (THE CAT IS ON THE MAT) x NIL) The result of this computation is the theorem (S (THE CAT IS ON THE MAT) (IS ACTOR (CAT DET THE) LOC (ON LOC (MAT DET THE))) NIL) VI THEORY OF OPERATION -- A. General Organization HCPRVR works essentially by the problem reduction principle. Each atomic formula can be thought of as a problem. Those that appear as facts in the list of axioms represent problems that have been solved, while those that appear as conclusions can be reduced to the list of problems represented by the premisses. Starting from the formula to be proved, HCPRVR reduces each problem to lists of subproblems and then reduces each of the subproblems in turn until they have all been reduced to the previously solved problems, the "facts" on the axiom list. The key functions in HCPRVR that do all this are PROVE and MATCH. B. The PROVE Function -~ PROVE is the function that controls the problem reduction process. It has one argument, a stack of subproblem structures. Each subproblem structure has the following format: ( <list of subproblems>.<binding list> ) where the list of subproblems is a sublist of the premisses in some axiom and the CAR of the binding list is a list of variables occurring in the subproblems, paired with their assigned values. When PROVE is initially called by TRY, it begins with the stack ( ( ( <formula> ) NIL ) ) The algorithm of PROVE works in depth-first fashion, solving subproblems in the same left-to-right order as they occur in the axioms and applying the axioms as problem reduction rules in the same order as they are listed. PROVE begins by examining the first subproblem structure on its stack. If the list of subproblems in that structure is empty, PROVE either returns the binding list, if there are no other structures on the stack, i.e., if the original problem has been solved, or removes the first structure from the stack and examines the stack again. If the list of subproblems of the first subproblem structure is __ empty, not PROVE examines the first subproblem on the list. If the predicate name in it is a function, the function is applied to the arguments. If the function returns NIL, PROVE fails; otherwise the subproblem is removed from the list and PROVE begins all over again with the modified structure. 94 When the predicate name of the first subproblem in the list in the first subproblem stucture is not a function, PROVE gets all the axioms that are stored under that predicate name and assigns them to the local variable Y. At this point PROVE goes into a loop in which it tries to apply each axiom in turn until one is found that leads to a solution to the original problem. It does this by calling the MATCH function to compare the conclusion of an axiom with the first subproblem. If the match fails, it tries the next axiom. If the match succeeds, the first subproblem is removed from the first subproblem structure, then a new subproblem structure is put on the stack in front of that structure. This new subproblem structure consists of the list of premisses from the axiom and the binding list that was created at the time MATCH was called. Then PROVE calls itself with this newly formed stack. If this call returns a binding list, it is returned as the value of PROVE. If the call returns NIL, everything is restored to what it was before the axiom was applied and PROVE tries to apply the next axiom. The way that PROVE applies an axiom might be better understood by considering the following illustration. Suppose that the stack looks like this: ( ( (Cl C2).<blist> ) ..* ) The first subproblem in the first subproblem structure is Cl. Let the axiom to be applied be (C < Efl P2 P3) PROVE applies it by creating a new binding list blist', initially empty, and then matching C with Cl with the call (MATCH C <blist'> Cl <blist>). If this call is successful, the following stack is formed: ( ( (Pl P2 P3).<blist'> ) ( (C2).<blist> ) e.. ) Thus problem Cl has been reduced to problems Pl, P2 and P3 as modified by the binding list blist'. PROVE now applies PROVE to this stack in the hope that all the subproblems in it can be solved. In the event that the axiom to be applied is cc>, that is, the axiom is just a fact, the new stack that is formed is ( ( ().<blist'> ) ( (C2).<blist> ) . . . ) When PROVE is called with this stack, it removes the first subproblem stucture and begins working on problem C2. c. The MATCH Function -- The MATCH function is a version of the unification algorithm that has been modified so that renaming of variables and substitutions of variable values back into formulas are avoided. The key idea is that the identity of a variable is determined by both the variable name and the binding list on which its value will be stored. The value of a variable is also a pair: the term that will replace the variable and the binding list associated with the term. The binding list associated with the term is used to find the values of variables occurring in the term when needed. Notice that variables do not have to be renamed because MATCH is always called (initially) with two distinct binding lists, giving distinct identities to the variables in the two expressions to be matched, even if the same variable name occurs in both of them. MATCH assigns a value to a variable by CONSing it to the CAR of the variable's binding list using the RPLACA function; it also puts that binding list on the list bound to the Lisp variable SAVE in PROVE. This is done so-that the effects of MATCH can be undone when PROVE backtracks to recover from a failed application of an axiom. VII EFFICIENCY HCPRVR is surprisingly efficient for its simplicity. The compiled code fits in 2000 octal words of binary programming space and runs as fast t: the Prolog interpreter. Although the speed can further improved by more sophisticated programming, we have not done so because it is adequate for our present needs. A version of HCPRVR has been written in C; it occupies 4k words on a PDP11/60 and appears to run about half as fast as the compiled Lisp version does on a DEC KIlO. The most important kind of efficiency we have noticed, however, is program development efficiency, the ease with which logic programs can be written and debugged. We have found it easier to write natural language processing systems in logic than in any other formalism we have tried. Grammar rules can be easily written as axioms, with an unrestricted mixture of syntactic and non-syntactic computations. Furthermore, the same grammar rules can be used for parsing or generation of sentences with no change in the algorithm that applies them. Other forms of natural language processing are similarly easy to program in logic, including schema instantiation, question-answering and text summary. We have found HCPRVR very useful for gaining experience in writing logic programs. REFERENCES [l] Kowalski, R. A. "Algorithm = logic + control." CACM 22, 7, July, 1979, 424-436. [2] Warren, D. H., L. M. Pereira, and F. Pereira. "PROLOG - the language and its implementation compared with lisp." Proc. SymP' AI and Prog. Langs., SIGPLAN 12, -___ 8/sIGARTx4, August, 1977, 109-115. 95	1980	11
3	FIRST EXPERIMENTS WITH RUE AUTOMATED DEDUCTION Vincent J. Digricoli The Courant Institute and Hofstra University 251 Mercer Street, New York, N.Y. 10012 ABSTRACT RUE resolution represents a reformulation of binary resolution so that the basic rules of inference (RUE and NRF) incorporate the axioms of equality. An RUE theorem prover has been imple- mented and experimental results indicate that this method represents a significant advance in the handling of equality in resolution. A. Introduction In (1) the author presented the complete theory of Resolution by Unification and Equality which incorporates the axioms of equality into two inference rules which are sound and complete to prove E-unsatisfiability. Our purpose here is to present systematically the results of experiments with an RUE theorem prover. The experiments chosen were those of McCharen, Overbeek and Wos (2), ahd in particular we are interested in comparing the results achieved by these two theorem provers. In MOW, the equality axioms were used expli- citly for all theorems involving equality and apparently no use was made of paramodulation. In RUE, where proofs are much shorter, the inference rules themselves make implicit use of the equality axioms which do not appear in a refutation and also no use of paramodulation is made. Both systems are pure resolution-based systems. Before considering the experiments, we first review and summarize the theory of resolution by unification and equality as presented in (1). There we define the concept of a disagreement set, the inference rules RUE and NRF, the notion of viability, the RUE unifying substitution and an equality restriction which inhibits redundant inferences. Here we simply introduce the concept of a disagreement set and define the rules of inference. A disagreement set of a pair of terms (tl,t2) is defined in the following manner : is the on y disagreement set and if "If (tl,ts) are identical, the empty ;;st2) differ, the set of one pair { (tl,t2) is the origin disagreement set. i, Furt ermore, if tl has the form f(al,...,ak) and t2 the form f(bl,...,bk), then the set of pairs of corresponding arguments which are not iden- tical is the topmost disagreement set -0 In the simple example : tl = f( at g(WW) 1 t2 = f( a', g(b',h(c')) 1 besides the origin disagreement, there are the disagreement sets : Dl = i (ana'), ( g(b,h(c)) ,g(b',h(c')) 1) D2 = { (a,a'), (b,b'), (h(c) ,h(c') 1) D3 = ( tara'), (b,b'), (c,c') 1 This definition merely defines all possible ways of proving t =t , i.e. we can prove t =t by proving equali 4 12 y ?n every pair of any one disagreement set. An input clause set, for ex- ample, may imply equality in D D3' but not in D2 Or it may most directly phove tl=t2 by or proving equality in D3. We proceed to define a disagreement set of -- complementary literals : W1,...,sn) f;(t, v.,tn) as the union of disagreement sets : D = .U D i=l,n i where D i is a disagreement set of (si,ti). We see immediately that : P(sl"",sn) A ht1 I-O&~) * D where D now represents the disjunction of inequal- ities specified by a disagreement set of P,;, and furthermore, that : f(al~~~~,ak) # f(blto..,bk) --$ D where D is the disjunction of inequalities speci- fied by a disagreement set of f(al,...,ak), f(bl,-wbk) 0 For example, p(f(a,g(b,h(c)))) A F(f(a',g(b',h(c')))) 4 afa' A bfb' A cfc' . The reader is invited to read (1) which states the complete theory of RUE resolution with many examples. Our primary concern here is to discuss experiments with an RUE theorem prover and to begin to assess the effectiveness of this 96 inference system. we Experiments Our experiments deal with Boolean Algebra are asked to prove from the eight axioms : Al :x+0=x A2 :x1=x A3 :x+Z=l A4 : x xt=o A5 : x(y+z) = xy +x2 A6 : x + yz = (x+y) (x+2) A7 :x+y=y+x A0 :xy=yx (we are denoting logical or by +r logical and by * or juxtaposition, and negation by overbar), the following theorems : and a0 # 0 I - xx = 0 a: # a0 d = (a/x} - x(y+z) = xy + xz (7 ={a/x} y+z # ;; ay+az # a0 t- x+0=x o- = E a/y,o/z a/x j a: + a0 # a0 t- 0+x=x 0- =laO/x) Tl : Z=l T2 :x+1=1 T3 :x0=0 T4 : x + xy = x T5 : x(x+y?" = x T6 :x+x=x T7 :xx=x T8 : (x+y) +z = x+(y+z) T9 : (xy)z = x* (yz) TlO : the complement of x is unique (xa=O) (x+a=l) (xb=O) (x+b=l) 4 a= b Tll :z=x -- T12 :x+y =xy De Morgan's Law I -- T13 :xy=x+;; De Morgan's Law II These theorems are stated in the order of increasing complexity of proof, with 6= 1 being trivially easy for a human to prove and De Morgan's Laws being very difficult for a human to deduce from the axioms. George and Garrett Birkhoff have a paper on the above proofs published in the Transactions of the American Mathematical Society (3) and Halmos comments on the significantly difficult character of the proofs in his Lectures on Boolean Algebras (4) 0 The following is a machine deduced, five step RUE refutation which proves x0 = 0 : 0 # aa - 0 = x; D o-=<a/x] The above experiments together with many others (dealing with group theory, ring theory, geometry, Henken Models, set theory and program verification) were proposed as benchmarks by McCharen,Overbeek and Wos, who in (2) published the results of their own experiments. We here tabulate the comparative performance of the RUE and MOW theorem provers on the above thearems. The MOW theorem prover uses binary resolution with explicit use of the equality axioms and is implemented in Assembly language on the IBM System 370-Model 195. Great effort was made to enhance the efficiency of their theorem prover and this is described in (2). The RUE theorem prover, on the other hand, represents a first implementation in PLl on a CDC 6600 machine which is much slower than the Model 195. In the experiments each theorem is treated as an independent problem and cannot use earlier theorems as lemmas, so that for example in proving associativity (T8), we need to prove (T2,T3,T4,T5) as sub-theorems. The total number of unifications performed is suggested as the primary measure of comparison rather than time. The comparative results are given in Table 1. From Tl to T7, The RUE theorem prover was very successfull, but at T8 (associativity) results have yet to be obtained since refinements in the heuristic pruning procedure are required and are being developed with the expectation that more advanced results will be available at the conference. RUE represents one of several important methods for handling equality in resolution and it is important to emphasize that it is a complete method whose power is currently being tested in stand-alone fashion. However, it is not precluded that we can combine this method with other tech- niques such as demodulation,paramodulation and reduction theory to achieve a mutually enhanced effect. 97 TABLE 1. THEOREM TOTAL NUMBER OF UNIFICATIONS RUE : MOW Tl 'i=l 77 26,702 T2 x+1=1 688 46,137 T3 x0=0 676 46,371 T4 x+xy=x 3,152 see below T5 X(X+Y) = x -3,113 tl I! T7 XX=X 2,145 n m T6rrT7 4,326(l)105,839 T8 (x+y)+z = x+(y+z) IP 413,455 T9 (xy)z = x(yz) IP NPR TlO ';:;~;I ';;-;I+ a=b IP NPR ( -I( 1 e Tll z =x IP NPR T12 x+y = ;; * ;; IP NPR - - - T13 xy=x+y IP NPR TIME LENGTH OF PROOF (SECONDS) RUE : MOW RUE : MOW 10.1 27.5 7 51.5 " " 12 102.9 57.0 24 41.6 see below 13 Note 1 : To prove the double theorem, T4,,T5, x+xy=x A x(x+y)=x, we add the negated theorem as a single clause, a+ab#a v a(a+b)#a , to the input clause set. It is evident that the erasure of these two literals in a refutation decomposes into two independent subproblems since no variables appear in the clause. Hence, the refutations for a+ab#a and a(a+b)#a obtained in separate experi- ments T4,T5 can be concatenated and the results of these experiments simply summed which is what we have done to state the RUE results for the double theorem. The same holds true for T6hT7. The estimated length of MOW proofs with the equality axioms is twice as long as corresponding RUE proofs. REFERENCES : 1. "Automatic Deduction and Equality" by Vincent J. Digricoli, Proceedings of the Oct., 1979, Annual Conference of the ACM, 240-250. 2. "Problems and Experiments for and with Automat- ed Theorem-Proving Programs", by McCharen, Overbeek and Wos, IEEE Transactions on Comput- ers, Vol C-25,No.8,August 1976. 3. "Distributive Postulates for Systems Like Boolean Algebras", Transactions of the American Mathematical Society, Volume 60, July-Dec. 1946. 4. "Lectures on Boolean Algebras", by Paul R. Halmos, Section 2, pages 3-5, D.Van Nostrand. The completion of these Boolean experiments together with other experiments with a more fully delineated comparative analysis with systems other than MOW represents work the author will report upon in the future. 98	1980	12
4	WHAT’S WRONG WITH NON-MONOTONIC LOGIC? David J . Israel Bolt Beranek and Newman Inc. 50 Moulton St. Cambridge, Mass. 02238 ABSTRACT In this paper ’ I ask, and attempt to answer, the following question : What’s Wrong with Non-Monotonic Logic? The answer, briefly’ is that the motivation behind the wonderfully impressive work involved in its development is based on a confusion of proof-theoretic with epistemological issues. ------------ What’s wrong with non-monotonic logic (and fo! that matter, with the logic of default reasoning)? The first question we should ask is: What’s supposed to be wrong with “standard”, monotonic logic? In recent - and extremely impressive - work, Doyle and McDermott [ 1 I, McDermott C21, and Reiter C31 have argued that classical logic - in virtue of its monotoniqity - is incapable of adequately capturing or representing certain crucial features of real live reasoning and inference. In particular’ they note that our knowledge is always incomplete, and is almost always known to be so ; that, in pursuing our goals - both practical and theoretical - we are forced to make assumptions or to draw conclusions on the basis of incomplete evidence ; conclusions and assumptions which we may have to withdraw in the light of either new evidence or further cogitation on what we already believe. An essential point here is that new evidence or new inference may lead us to reject previously held beliefs, especially those that we knew to be inadequately supported or merely presumptively assumed. In sum, our theories of the world are revisable; and thus our attitudes towards at least some our beliefs must likewise be revisable. Now what has all this to do with logic and its monotonicity? Both Reiter and Doyle-McDermott characterize the monotonicity of standard logic in syntactic or proof-theoretic terms. If A and B are two theories, and A is a subset of B, then the ‘The research reported in this paper was supported in part by the Advanced Research Projects Agency, and was monitored by ONR under Contract No. NOOO14-77-C-0378. To remedy this lack, Doyle and McDermott introduce into an otherwise standard first order language a modal operator “M” which, they say, is to be read as “It is consistent with everything that is believed that.. .” (Reiter’s “M”, which is not a symbol of the object language, is also supposed to be read “It is consistent to assume that..“. I think there is some unclarity on Reiter’s part about his “M”. He speaks of it in ways conducive to interpreting it as a metalinguistic predicate on sentences of the object language ; and hence not as an operator at all, either object-language or metalanguage. So his default rules are expressed in a language whose object-language contains sentences of the form l’Mp” , i .e . , in a language which, relative to the original first-order object language, is a meta-meta-language .) Now in fact this reading isn’t quite right. The suggested reading doesn’ t capture the notion Doyle-McDermott and Reiter seem to have in mind. What they have in mind is, to put it non-linguistically (and hence, of course, non-syntactically) : that property that a belief has just in case it is both compatible with everything a given subject believes at a given time and remains so when the subject’s belief set undergoes certain kinds of changes under the pressure of both new information and further thought, and where those changes are the result of rational epistemic -----_ policies. 99 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. I’ve Put the notion in this very epistemologically oriented way precisely to hone in on what I take to be the basic misconception underlying the work on non-monotonic logic and the logic of default reasoning. The researchers in question seem to believe that logic - deductive logic , for there is no other kind - is centrally and crucially involved in the fixation and revision of belief. Or to put it more poignantly, they mistake so-called deductive rules of inference for real, honest-to-goodness rules of inference. Real rules of inference are precisely rules of belief fix ation and revision ; deductive rules of transformation are precisely not. Consider that old favorite : modus ( ponendo) ponens. It is not a rule that should be understood as enjoining us as follows : whenever you believe that p and believe that if p then q, then believe that q. This, after all, is one lousy policy. What if you have overwhelmingly good reasons for rejecting the belief that q? All logic tells you is that you had best reconsider your belief that p and/or your belief that if p then q (or, to be fair, your previously settled beliefs on the basis of which you were convinced that not-q); it is perforce silent on how to revise your set of beliefs so as to . . to what? Surely, to come up with a good theory that fits the evidence, is coherent, simple, of general applicability, reliable, fruitful of further testable hypotheses, etc. Nor is it the case that if one is justified in believing that p and justified in believing that if p then q (or even justified in believing that p entails q) , is one justified in be1 iev ing (inferring) that c~. Unless, of course, one has no other relevant ---- beliefs. Butone always does. ---- The rule of modus ponens is, first and foremost, a rule that permits one to perform certain kinds of syntactical transformations on (sets of) formally characterized syntactic entities. (Actually, first and foremost, it is not really a rule at all; it is f’reallyff just a two-place relation between on the one hand an ordered pair of wffs., and on the other, a wff .> It is an important fact about it that, relative to any one of a family of interpretations of the conditional, the rule is provably sound, that is Nor is it quite clear. By “consistentff are we to mean syntactically consistent in the standard monotonic sense of syntactic derivability or in the to-be-explicated non-monotonic sense? Or is it semantic consistency of one brand or another that is in question? This unclarity is fairly quickly remedied . We are to understand by If consistencyff standard syntactic consistency, which in standard systems can be understood either as follows: A theory is syntactically consistent iff there is no formula p of its language such that both p and its negation are theorems, or as follows : iff there is at least one sentence of its language which is not a theorem. There are otherwise standard, that is, monotonic, systems for which the equivalence of these two notions does not hold; and note that the first applies only to a theory whose language includes a negation operator. truth (in an interpretation)-preserving . The crucial point here, though, is that adherence to a set of deductive rules of transformation is not a sufficient condition for rational belief; it is sufficient (and necessary) only for producing derivations in some formal system or other. Real rules of inference are rules (better : policies) guiding belief fixation and revision. Indeed, if one is sufficiently simple-minded, one can even substitute for the phrase ” good rules of inference”, the phrase ‘I( rules of) scientific procedure” or even “scientific method”. And, of tour se, there is no clear sense to the phrase “good rules of transformation”. (Unless ffgoodff here means ffcompleteff - but with respect to what? Truth? > Given this conception of the problem to which Doyle-McDermott and Reiter are addressing themselves, certain of the strange properties of, on the one hand, non-monotonic logic and on the other, the logic of default reasoning, are only to be expected. In particular, the fact that the proof relation is not in general decidable. The way the “Mfl operator is understood, we believers are represented as follows: to make an assumption that p or to put forth a presumption that p is to be1 iev e a proposition to the effect that p is consistent with everything that is presently believed and that it will remain so even as my beliefs undergo certain kinds of revisions. And in general we can prove that p only if we can prove at least that p is consistent with everything we now be1 iev e . But, of course, by Church’s theorem there is no uniform decision procedure for settling the question of the consistency of a set of first-order formulae . (Never mind that the problem of determining the consistency of arbitrary sets of formulae of the sentential calculus is NP-complete . > This is surely wrong-headed : assumptions or hypotheses or presumptions are not propositions we accept only after deciding that they are compatible with everything else we be1 iev e , not to speak of having to establish that they won’t be discredited by future evidence or further reasoning. When we assume p, it is just p that we assume, not some complicated proposition about the semantic relations in which it stands to all our other beliefs, and certainly not some complicated belief about the syntactic relations any one of its linguistic expressions has to the sentences which express all those other beliefs. (Indeed, there is a problem with respect to the consistency requirement, especially if we allow be1 ief s about beliefs. Surely, any rational subject will believe that s/he has some false be1 iefs , or more to the point, any such subject will be disposed to accept that belief upon reflection. By doing so, however, the subject guarantees itself an inconsistent belief-set; there is no possible interpretation under which all of its beliefs are true. Should this fact by itself worry it (or us?) .) After Reiter has proved that the problem of determining whether an arbitrary sentence is in an extension for a given default theory is undecidable, he comments: (A)ny proof theory whatever for... the facts? (Are the rules provably sound rules of transformation?) Or are the conclusions legitimate because they constitute essential (non-redundant) parts of the best of the competing explanatory accounts of the original data; the best by our own, no doubt somewhat dim, lights? (Are the rules arguably rules of rational acceptance?) At the conclusion of his paper, McCarthy disambiguates and opts for the right reading. In the context of an imaginative discussion of the Game of Life cellular automaton, he notes that "the program in such a computer could study the physics of its world by making theories and experiments to test them and might eventually come up with the theory that its fundamental physics is that of the Life cellular automaton. We can test our theories of epistemology and common sense reasoning by asking if they would permit the Life-world computer to conclude, on the basis of its experiments, that its physics was that of Life." McCarthy continues: default theories must somehow appeal to some inherently non semi-decidable process. [That is, the -proof-relation, not just the proof predicate, is non recursive; the proofs, not just the theorems, are not Why such a beast recursively enumerable. is to be called a logic is somewhat beyond me - DI.1 This extremely pessimistic result forces the conclusion that any computational treatment of defaults must necessarily have an heuristic component and will, on occasion, lead to mistaken beliefs. Given the faulty nature of human common sense reasoning, this is perhaps the best one could hope for in any event. Now once ag ain substitute in the above "(scient ific or common sense) reasoni ng I1 for "defaulted and then reflect on how odd it is to think that there could be a purely proof-theoretic treatment of scientific reasoning. A heuristic treatment, that is a treatment in terms of rational epistemic More generally, we can imagine a metaphilosophy that has the same relation to philosophy that metamathematics has to mathematics. Metaphilosophy would study mathematical (? - D.1.) systems consisting of an 'fepistemologist'f seeking knowledge in accordance with the epistemology to be tested and interacting with a ffworldff. It would study what information about the world a given philosophy would obtain. This would depend also on the structure of the world and the ffepistemologist'sff opportunities to interact. AI could benefit from building some very simple systems of this kind, and so might philosophy. policies, is not just the best we could hope for. It is the only thing that makes sense. (Of course, if we are very fortunate a "syntactic" encoding we may be able to develop f these policies; but we 0 certainly mustn't expect to come up with rules for rational belief fixation that are actually provably truth-preserving. Once again, the only thing that makes sense is to hope to form ulate a set of rules which, from within our current theory of the world of ourselves as both objects within and inouirers about that world, can be argued to embody rational policies for extending our admittedly imperfect grasp of things.) Inference (reasoning) is non-monotonic: New information (evidence) and further reasoning on old beliefs (including, but by no means limited to, reasoning about the semantic relationships - e.g., of entailment - among beliefs) can and does lead to Amen; but might I note that such a metaphilosophy does exist. Do some substituting again: for " hilosophyf' P (except in its last occurrence), substitute ff~~iencef'; for 'fepistemologistff, ffscientistff; for ffepistemologyff, either "philosophy of science" or "scientific methodology". The moral is, I hope, clear. Here is my constructive proposal: AI researchers interested in "the epistemological problem" should look, neither to formal semantics nor to proof-theory; but to - of all things - the philosophy of science and epistemology. the revision of our theories and, of course, to revision bv f'subtractionff as well as by ffaddition'f. Entailment- and derivability are monotonic. That is, logic - the logic we have, know, and - if we understand its place in the scheme of things - have every reason to love, is monotonic. BRIEF POSTSCRIPT I've been told that the tone of this paper is REFERENCES overly critical; or rather, that it - lacks constructive content. A brief postscript is not [IlMcDermott, D., Doyle, J. "Non-Monotonic Logic I" , AI Memo 486, MIT Artificial Intelligence Laboratory, Cambridge, Mass., August 1978. C2lMcDermott. D. "Non-Monotonic Logic II", Research Report 174, Yale University Department of Computer Science, New Haven, Conn., February 1980. the appropr iate locus for correcting this defect; but it may be an appropri ate place for cas ting my vote for a suggestion made by John McCarthy. In his "Epistemological Problems of Artificial Intelligence" [41. McCarthy characterizes the epistemological part of "the AI problem" as follows: "(it) studies what kinds of facts about the world are available to an observer with given opportunities to observe, how these facts can be represented in the memory of a computer, and what rules permit legitimate conclusions to be drawn - - from these facts." [Emphasis added.] ThisTthough brief, is just about right, except for a perhaps C3lReiter, R. "A Logic for Default Reasoning", Technical Report 79-8, University of British Columbia Department of Computer Science, Vancouver, B.C., July 1979. 143McCarthy' J. "Epistemological Problems of studied ambiguity in that final clause. Are the Artificial Intelligence",-In Proc. IJCAI-77. Cambridge, Mass., August, 1977, pp. 1038-1044. conclusions legitimate because they are entailed by 101	1980	13
5	PATHOLOGY ON GAME TREES: A SUMMARY OF RESULTS* Dana S. Nau Department of Computer Science University of Maryland College Park, MD 20742 ABSTRACT Game trees are widely used as models of various decision-making situations. Empirical results with game-playing computer programs have led to the general belief that searching deeper on a game tree improves the quality of a decision. The surprising result of the research summarized in this paper is that there is an infinite class of game trees for which increasing the search depth does not improve the decision quality, but instead makes the decision more and more random. I INTRODUCTION - Many decision-making processes are naturally modeled as perfect information games between two players [3, 71. Such games are generally represented as trees whose paths represent various courses the game might take. In artificial intelligence, the well-known minimax procedure [2, 71 is generally used to choose moves on such trees. If a correct decision is to be guaranteed using minimaxing, substantial portions of the game tree must be searched, even when using tree- pruning techniques such as alpha-beta [2, 71. This is physically impossible for large game trees. However, good results have been obtained by searching the tree to some limited depth, estimating the minimax values of the nodes at that depth using a heuristic evaluation function, and computing the minimax values for shallower nodes as if the estimated values were correct [2, 71. There is almost universal agreement that when this is done, increasing the search depth increases the quality of the decision. This has been dramatically illustrated with game-playing computer programs [l, 8, 91, but such results are purely empirical. the The author has developed a mathematical ory modeling the effects of search depth on the * This work was supported in part by a National Science Foundation graduate fellowship, in part by a James B. Duke graduate fellowship, and in part by N.S.F. grant number ENG-7822159 to the Laboratory for Pattern Analysis at the University of Maryland. The results discussed in this paper are presented in detail in the author's Ph.D. dissertation [4]. probability of making a correct decision. This research has produced the surprising result that there is an infinite class of game trees for which as long as the search does not reach the end of the tree (in which case the best possible decision could be guaranteed), deeper search does not improve the decision quality, but instead makes the decision more and more random. For example, probability of search depth 0 2 4 6 8 10 12 14 16 FIGURE l.--Probability of correct decision as a function of search depth on the game tree G(l,l), for five different evaluation functions. On G(l,l), the probability of correct decision is 0.5 if the choice is made at random. For each of the five functions, this value is approached as the search depth increases. 102 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Figure 1 illustrates how the probability of correct decision varies. Section 2 of this paper summarizes the mathematical model used in this research, Section 3 presents the main result, and Section 4 contains concluding remarks. II THE MATHEMATICAL MODEL -- Let G be a game tree for a game between two players named Max and Min. Nodes where it is Max's or Min's move are called max and min nodes, respectively. Assume that G has no draws (this restriction can easily be removed, but it simplifies the mathematics). Then if G is finite, every node of G is a forced win for either Max or Min. Such nodes are called "+I' nodes and 1(-11 nodes, respectively. If G is infinite, not every node need be a "+,. or II- II node, but the "+" and M-n labeling can easily be extended to all nodes of G in a way which is consistent with all finite truncations of G. Correct decisions for Max and Min are moves leading to "+" and u-11 nodes, respectively. "+" max nodes (which we call S nodes) may have both "+" and 1(-(1 children; "+" min nodes (T nodes) have only "+" children; -II min nodes (U nodes) may have both "+" and 11-(1 children; and --II max nodes (V nodes) have only 11-11 children. Thus it is only at the S and U nodes that it makes a difference what decision is made. These nodes are called critical nodes. An evaluation function on G may be any mapping e from the nodes of G into a set of numbers indicating how good the positions are estimated to be. For computer implementation, the range of e must be finite. We take this finite set to be fO,l,...,r}, where r is an integer. Ideally, e(g) would equal r if g were a "+" node and 0 if g were a 11-u node, but evaluation functions are usually somewhat (and sometimes drastically) in error. Increasing the error means decreasing e(g) if g is a "+" node and increasing e(g) if g is a 11-11 node. Thus if we assume that the errors made by e are independent and identically distributed, the p.d.f. f for the values e returns on "+" nodes is a mirror image of the p.d.f. h for the values e returns on 1)-11 nodes; i.e., f(x) = h(r-x), x = O,l,...,r. f may be represented by the vector P= (f(O),f(l),...,f(r>>, which is called the probability vector for e. --- 111 RESULTS -- The probability vector for e induces probability vectors on the minimax values of the nodes of G, and the probability of making a correct decision at any critical node g of G is a function of the probability vectors for the minimax values of the children of g. This probability is thus determined by the structure of the subtree rooted at g, and little can be said about it in general. However, if G has a sufficiently regular structure, the properties of this probability can be analyzed. Let m and n be positive integers, and let G(m,n) be the unique game tree for which 1. the root is an S node (this choice is arbitrary and the results to follow are independent of it); 2. each critical node has m children of the same sign and n children of opposite sign; T . . . T U . . . U U . . . U m+n A-A . . . . . . FIGURE 2.--The game tree G(m,n). Min nodes are indi- cated by the horizontal line segments drawn beneath them. 103 3. every node has m+n children. G(m,n) is illustrated in Figure 2. If moves are chosen at random on G(m,n), the probability that a correct choice is made at a critical node is obviously m/(m+n). If the choice is made using a depth d minimax search and an evaluation function with probability vector P, it is proved [4, 51 that the probability that the decision is correct depends only on m, n, P, and d. We denote this probability by &,n(P,d). The trees G(m,n) have the following surprising property. Theorem 1. For almost every probability ~- vector P and for all but finitely many values of m and n, lim &,n(P,d) = m/(m+n). d-3 co Thus, as the search depth increases, the probability of correct decision converges to what it would be if moves were being chosen at random. This pathological behavior occurs because as the search depth increases it becomes increasingly likely that all children of a critical node receive the same minimax value, whence a choice must be made at random among them. Figure 1 illustrates Theorem 1 on the game tree G(l,l), using five different values of P. The significance of Theorem 1 for finite games is that infinitely many finite games can be generated by truncating G(m,n) in whatever way desired. Deeper search on these trees will yield increasingly random decisions as long as the search does not reach the end of the tree. Additional theoretical and experimental results reported elsewhere [4, 5, 61 provide additional information about which of the G(m,n) are pathological and why. Theorem 1 almost certainly extends to a much larger class of game trees, but the irregular structure of most game trees would require a much more complicated proof. IV CONCLUSIONS - The author believes that the pathology of the trees G(m,n) indicates an underlying pathological tendency present in most game trees. However, in most games this tendency appears to be overridden by other factors. Pathology does not appear to occur in games such as chess or checkers [l, 8, 91, but it is no longer possible blithely to assume (as has been done in the past) that searching deeper will always result in a better decision. 1. 2. 3. 4. 5. 6. 7. 8. 9. REFERENCES Biermann, A. W. Theoretical issues related to computer game playing programs. Personal Comput. (Sept. 1978), 86-88. Rnuth, D. E., and Moore, R. W. An analysis of alpha-beta pruning. Artif. Intel. 6 - - (1975), 293-326. LaValle, I. H. Fundamentals of Decision Analysis. Holt, Rinehart and Winston, New York, 1978. Nau, D. S. Quality of decision versus depth of search Dissertation, Dull Uni~?m~Aug~~~eSb79). Ph.D. Nau, D. S. Decision quality as a function of search depth on game trees. Tech. Report TR-866, Computer Sci. Dept., Univ. of Md. (Feb. 1980). Submitted for publication. Nau, D. S. The last player theorem. Tech. Report TR-865, Computer Sci. Dept., Univ. of Md. (Feb. 1980). Submitted for publication. Nilsson, N. J. Problem-Solving Methods in Artificial Intelligence. --ram NG York, 1971. Robinson, A. L. Tournament competition fuels computer chess. Science 204 (1979), 1396-l 398. Truscott, T. R. Minimum variance tree searching. Proc. First Int. Symp. on Policy Anal. and IX------- Syst. (1979X 203-209.- - - * A property holds for almost every member of a set if it holds everywhere but on a subset of measure zero. Thus for any continuous p.d.f. on the set, the probability of choosing a member of the set to which the property does not apply is 0. 104	1980	14
6	ISEOC9PFREE S. W. Ng and Adrian Walker Work performed at Rutgers University* ABSTRACT If a system uses assertions of the general form x causes y , (e.g. MYCIN rules) then loop situations in which X, causes X2, X2 causes X3, . . . . , X, causes X,, are, intuitively, best avoided. If an assertion has an attached confidence weight, as in x (0.8)-causes y , then one can choose to say that the confidence in a chain of such assertions is as strong as the weakest link in the chain. If there are several chains of assertions from X to Z, then one can choose to say that X causes Z with a confidence equal to that of the strongest chain. From these cfioices, it follows that the confidence that X causes Z corresponds to a loop-free chain of assertions. This is true even if there are chains from X to Z with common subchains and loops within loops. An algorithm for computing the confidence is described. I INTRODUCTION and TERMINOLOGY There is currently considerable interest in representing knowledge about a practical situation in the form of weighted cause-effect or situation-action rules, and in using the knowledge so represented in decision- making systems. For example, in medical decision mak- ing systems, the rules may represent causal trends in a disease process in a patient [6], or the rules may represent trends in the decision process of a physician who is diagnosing and treating a patient [2,4]. In such representations, the chaining together of rules can be written as a weighted, directed graph. In MYCIN [2] the graphs are like and-or trees, while in OCKHAM [3,4,5] the graphs may have loops. This paper presents a result which appears in [l]. From the result it follows that, using the max and min operations, a graph con- taining loops can be interpreted as though it were loop- free. Authors’ pres ent addresses: S. W. Ng, 6F Wing Hing Street, Hong Kong. Adrian Walker, Bell Laboratories, Murray Hill, NJ. The kind of graph in question is a stochastic graph (sg) consisting of nodes N = {1,2,...,n) and a function P from N X N to the real numbers W, 0 5 w 11. P is such that, for each i E N, 2 P(i,j) 5 1. If P(i,j) = W, then w J=i is the weight of the arc from node i to node j. A path in an sg is a string n, . . nl c N+ such that P(nk,nk+,) > 0 for lsk-cl. n2,...,ni-l are intermediate nodes of . . . q. A path n, . ;dop if n, = n nl of a graph is said to have a J for some i.j such that either 1 li < j <I or I <i < j I I. Otherwise the path is loop-free. The weight of a path n, n2 . nl of an sg is the minimum over 1 <i < I of the weight of the arc from n, to n,+l. The k-weight w: from node i to node j of a graph, is the maximum of the weights of all the paths from i to j having no intermediate node with number higher than k. The weight w,/ from node i to node j of an sg is w,;. II EXAMPLES This section gives examples of potential causal loops, in MYCIN [2] and in OCKHAM [3,4,5], and it shows how these loops are avoided by the use of the maximum and minimum operations. A. A MYCIN Example Consider a set of MYCIN rules B A C (l.O)- A B (l.O)- D D V E (0.5)- B G AH (0.5)- B and suppose that C, E, G, and H are known with confidences 0.9, 0.8, 0.5, 0.4, respectively. Writing C(X) for the confidence in X, confidences propagate through rules by: 105 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. c(Z) = w . max(c(X),c(Y)) for xv Y (w)+ 2 III ALGORITHM and RESULTS and, c(Z) = w . min(c(X),c(Y)) for x A Y (w)-2. The greatest confidence which can be computed in A is c (A ) = 0.4 by the tree A+((Bt-(D+(BG AH)) V E) A C) B occurs twice, so the tree can be thought of as a graph with a loop. However, the value of C(A) depends only on the loop-free path EBA. B. An OCKHAM Example The following set of OCKHAM [3,4,5] rules is intended to show a strategy of a person who is deciding whether to stay at home, go to the office, or to try for a standby flight to go on vacation. The external factors project deadline, sno&torm, project completed, and another flight influence the choice by placing the arc(s) so labelled in a stochastic graph. The rules are: HOME (project deadline, 1 .O)+ OFFICE OFFICE (snowstorm, 0.5)--+ HOME OFFICE (project completed, 0.5)-+ AIRPORT-STANDBY AIRPORT-STANDBY (another flight, 0.25)+ AIRPORT-STANDBY AIRPORT-STANDBY (snowstorm, 0.75)-+ HOME These rules make up a stochastic graph with nodes HOME, OFFICE, and AIRPORT-STANDBY. If all of the external factors project deadline, snowstorm, project com- pleted, and another flight are true, then the graph has five arcs and multiple loops. If the weight from HOME to AIRPORT-STANDBY is considered, then it turns out to be 0.5. The corresponding path, HOME-OFFICE-AIRPORT-STANDBY, is loop-free. The algorithm MAXMIN, shown below, computes the weight from a given node to another node in an sg. Note that, by Step 2, MAXMIN runs in o(n3) time. MAXMIN Input: A stochastic graph of n nodes Output: n2 real numbers Step 1: for 1 5 i,j 5 n do B,,O := P(i,j) Step 2: for k:=l to n do for 1 1. i,j I n do B{ := max (BG-‘, min (B,kk-‘, BjJ-‘)) Step 3: for 1 I ij I n do output Bl The properties of paths, path weights, and the values Blf, described in the Lemma below, are esta- blished in Appendix I. Lemma In an sg of n nodes, the following hold for 1 5 i,j 5 n andforork (n: statements (i) If WI; > 0, then there exists i to j whose weight is w;, a loop-free path from (ii) B: = WI:. Setting k=n in parts (i) and (ii) of the Lemma yields the two results: Result I In any sg the weight wlJ, that is, the maximum path weight over all paths from i to j, is equal the max- imum over only the loop-free paths from i to j. Result 2 If MAXMIN receives as input an sg with n nodes, then, for 1 I ij I n, the output Bl; is equal the weight wi, from node i to node j of the graph. Result 1 establishes a property of any sg, namely that the weight from one node to another is the weight of some loop-free path, while Result 2 establishes that MAXMIN is one possible algorithm for finding such weights. 106 This is because (A) and (B) exhaust all possibilities. IV CONCLUSIONS In a system in which weighted causal assertions can be combined into causal paths and graphs, causal loops can occur. Common sense about everyday causality suggests that such loops are best avoided. If the weight of a path is chosen to be the minimum of the individual arc weights, and the net effect of a start node on a final node is chosen to be the maximum of the path weights from the start node to the final node, then the weights (by whatever algorithm they are computed) are independent of the presence of loops in the underlying graph. There is a simple O(n3) algorithm to compute these weights. ACKNOWLEDGEMENT Thanks are due to A. Van der Mude for helpful comments. APPENDIX I Proof of Lemma Let k = 0. If wij > 0, then by definition, there exists a path from i to j having no intermediate nodes, whose weight is wiio. Clearly this path is ij, which is loop-free. So we may write r,$’ = ij, where 76 denotes a path from i to j having no inter- mediate node with number greater than k. Then WijO = P (i j) = Biio. If wi! = 0, then there is no such path, and BijO = 0. Suppose, by way of inductive hypothesis, that for 1 5 ij I n and for some (k-l) -c n, (i) if wi’f-’ > o then there is a loop-free path r$-‘, from i to j with each intermediate node at most k-l, whose weight is w,$-l, and (ii) B,!j-l = ~4~~. If w,$-’ > o then there is a path 7 from i to j whose weight is w$. y is such that either (A) each intermediate node of 7 is at most (k-l), or (B) y goes from i to k; from k to k some number of times, then from k to j, with each intermediate node of each subpath being at most (k-l). In case (A) it is clear that wlf = WI:-‘, and the induc- tive step for part (i) of the Lemma is completed with k- -f?I k-‘. - Y,, In case (B), it follows from our induction hypothesis that there exist loop-free paths r,kk-‘, r&‘, rkk,-’ with weights wi-‘, wrtk- ’ , wkJ k-’ respectively. Let w = min(w,i-‘,wl,-‘) and w’ = w,&‘, and consider the sub- cases (Bl) in which y goes from k to k zero times, and (B2) in which y goes from k to k one or more times. In (Bl) the weight of y is clearly W, while in (B2) it is min(w ,ti). Hence, from the definition of WI:, we have wif = max(w ,min (w ,ti)), which is simply w. So part (i) of the Lemma holds with 7,: = r,kk-’ riJ-‘. From part (ii) of the inductive hypothesis, and from Step 2 of the MAX- MIN algorithm, it follows that B; = max(wJ-‘,w). So Bk = max(wk-’ wkj = wk since it dlfinition zf k{that ‘z; zz w,:-‘. follows from the So in either of the cases (A) and (B) B[ = w,f, which establishes part (ii) of the Lemma for the case W; > O. If w; = 0 then there is no path from i to j. Suppose Bt f 0. Then either w,t-’ # 0, or both of w;-‘, wiJ-’ are nonzero. In each case there is a path from i to j, a con- tradiction. So if wlf = o then B: = w:. q HI PI [31 [41 [51 h5 REFERENCES Ng, S. W., and A. Walker “Max-min Chaining of Weighted Assertions is Loop-free”, CBM-TR-73, Dept. of Comp. Sci., Rutgers University, 1977. Shortliffe, E. Computer Based Medical Consulta- tions: MYCIN. American Elsevier, 1976. Van der Mude, A. and A. Walker “On the Infer- ence of Stochastic Regular C-7dmmars” Information and Control, 38:3 (1978) 310-329. Walker, A. “A Framework for Model Construc- tion and Model Based Deduction in a System with Causal Loops” In Proc. Thi ,i Illinois Conf. Med. Info. Syst., 1976. Walker, A. “On the Induction of a Decision Mak- ing System from a Data Base”, CBM-TR-80, Dept. of Comp. Sci., Rutgers University, 1977. Weiss, S. “Medical Modeling and Decision Mak- ing”, CBM-TR-27, Dept. of Comp. Sci., Rutgers University, 1974. 107	1980	15
7	Applying General Induction Methods to the Card Game Eleusis Thomas G. Dietterich Department of Computer Science Stanford University Stanford, CA 94305 Abstract Research was undertaken with the goal of applying general universally-applicable induction methods to complex real-world problems. The goal was only partially met. The chosen domain-the card game Eleusis-was still somewhat artificial, and the universally-applicable induction methods were found to be lacking in important ways. However, the resulting Eleusis program does show that by using knowledge-based data interpretation and rule evaluation techniques and model-fitting induction techniques, general induction methods can be used to solve complex problems. Introduction Work in the area of computer induction is characterized by a continuum from general, universally-applicable methods [5, 6, 7, 9, 10, 121 to specific, problem-oriented methods [2, 8, 111. The general-purpose methods have been criticized for lacking the power to operate in real-world domains. Problem-oriented methods have been criticized for being too specialized to be applied to any problems outside their original domains. This paper describes an attempt to bridge this gap by applying general-purpose induction algorithms to the problem of inducing secret rules in the card game Eleusis. Further details are available in [3]. A Program for Eleusis Eleusis (developed by Robert Abbott [l, 41) is a card game in which players attempt to induce a secret rule invented by the dealer. The secret rule describes a linear sequence of cards. In their turns, the players attempt to extend this sequence by playing additional cards from their hands. The dealer gives no information aside from indicating whether or not each play is correct. Players are penalized for incorrect plays by having additional cards added to their hands. The game ends when a player empties his hand. A record of the play is maintained as a layout (Figure 1) in which the top row, or mainline, contains all of the correctly-played cards in sequence. Incorrect cards are placed in side lines below the main line card which they follow. mainline: 3H QS 4C JD 2C 10D 8H 7H 2C 5H sidelines: JD AH AS IOH 5D 8H 10s QD Rule 1: “If the last card is odd, play black, if the last card is even, play red.” Figure 1. Sample Eleusis Layout (after El]). This research sought to develop a program which could serve as an intelligent assistant to a human Eleusis player. The program needed to be able to: ) discover rules which plausibly describe the layout, ) accept rules typed by the user and test &hem against the layout, ) extend the layout by suggesting cards to be played from the player’s hand. Although Eleusis is artificial and noise-free, it is sufficiently complex to provide a reasonable test bed for inductive techniques. The development of an intelligent assistant required not only basic induction methods but also extensive deduction techniques for testing rules and extending the layout. Problems with Existing Induction Methods While designing the rule-discovery portion of the Eleusis program, existing induction algorithms [5, 6, 7, 9, 10, 121 were examined and found to be lacking in three fundamental ways. The first major problem with some of these algorithms is their emphasis on conjunctive generalizations. Many Eleusis rules are disjunctive. For example, Rule 1 can be written as: tli {odd(cardi-1) A black(cardi) V even(cardi-1) A red(cardi)} The second major problem with these algorithms is that they make implicit assumptions concerning plausible generalizations-assumptions which are not easily modified. Of the algorithms examined, only Mitchell’s version space algorithm [lo] maintains information concerning all rules consistent with the data (and his algorithm is still oriented toward conjunctive generalization). The algorithms of Hayes-Roth and Vere both seek the most specific rule consistent with the data, while Michalski’s Aq algorithm seeks a disjunctive description with the fewest conjunctive terms. In contrast, the plausibility heuristics for Eleusis are: Choose rules with intermediate degree of generality. (Justification: the dealer is unlikely to choose a rule which is overly general because it would be too difficult to discover. Conversely, overly specific rules are easily discovered because they lead to the creation of numerous counter-examples during play.) Choose disjunctive rules based on symmetry. (Justification: Rule 1 is an excellent example of a symmetric disjunctive rule. Most often in Eleusis, the terms of a disjunction define mutually exclusive cases which have some symmetric relationship to each other. The dealer is very likely to choose such rules because they are not too hard-nor too easy--to discover.) (These plausibility heuristics are based on the assumption that the dealer is rational and that he is attempting to maximize his own score (according to the rules of the game). This is an artificial assumption. It is very rare in science that we have such insight into nature. However, in all domains plausibility criteria must be available-otherwise, we don’t know what we are searching for.) The third major problem with using general-purpose induction techniques in Eleusis is that the raw data of the Eleusis layout are not in a form suitable for generalization. (Many researcliers [2, 111 have pointed out this problem in other domains.) One aspect of this problem is evident in Rule 1: neither color nor parify is explicit in the representation of the cards. Another difficulty is that the sequential ordering of the cards is implicit in their position in the layout. It must be made explicit in order to discover rules like Rule 1. Two techniques were developed to address these problems. First, in order to avoid an exhaustive search of rule space and at the same 218 time avoid the “tunnel vision” of existing algorithms, rule models were developed to guide the induction process. Secondly, in order to transform the input data info a form appropriate for generalization, a series of knowledge-based processing layers were created. Induction by Model-Fitting By analogy with traditional statistical time-series analysis, the program uses a model-fitting approach to induction. The term model denotes a syntactic or functional skeleton which is fleshed out by the induction algorithms to form a rule. In traditional regression analysis, for example, the model is the regression polynomial whose coefficients must be determined by induction from the data. Properly chosen models can strongly constrain the search required for induction. After looking at several Eleusis games, the following models were designed for the Eleusis program: B Decomposition. This model specifies that the rule must take the form of an exclusive disjunction of if-lhen rules. The condifion parts of the rules must refer only to cards prior to the card to be predicted. The action parts of the if-then rules describe correct plays given that the condition parts are true. The condition parts must be mutually exclusive conjunctive descriptions. The action parts are also conjunctions. Rule 1 fits the decomposition model: Vi odd(cardi-1) * black(cmdi) V even( cardi- 1) =P red(cardJ ) Periodic. A rule of this model describes the layout as a periodic function. For example, Rule 2 (Figure 2) is a periodic rule. The layout is split into phases according to the length of the proposed period. The periodic model requires that each phase have a conjunctive description. JC 4D QH 3s QD 9H QC 7H QD 9D QC 3H KC 5s 4s 10D 7s M phase 0: JC QH QD QC QD QC 5s 4s 1OD 7s phase 1: 0 4D 3s 9H 7H 9D 3H KC Rule 2: (periodic rule with length 2): phase 0: Vi faced(cardi) phase 1: tli nonfaced(cardi) Figure 2. A Periodic Rule. ) Disjunctive Normal Form (DNF) with fewest terms. The Aq algorithm (Michalski [9]) is used to discover rules which have the fewest number of separate conjunctions. The Aq algorithm was given heuristics to guide it towards symmetric, disjoint disjunctive terms. By definition, not all Eleusis rules can be represented using these three models. But, these models, when combined with segmentation (see below), cover all but one or two of the Eleusis rules which I have seen. For each of these models, an algorithm was developed to fit the data to the model. In order to fit the data to the decomposition model, the program must determine which variables to decompose on, i.e. which variables to test in the condition part of the rule (Rule 1 decomposes on parity E {odd, even)). The program must also decide how far into the past this decomposition should apply (i.e. do we look at just the most recent card, or the two most recent cards, . . . . etc.). Once the decomposition variables and the degree of lookback are determined, the algorithe must find a conjunctive description for the action parts of the rules. The program uses a generate-and-test approach. First, it considers rules which look back only one card, then two cards, and so on until a rule consistent with the data is found. To determine the 219 decomposition variable(s), it generates trial rules by decomposing on each variable in turn and chooses the variable which gives the simplest rule. If the resulting rule is not consistent with the data, the layout is decomposed into sub-layouts based on the chosen variable, and a second decomposition variable is again determined by generating trial decompositions and selecting the simplest. This process is repeated until a rule consistent with the data is found. (This is a beam search with a beam width of 1). In order to fit the periodic model, the program chooses a length for the period, splits the layout into separate phases, and finds a conjunctive description of each phase. Since the rule is more plausible if the descriptions of each pha’se are mutually exclusive, the algorithm attempts to remove overlapping conditions in the descriptions of the different phases. Again, a generate-and-test approach is used to generate periodic rules with different length periods (from length 1 upwards) until an acceptable rule is discovered. The Aq aigorithm is used to fit dati to the DNF model. Knowledge-layer Structure Like many other AI systems, the Eleusis program is structured as a set of layers, or more accurately, rings, based on the kinds of knowledge used to solve the problem (Figure 3). Each layer takes input data from the outside, transforms the data using knowledge appropriate to this layer, performs generalization by searching the space of rules at this level of abstraction, and evaluates the obtained rules. Rules which are sufficiently plausible are returned to the outer layers. Each layer calls the inner layers to perform tasks requiring knowledge appropriate to those inner layers. Figure 4 shows the five layers of the Eleusis program. Notice that in Eleusis, the outermost ring is very specific to Elcusis, while the inner-most rings contain only the general model-fitting induction algorithms. This design is intended to allow the program to be easily applied to similar problems. Since all Eleusis-specific knowledge is in the outer-most two layers, these could be stripped off and replaced by layers which apply different kinds of data transformations to solve different problems (e.g. letter series completion, sequence extrapolation). Figure 3. The Knowledge-layer Scheme. 5 User Interface 4 Eleusis KnowledPe , ; Seg~ Sequential Analysis 1 Basic induction Most Specific CL Most General Figure 4. Layered Structure of Eleusis Program. The five layers fiinction as follows. The outer-most layer provides an interface for the user. Layer 4 transforms the layout by making explicit such things as the color, parity, prime-ness, and faced-ness of the cards. Layer 3 segments the layout. Segmentation is used to discover rules such as Rule 3 (Figure 5) which involve first splitting up the layout into segments according to some criterion (like constant color) and deriving a new layout based on the lengths of these segments. Layer 2 makes the order of the events explicit [4] Gardner, Martin, “On Playing the New EIeusis, the game that either by creating “first difference” variables (e.g. Avalue(caQ = simulates the search for truth,” Scientific American, 237, value(c&$ - value(cardi_l)) or by breaking the layout into separate October. 1977, pp 18-25. phases (for periodic rules). The result of the preprocessing of layers 5 through 2 is that layer 1 is called with a specific model for which [S] Hayes-Roth, F:; J. McDermotf “An Interference Matching Technique for Inducing Abstractions”, Communicufions of fhe the degree of Zookback and (optionally) length of period have been specified and with a set of unordered events to which the model is to be fitted. Layer 1 actually performs the model-fitting using one of the three model-fitting induction algorithms. Once the rules have been induced, they are passed back up through the layers for evaluation. Layer 2 evaluates the rules using knowledge about ordering (e:g. guaranteeing that the rule doesn’t lead to a dead end). Layer 3 checks that the rules are consistent with the scgmcntation it performed (in particular, the boundary values cause some problems). Layer 4 evaluates the rules according to the heuristics for plausible rules in Elcusis. Finally, layer 5 prints any rules which survive this evaluation process. ACM, 21:5, 1978, pp. 401-410. [6] Hunt, E.B.. Experiments in Induclion, Academic Press, 1966. f’7] Larson, J., “Inductive Inference in the Variable Valued Predicate Logic System VL21 : Methodology and Computer Implementation”, Rept. No. 869, Dept. of Comp. Sci., Univ. of III., Urbana, May 1977. [8] Lenat, D., “AM: An Artificial Intelligence Approach to Discovery in Mathematics as Heuristic Search,” Comp. Sci. Dept., Rept. STAN-CS-76-570, Stanford University, July 1976. 191 Michalski. R. S., “Algorithm Ag for the Quasi-Minimal Solution of the Covering Problem,‘* Archiwum -Automafyki i Telemechaniki, No. 4, Polish Academy of Sciences, 1969 (in Polish). [lo] Mitchell, T. M., “Version Spaces: an Approach to Concept Learning,” Comp. Sci. Dept. Rept. STAN-CS-78711, Stanford University, December 1978. I AH 7C 6C 9s 10H 7H 1 OD JC AD 4H 8D 7C KD 6s QD 3s I JH I Rule 3: “Play odd-length strings of cards where color is constant within each string.” 1 The segmenled layout looks like this (color, length): (red, 1) (black, 3)’ (red, 3) (black, 1) (red, 3) Figure 5. A Segmentation-based Rule. The program works well. The three rule models, when combined with segmentation, span a search space of roughly 1O183 possible rules (several control parameters affect the size of this space). The program generates and tests roughly 19 different parameterizations of the three models in order to choose three to five plausible rules. It runs quite quickly (less than seven seconds, on a Cyber 175, in the worst case so far). The rules developed are similar to those invented by humans playing the same games (15 complete games have been analyzed). Conclusion General induction techniques can be used to solve complex learning tasks, but they form only part of the solution. In the Eleusis domain, data interpretation, rule evaluation, and model-directed induction were all required to develop a satisfactory program, A degree of generality was obtained by segregating the functions of the program into layers according to the generality of the knowledge they required. This should allow the program to be applied to similar tasks merely by “peeling off’ and replacing its outer layers. Acknowledgments Thanks go to R. S. Michalski, my M.S. thesis advisor, for suggesting Eleusis as a domain and for providing numerous ideas including the basic idea for the decomposition algorithm. Thanks also to David Shur for proofreading this paper. NSF grant no. MCS-76-22940. This research was supported by References [l] Abbott, Robert, “The New Eleusis,” Available from Abbott at Box 1175, General Post Office, New York, NY 10001 ($1.00). [2] Buchanan, B.G., D. H. Smith, W. C. White! R. J. Gritter, E. A. Feigenbaum, J. Lederberg, C. Djerassr, Journal of the American Chemical Society, 98 (1976) p. 6168. [ll] Soloway, E., “Learning = Interpretation + Generalization: a case study in knowledge-directed learning,” PhD Thesis, COINS TR 78-13, University of Massachusetts, Amherst, MA., 1978. [12] Vere, S. A., “Induction of Relational Productions in the Presence of Background Information,” In Proceedings of the Fifrh International Joint Conference on Artificial Intelligence, MIT, Cambridge, MA., 1977. [3] Dietterich, Thomas G., “The Methodology of Knowledge Layers for Inducing Descriptions of Sequentially Orderid Events,” MS Thesis, Dent of Corn. Sci., Univ. of Illinois, Urbana, October, 1979. - 220	1980	16
8	MODELLING STUDENT ACQUISITION OF PROBLEM-SOLVING SKILLS Robert Smith Department of Computer Science Rutgers University New Brunswick, N. J. 08903 ABSTRACT This paper describes the design of a system that simulates a human student learning to prove tneorems in logic by interacting with a curriculum designed to teach those skills. The paper argues that sequences in this curriculum use instructional strateu, and that the student recognizgs these strategies in driving the learning process. I. INTRODUCTION A central issue in the design of learning systems (LS's) is the classification of the sources of information that the system uses for its acquisition. The general notion is that an LS begins with certain knowledge and capabilities, and then extracts information from training sequences or experiments in the acquisition process. Implicit within this general characterization is the idea that the LS enforces some kind of interpretation on the training sequence by way of driving the acquisition process. Often the nature of the interpretation is left implicit. An example of an LS that makes somewhat explicit the interpretation of its training sequence is Winston's program for learning structure descriptions, where the program interprets the near m example as providing key information about the structure being learned [lr]. We speculate that much human learning takes place in a more richly structured environment, wherein the human learner is interpreting the instructional seauences provided to him in a richer way than LS's typically envision. Indeed, most LS's nave made few if any explicit assumptions about the structure of the environment in which the training sequence occurs. One particularly rich environment is learning & teaching. We suggest that teachers use certain instructional strategies in presenting material, and that students recognize these strategies. This paper describes the motivation for an LS called REDHOT. REDHOT is a simulation of a student -------- * I would like to thank Phyllis Walker for analysis of the curriculum and student protocols; Saul Amarel, Tom Mitchell, Don Smith, and N. Sridharan for many ideas and assistance. The research reported here is sponsored by the Office of Naval Research under contract N00014-79-C-0780. We gratefully acknowledge their support for this work. acquiring the skill of constructing proofs in elementary logic. We characterize this skill as consisting of (1) primitive operators in the form of natural-deduction rules of inference, (2) "macro moves" consisting of several rules of inference, and (3) application heuristics that describe when to use the rules. The central theme of this research is to model the acquisition of these skills around the recognition of instructional strategies in a curriculum designed to teach the student. II. CURRICULUM FOR REDHOT We are using the curriculum from the computer- assisted instruction (CA11 course developed at Stanford University by Patrick Suppes and co- workers. (See [3] for details.) This CA1 system is used as the sole mode of instruction for approximately 300 Stanford students each year. We chose a CA1 curriculum because we thought that the explicitness inherent in a successful CA1 system developed and tested over a number of years might make the instructional strategies relatively clear. The curriculum contains explanatory text, examples, exercises, and hints. The explanatory text is rather large, and uses both computer- generated audio and display as modes of presentation. The presentation strategy used by the actual CA1 system is linear through the curriculum. For use with REDHOT, we have developed a stylized curriculum in an artificial language CL. It contains the examples, exercises, partially formed rules, and hints. The exercises are the central part of the curriculum. There are approximately 500 theorems that the student is asked to prove, with about 200 in propositional logic, 200 in elementary algebra, and 100 in the theory of quantification. The human student performs these exercises by giving the steps of the developing proof to an interactive proof checker. This proof checker is the heart of the original instructional system. We developed a version of this proof checker for use with the REDHOT student simulation. III. THE DESIGN OF REDHOT REDHOT learns rules for proving theorems. These rules are initially the natural deduction 331 rules of many logic systems. The student improves upon these rules by building macro operators and by adding heuristics to existing rules--i.e., giving strategic advice in the left-hand-sides of the production rules.* For example, the rule AA ("affirm the antecedent", the system's version of modus ponens) can be stated as the following rule: Rule AA I GOAL: Derive Q I Prerequisites: f P already on some line i I P -> Q on some line j I Method: f AA command on lines i and j I Heuristics: None (yet) I Effectiveness: Perfect I In the above, we have adopted a style of rule presentation that is similar to the rules of a production system. The letters P and Q stand for arbitrary formulas, and i and j for arbitrary lines of the already existing proof. The aoal tells what the rule will produce. The mreauisites tell us that two formulas of the appropriate forms are needed, and the method gives the schematic command to the proof checker that will accomplish this. The heuristics associated with a rule are learned by the system, and indicate how the rule should be used. Effectiveness of the rule is also learned, and indicates how effective the rule will be in achieving its goal given its prerequisites. The effectiveness of this rule is "perfect" since the rule is given as such in the curriculum. The underlying problem solver for REDHOT is a depth-first, heuristic problem solving over the existing rules. It is assumed that: (a) the system has sufficient memory and CPU facilities for this search, if necessary; and (b) that the underlying pattern matchers are sufficient to match arbitrary formulas, line numbers, etc. Both of these assumptions are open to criticism on the grounds of being psychologically unrealistic. One of the goals of the construction of REDHOT is to decide on plausible ways to restrict the problem solver and pattern matcher to make a more realistic system. REDHOT learns the heuristics for the application of the rule. These heuristics are stated in a heuristic-language HL, which is strongly tied to the curriculum language CL. The heuristics associated with a rule guide the student as to when to try that rule or not to try it. For example, the curriculum appears to teach the student that the AA rule is a plausible thing to try when the prerequisites are available -------- * See [II for a conceptual discussion of the levels through which this process might proceed. One way to regard this research is a suggestion of the mechanism for this acquisition of heuristics and macro moves. (whether or not the goal is immediately desired). This is one of the primitives of the HL heuristics language. An example of a macro operator that is not a primitive rule is the "multiple AA" macro move. A paraphrase of this macro operator might be: I Multiple-AA Macro Move I I IF you want to prove Q : I AND I I have P, P -> P1, PI -> P2, . . . . Pn -> Q I : THEN I I make multiple AA applications, I I which is guaranteed to succeed I We discuss below t.he training sequence that teaches this macro move. IV. m RECOGNITION @ INSTRUCTIONAL STRATEGIES REDHOT bases its acquisition of application heuristics and macro operators on its recognition of instructional strategies in the training sequence. For example, consider the sequences of exercises in Figure I, taken from the actual curriculum. The sequence, which is at the beginning of the whole curriculum, gives the primitive rule of inference AA, then shows successive elaborations of the use of that rule of inference. REDHOT detects this to be a use of a strategy called focus a& elaborate, in which a rule is first focussed, and then a particular elaboration is given. Teacher: Here is a rule called AA. Teacher: Here are some exercises involving AA: 1. Goal: Q Premises: S -> Q, S 2-5 [Several more exercises with different formulas. 1 6. Goal: W Premises: S -> Q, Q -> W, S 7. Goal: S Premises: R -> S, Q -> W, W -> R, Q 899 [Two more similar exercises involving multiple applications of AA.] Figure _1_ Sequence of Exercises for Learning Multiple Application of AA Command In the above training sequence, REDHOT takes steps l-5 as focussing on the AA rule, and steps 6- 9 as providing a useful elaboration of that rule, in the form of the macro operator for multiple application. 222 A second example of the use of an instructional strategy concerns removing possible bugs in learned heuristics and macro operators. We illustrate this with the macro rule for conditional proof, a common strategy in logic and mathematics, which we paraphrase as follows: I Conditional Proof MACRO MOVE I 1 IF you want to prove P -> Q I I THEN I I ASSUME P as a working premise I I PROVE Q (likely from P); I I APPLY "CP" rule, removing premise I I The actual instructional sequence goes to great length to teach this principle, and students typically have a difficult time with the principle; one defective version of the rule that students seem to learn is the following: I "Defective" MACRO MOVE I I IF you have a formula P -> Q I I AND you want to prove Q I I THEN I I ASSUME P as a working premise; I I PROVE Q using AA; I This is a very poor strategy; but the evidence suggest that over half of the students learn it. The following exercise seems to help students debug the rule "Defective" . Derive: S -> (Q OR R) Premise: (R -> R) -> (Q OR R) In examining student protocols, we see that many students will try several times to apply the "Defective" rule to this exercise. Finally, (we speculate) they realize that (R -> R) is already something that they know how to prove, using a previously learned macro operator. Then, the actual proof becomes obvious, the student corrects the defective rule, and goes on to handle similar exercises correctly. We call this instructional strategy "focus and frustrate" wherein a student discovers --somewhat traumatically--that a rule he learned is defective. Therefore, an exercise such as the above is not just randomly selected, but instead tests possible student "bugs" in an exact way. Notice that it is one of the simplest exercises that will discriminate between the correct and defective formulations of the macro rule for conditionaL proof. (See [2] for a discussion of "debugging" student skills.) to V. REDHOT AND LEARNING SYSTEMS Like many LS's, REDHOT starts wi th the ab ility state everything it will "learn" I in some sense at least. The initial rules for solving the problem (the natural deduction rules for logic) are complete with respect to the underlying problem solver --unless it is restricted in time/space (in practice it is). The heuristic and macro languages are also given in advance, and they of course define a space of the possible rules that might be learned. So, the object is to select among heuristics and macro rules in this space. One way to formulate doing this is by experimentation or exploration. REDHOT selects objects from this meta-space by being taught. Learning by being taught consists of the "teacher" laying out exercises in an organized and structured way, and the student recognizing something of that structure. The student makes-- believes that he is entitled to make--fairly bold hypotheses about the rules he is learning, and relies on the training sequence to contain exercises that will check for common errors that he the student may have made in formulating these rules. REDHOT compares somewhat radically to many LS's that rely on a somewhat slow, computationally coherent delimitation of the rule (or concept) involved. We speculate that learning by "discovery" or "experimentation" is a slow process for even humans, done over the eons of time and through social interaction. Most human learning is by being taught, and one can argue that AI should give attention to the relation between learning and teaching, in terms of modelling the acquisition of concepts, problem-solving skills, and natural language. We further speculate that learning by "discovery" will be aided by extracting as much information as possible from the structure of the environment in which the LS operates. IllI L-21 [31 II41 REFERENCES Amarel, Saul, "An Approach to Problem Solving and Theorem Proving in the Propositional Calculus", in Svstems and Computer Science, (Hart and Takasu, eds.), Toronto: University of Toronto Press, 1967. Brown, John Seely, Burton, Richard R., and Larkin, Kathy M., "Representing and Using Procedural Bugs for Educational Purposes", in Proceedings of the 1977 Annual Conference of the Association for Computing Machinerv, New York, 1977. Suppes, P., Smith, R. L., and Beard, M., "University-level Computer-assisted Instruction at Stanford: 1975", in Instructional Science, 1977, 4, 151-185. Winston, Patrick Henry, "Learning Structural Descriptions from Examples", Ph.D. thesis, in The Psvchology of Commuter Vision, (Patrick Henry Winston, ed.), McGraw-Hill, New York, 1975. 223	1980	17
9	A Computer Model of Child Language Learning Mallory Selfridge Yale University Abstract A computer program modelling a child between the ages of 1 and 2 years is described. This program is based on observations of the knowledge this child had at age 1, the comprehension abilities he had at age 2, and the language experiences he had between these ages. The computer program described begins at the age 1 level, is given similar language experiences, and uses inference and learning rules to acquire comprehension at the age 2 level. Introduction This paper describes a computer model of the development of comprehension abilities in a child, Joshua, between the ages of one and two years. The program begins with the kind of knowledge that Joshua had at age 1, when he under- stood no language, and learns to understand com- mands involving action, object, and spatial rela- tion words at Joshua's age 2 level. It does so by being given the kind of language experiences Joshua had between the ages of 1 and 2, and making use of rules to 1) infer the meaning of utter- ances, 2) attend to words, and 3) learn language meaning and structure. The program passes through a reasonable developmental sequence and makes the same kind of errors that children make at inter- mediate stages. This work suggests that language learning to the 2 year old level can be accounted for primari- ly by the learning of word meaning and structure, that world knowledge is crucial to enable the child to infer the meaning of utterances, and that children hear language in situations which enable them to perform such inferences. The success of the program in modelling Joshua's language development -- both its progression and its errors -- suggests that it embodies a plausible theory of how Joshua learned to understand language. While there are several aspects of the model which are unrealistic (for example, segmented input, no am- biguous words, no simultanious conceptual develop- ment), there is reason to believe that future work can sucessfully address these issues. Further de- tails can be found in Selfridge (1980). This paper first considers Joshua's initial state of knowledge at age 1, and then his comprehension abilities at age 2. It describes the kind of language experiences he had, and several kinds of learning rules which can account for Joshua's development. The computer program incor- porating these observations and rules * described, and finally some conclusions ai: presented. Joshua's Initial Knowledge The first component of --- a computer model of the development of Joshua's comprehension is Joshua's knowledge prior to his language learning. Observations like the follow- ing suggest that Joshua had considerable knowledge of objects, actions, spatial relations, and ges- tures at age 1 (ages are given in YEARS:MONTHS:DAYS): 0:11:19 Joshua and I are in the playroom. I build a few block towers for him to knock down, but he doesn't do so; rather, he dismantles them removing the blocks from the top, one at a time 1:0:16 Joshua and I are in the playroom . Joshua takes a toy cup, and pretends to drink out of it. 1:2 Joshua is sitting in the living room playing with a ball. I hold my hand out to him, and he gives me the ball. The above observations show that Joshua knew the properties and functions of objects like blocks, cups and balls. He knew actions that could be per- formed with them, and various spatial relations that they could enter into. Finally, he knew that behavior can be signaled through gestures by other people. Thus, a language learning program must be equipped with this kind of knowledge. Joshua's Comprehension Abilities at Age2 At age 2, Joshua could respond correctly to commands with unlikely meaning and structure. His correct responses suggests full understanding of them. For example, consider the following: 2:0:5 We walk into the living room and Joshua shows us his slippers. His mother says "Put your slippers on the piano." Joshua picks up the slippers and puts them on the piano keys, looking at his mother. She laughs and says "Thats silly." Joshua removes the slippers. The meaning of this utterance is unlikely since slippers do not generally go on piano keys, and piano keys don't generally have things put on them. His response suggests that he was guided by full understanding of the meanings of the words in "Put your slippers on the piano." At age 2 Joshua also understood language structure, as the following example shows: 2:o:o Joshua and I are in the p layroom, my tape recorder i on the floor in front of me. I say "Get on the tape recorder, Joshua". Joshua looks at me oddly, and looks at the the tape recorder. I repeat "Get on the tape recorder." Joshua moves next to the tape tape recorder. I once more repeat 'Get on the the tape recorder." Joshua watches me intently, and lifts his foot up and slowly moves it over the tape recorder to step on it. I laugh and pull the tape recorder away. It seems that Joshua understood "Get on the tape recorder" the first time I said it, and that his reluctance to comply reflected his knowledge that what I was asking was very unlikely. That is, Joshua understood that the tape recorder was the object to be underneath him, although this is un- likely given his experience with it. This, in turn, suggests that Joshua understood the struc- ture of the word "on", namely, that the word whose meaning is the supporting surface follows "on". Thus a program modelling Joshua at age 2 must understand utterances using language structure. Joshua's Language Experiences In the year between -- the ages of 1 and 2, Joshua experienced situations which allowed him to make inferences concerning the utterances he heard. In this section, three examples of such situations are given, and infer- ence rules accounting for Joshua's response and attention to words are presented. In the first example, I am using an utterance and simultaniously signalling the meaning of that utterance through gestures: 1:2:17 We are sitting in the living room, Joshua is holding a book. I look at Joshua, maintain eye contact for a moment, hold my hand out to him and say "Give me the book, Joshua." Joshua holds the book out to me. In this situation, Joshua probably inferred that the meaning of "Give me the book, Joshua." was the same as that signalled by the gestures. The fol- lowing rule captures this idea: Gestural Meaning Inference If an utterance is accompanied by gestures with associated meanings then infer that the the utterance means the same as the gestures. Knowledge of object function and properties helped Joshua infer responses in other situations. In the following, Joshua used his knowledge that books can be opened in his response: 1:0:9 Joshua has a book in his hand, and is looking at it, turning it over, and examining it. His mother says 'open the book, open the book..." Joshua opens the book. She says, "Good Joshua, good." A rule summarizing this inference is the follow- ing: Function/Property Inference If an utterance is heard while interacting with an object then the meaning of the utterance involves a function or property of that object. Parent speech to children posesses many attention-focussing characteristics (e.g. Newport, 1973). The following example is typical: 1:18:0 Joshua's father is trying to demonstrate that Joshua knows the names of the upstairs rooms, and has put a toy lawnmower in the bathroom. He says "Where is the lawnmower, Josh? Its in the BATHROOM. The LAWNMOWER is in the BATHROOM. BATHROOM!" Joshua's attention to "bathroom" in this example can be explained by the following rule: Attention Inference If a word is emphasised, repeated, or said in isolaytion, then attend to it. These are the kind of rules which I postulate enabled Joshua to infer the meaning of utterances from context, and attend to part of the utterance. The program must be equipped with such rules and must be given input in similar contexts. Learning Rules This section will consider Joshua's learning of action, object, and relation words, and language structure. It presents accounts of how Joshua might have learned each of these. Most of the rules have their roots in the learning strategies proposed by Bruner, Goodnow, and Austin ('956). One way Joshua learned the names of ob- jects is by having them named for him, as in the following example: 1:0:0 Joshua is crying. His mother picks him up and goes over to the refrigerator. She gets some juice, holds it up, and asks, "Do you want some JUICE?" Joshua keeps crying. She gets a banana and asks, "Do you want some BANANA, Joshua?" Joshua reaches for it. The following rule models Joshua's ability to learn by having objects named: Direct Naming Inference If a word and an object are both brought to attention, infer the word is the object's name. This rule, and other object word learning rules, can account for how Joshua learned object words such a "slippers", "piano", "ball", and "table". Action words can be learned via inferences about other known words in the utterance. In the following example, summarized from Schank and Sel- fridge (1977), Hana could have inferred the mean- ing of "put" based on her knowledge of the mean- ings of "finger" and "ear." (age 1) Hana knows the words "finger" and "ear", but not "put." She was asked to "Put your finger in your ear," and she did so. The following two rules can account for learning "Put" in situations like this. The first suggests that "put" would initially be learned as "put something in something else." The second, applied after the first in a slightly different situation, would refine the meaning of "put" to "put some- thing someplace". Response Completion Inference Infer the meaning of an unknown word to be the meaning of the entire utterance with the meanings of the known words factored out. Meaning Refinement Inference If part of the meaning of a word is not part of the meaning of an utterance it occurs in, remove that part from the word's meaning. Rules like the above can account for Joshua learning action words like "put", "bring", "give", and so on. However, they can also account for Joshua learning relation words, such as "on" and “in”. If Joshua knew "put", "ball", and "box", say, and was asked to "put the ball in the box", these rules would account for his learning that "in" referred to the "contained" relation. These, then, are the sort of rules the pro- gram uses to learn word meanings. The program's rule for learning language structure is more direct. It is based around the two structural predicates, PRECEDES and FOLLOWS, which relate the positions of words and concepts in short-term memory. This rule models Joshua's acquisition of structural information upon hearing utterances he understands, and appears below: Structure Learning Rule If a slot filler occurs preceding or fol lowing a word' s meaning then update the word's defini .tion that information. This rule accounts for Joshua learning that the filler of the VAL slot of "in"'s meaning -- (CONT VAL (NIL)) -- is found FOLLOWing "in" in the utterance. The Program This section presents four excerpts from a run of the program, written in LISP on a DECSYSTEM-20. Each represents the program at a different stage in development as it progresses from Joshua's age 1 abilities to Joshua's age 2 abilities, using the inference rules described previously. The knowledge representation used is Conceptual Dependancy (Schank, 19731, and the language understanding process embedded in the program is similar to that in Birnbaum and Sel- fridge (1979). The first stage of the program corresponds to Joshua at age 1. At this stage, the program had only the knowledge ascribed to Joshua at that age. In the excerpt below, the "parent" types a lower- case utterance to the program, and the program responds with a message stating its lack of under- standing. When the parent provides gestures via simulated visual input, however, the program understands, and prints the CD representation of its response. IPARENT SAYS: give me the ball ICHILD STARES BLANKLY AT PARENT ICHILD RETURNS TO PLAY ICHILD sms: (PARENT HOLDS OUT HAND) (PARENT ~0oKs AT BALLS) ICHILD INFERS RESPONSE USING RULE: I GESTURAL MEANING ICHILD RESPONDS: (ATRL~NS ACTOR (CHILD) I OBJECT (BALLS) TO (~0~s VAL (PARENT)) In the second stage, shown in the excerpt below, the program has learned the meaning of several words, and understands some utterances correctly. In this case, it has learned the words "put", "ball", and "box". However, notice that although it responds correctly to the first utter- ance given by the parent, it misunderstands the second. This sort of error is reported in Hoogen- raad et al. (1976). Not knowing "on", the program incorrectly infers that the appropropriate rela- tionship is containment. IPARENT SAYS: put the ball in the box I ICHILD 11wERs RESPONSE USING RULE: I UTTERANCE UNDERSTANDING, FUNCTION/PROPERTY ICHILD RESPONDS: (PTRANS ACTOR (CHILD) I OBJECT (BALLS) TO (CONT VAL (BOXl)) I PARENT SAYS: put the ball on the box I iCHILD INFERS RESPONSE USING RULE: I UTTERANCE UNDERSTANDING, FUNCTION/PROPERTY ICHILD RESPONDS: (PTRANS ACTOR (CHILD) OBJECT (BALLS) TO (CONT VAL &0X1)) The transition from the second stage to the third is accomplished by teaching the program more words. In this case it has learned the additional words "slippers", "on", "piano", “ball”, and “table. ” At this stage, the program can now understand "Put the slippers on the piano", whereas at any earlier stage it would not have. The program also prints out a message showing that it recognizes this as an unusual request. However, although this stage represents Joshua's age 2 understanding of word meaning, the program has not yet learned language structure. The program interprets the second utterance in- correctly, however, in accord with its knowledge of the usual relationships between objects. This sort of error is similar to that reported in Stroher and Nelson (1974). IPARENT SAYS: put the slippers on the piano ICHILD LOOKS AT PARENT STRANGELY I CHILD INFERS RESPONSE USING RULE: I UTTERANCE UNDERSTANDING ICHILD RESPONDS: I (PTRANS ACTOR (CHILD) I OBJECT (SLIPPERS~) TO (TOP VAL (~1~~01))) IPARENT SAYS: put the table on the ball I iCHILD INFERS RESPONSE USING RULE: I UTTERANCE UNDERSTANDING ICHILD RESPONDS: (PTRANS ACTOR (CHILD) I OBJECT (BALLS) TO (TOP VAL (TABLET))) The fourth stage is shown in the excerpt below. The program has now learned the structure of "on", and can hence correctly understand "Put the table on the ball." In addition, it prints out 226 a message indicating its liarity of this reque st. awareness of IPARENT SAYS: put the table on the ball ~CHILD LAUGHS AT UNUSU~~L REQUEST ICHILD INFERS RESPONSE USING RULE: I UTTERANCE UNDERSTANDING ICHILD RESPONDS: (PTRANS ACTOR (CHILD) I OBJECT (TABLET) TO (TOP VAL (BALLS))) the pecu- At the four th stage, the program has success- fully learned to understand a subset of language at Joshua's age 2 level. It began with world knowledge similar to that Joshua began with, was equipped with reasonable learning and inference rules, and progressed as he did by being given language experiences similar to those he experi- enced. Conclusions This paper has described a computer model of a child learning to understand commands involving action, object, and relation words. The program learns language meaning and structure to the level attained by the child at age 2, by being initially given the same kind of knowledge the child had and by being exposed to language in the same kind of contexts as the child did. gram learned language according to a The pro- reasonable progress ion, making the same sort of error s that children do a t intermediate stages. No par ts of speech or traditional grammatical constructions are learned. It also acquires structural knowledge after knowledge of meaning, because no structural knowledge can be-associated with a word until the meaning of that word is learned. This This aspect of the model offers an explanation for why children learn structure following meaning (Wetstone and Friedlander, 1973). In addition to English, the program has been tested on comparable subsets of Japanese, Russian, Chinese, Hebrew, and Spanish. Its performance with these languages was equivalent to its learning of English, suggesting that the program has no English-specific mechan- isms. This research suggests several conclusions. It suggests that a large part of the language learning problem lies in accounting for how the child infers the meaning of the language he hears. It argues that the mechanisms underlying the learning of meaning and structure are the same. It questions the role of traditional grammatical models both in language learning and language understanding, and suggests that models of language learning must be based on strong models of language understanding. In particular, it ques- tions Chomsky's (1980) position that language is not learned. This work suggests that plausible learning models of language development are possi- ble. Further research should proceed in many directions. In particular, the program discussed here should be extended to model the development of comprehension of more complex constructions, such as relative clauses, and the generation of language. Acknowledgements Dr. Roger Schank's assistance in this work was in- valuable. Peter Selfridge provided useful com- ments on this paper. Bibliography Birnbaum, L., and Selfridge, M. (1979). Prob- lems in Conceptual Analysis of Natural Language. Research Report 168, Department of Computer Sci- ence, Yale University. Bruner, J.S., Goodnow, J. J., and Austin, G.A., (1956). A Study of Thinking. John Wiley and Sons, New York - Chomsky, N., (1980). Rules and Representations, excerpted from Rules and Representations. Colum- -- bia University Press, New York Hoodenraad, R., Grieve, R., Baldwin, P., and Campbell, R. (1976). Comprehension as an Interac- tive Process. In R. N. Campell and P. T. Smith (eds.) Recent Advances in the Psychology of Language., Plenum Press, NeFYor - Newport, E.L., (1976). Motherese: the Speech of Mothers to Young Children. in N.J. Castellan, D.B. Pisoni, and G.R. Potts, (eds.) Cognitive Theory: VI II., Lawrence Erlbaum Assoc., Hilsdale, N.J. Schank, R. C., (1973). Identification of Con- ceptualizations Underlying Natural Language. In R. C. Schank and K. M. Colby (eds.) Computer Models of Thought and Language W.H. Freeman and Co., San Fransisco. Schank, R. C., and Selfridge, M. (1977). How to Learn/What to Learn. in Proceedings of the International Joint Conference on Artificial In- telligence, Cambridge, Mass. Selfridge, M. (1980). A Process Model of Language Acquisition. Computer Science Technical Report 172, Yale University, New Haven, Ct. Strohrer, H. and Nelson, K.E., (1974). The Young Child's Development of Sentence Comprehen- sion: Influence of Event Probability. Non-verbal . . Context, Syntactic Form, and Strategies. Child m., 45: 567-576 Westone, H. and Friedlander, (1973). The Ef- fect of Word Order on Young Children's Responses to Simple Questions and Commands. Child Dev 44:734-740 - -0 227	1980	18
10	APPROACHES TO KNOWLEDGE ACQUISITION: THE INSTRUCTABLE PROQUCTION SYSTEM PROJECT Michael D. Rychener Carnegie-Mellon University Department of Computer Science Schenley Park Pittsburgh, PA 15213 Abstract Progress in building systems that acquire knowledge from a variety of sources depends on determining certain functional requirements and ways for them to be met. Experiments have been performed with learning systems having a variety of functional components. The results of these experiments have brought to light deficiencies of various sorts, in systems with various degrees of effectiveness. The components considered here are: interaction language; organization of procedural elements; explanation of system behavior; accommodation to new knowledge; connection of goals with system capabilities; reformulation (mapping) of knowledge; evaluation of behavior; and compilation to achieve efficiency and automaticity. A number of approaches to knowledge acquisition tried within the Instructable Production System (IPS) Project are sketched.* 1. The lnst ructable Production System Project The IPS project [6] attempts to build a knowledge acquisition system tinder a number of constraints. The instructor of the system gains all information about IPS by observing its interactions with its environment (including the instructor). Interaction is to take place in (restricted) natural language. The interaction is mixed initiative, with both participants free to try to influence the direction. Instruction may be about any topic or phenomenon in the system’s external or internal environment. Knowledge accumulates over the lifetime of the system. This research was sponsored by the Defense Advanced Research Projects Agency (DOD), ARPA Order No. 3597, monitored by the Air Force Avionics Laboratory Under Contract F33615-78-C-1551. The views and conclusions contained in this document ate those of the author and should not be interpreted as representing the official polictes, either expressed or implied, of the. Defense Advanced Research Projects Agency or the US Government. Throughout these IPS experiments, the underlying knowledge organization has been Production Systems (PSs) [2], a form of rule-based system in which learning is formulated as the addition to, and modification of, an unstructured collection of production rules. Behavior is obtained through a simple recognize-act cycle with a sophisticated set of principles for resolving conflicts among rules. The dynamic short-term memory of the system is the Working Memory (WM), whose contents are matched each cycle to the conditions of rules in the long-term memory, Production Memory. Study of seven major attempts to construct instructable PSs with various orientations leads to recognizing the centrality of eight functional components. Listing the components and their embodiment in various versions of IPS can contribute to research on learning systems in general, by clarifying some of the important subproblems. This discussion is the first overview of the work of the project to date, and indicates its evolutionary development. Members of the IPS project are no longer working together intensively to build an instructable F ‘S, but individual studies that will add to our knowledge about one or more of these components are continuing. Progress on the problem of efficiency of PSs has been important to the IPS project [3], but will not be discussed further here. 2. Essential Functional Components of Inst ructable Systems The components listed in this section are to be interpreted loosely as dimensions along which learning systems might vary. Interaction The content and form of communications -A between instructor and IPS can have a lot to do with ease and effectiveness of instruction. In particular, it is important to know how closely communications correspond to internal IPS structures. Similarly, we must ask how well the manifest behavior of IPS indicates its progress on a task. An IPS can have various orientations towards interactions, ranging from passive to active, with maintenance of consistency and assimilation into existing structures. Organization. Each version of IPS approaches the issue of obtaining coherent behavior by adopting some form of organization of its ‘procedural’ knowledge. This may involve such techniques as collecting sets of rules into ‘methods’ and using signal conventions for sequencing. Whether IPS can explain its 228 static organization and whether the instructor of procedural control are important subissues. can see the details Explanation. A key operation in an instructable system is that of explaining how the system has arrived at some behavior, be it correct or incorrect. In the case of wrong behavior, IPS must reveal enough of its processing to allow the more intelligent instructor to determine what knowledge IPS lacks. Accommodation. Once corrections to IPS’s knowledge have been formulated by the instructor, it remains for further interactions with IPS to augment or modify itself. In the IPS framework, these modifications are taken to be changes to the rules of the system, rather than changes to the less permanent WM. As with interaction, IPS can take a passive or active approach to this process. Connection. Manifest errors are not the only way a system indicates a need for instruction: inability to connect a current problem with existing knowledge that might help in solving it is perhaps a more fundamental one. An IPS needs ways to assimilate problems into an existing knowledge framework, and ways to recognize the applicability of, and discriminate among, existing methods. Reformulation. Another way that IPS can avoid requiring instruction is for it to reformulate existing knowledge to apply in new circumstances. There are two aspects to this function: finding knowledge that is potentially suitable for mapping, and performing the actual mapping. In contrast to connection, this component involves transformation of knowledge in rules, either permanently or dynamically. Evaluation. Since the instructor has limited access to what IPS is doing, it is important for IPS to be able to evaluate its own progress, recognizing deficiencies and errors ai they occur so that instruction can take place as closely as possible to the dynamic point of error. Defining what progress is and formulating relevant questions to ask to fill gaps in knowledge are two subissues. Compilation. Rules initially formed as a result of the instructor’s input may be amenable to refinements that improve IPS’s efficiency. This follows from several factors: during instruction, IPS may be engaged in search or other ‘interpretive’ execution&; instruction may provide IPS with fragments that can only be assembled into efficient form later; and IPS may form rules that are either too general or too specific. Improvement with practice is the psychological analog of this capability. Anderson et al [l] have formulated several approaches to compilation. 3. Survey of Approaches Kernell, ANA, Kernel2 and IPMSL have been fully implemented. The others were suspended at various earlier stages of development, for reasons that were rarely related to substantive or conceptual difficulties. Kernel Version 1 -ZL The starting point for IPS is the adoption of a pure means-ends strategy: given explicit goals, rules are the means to reducing or solving them. Four classes of rules are distinguished: means rules; recognizers of success; recognizers of failure; and evocation of goals from goal-free data. The Kernel1 [6] approach further organizes rules into methods, which group together (via patterns for the same goal) a number of means, tests and failure rules. Interaction consists of language strings that roughly correspond to these methods and to system goals (among which are queries). Keywords in the language give rise to the method sequencing tags and also serve to classify and bound rules. Explanation derives from the piecing together of various goals in WM, along with associated data. The major burden of putting together raw data that may be sufficient for explanation rests on the instructor, a serious weakness. Additive Successive Aouroximations [ASA). Some of the drawbacks of I<ernell can be remedied by orienting instruction towards fragments of methods that can be more readily refined at later times. Interaction consists of having the instructor point at items in IPS’s environment (especially WM) in four ways: condition (for data to be tested), action (for appropriate operators), relevant (for essential data items), and entity (to create a symbol for a new knowledge expression). These designations result in methods that are very loose collections of rules, each of which contributes some small amount towards achievement of the goal. Accommodation is done as post-modification of an existing method in its dynamic execution context, through ten method-modification methods. Analoclv. A concerted attempt to deal with issues of connection and accommodation is represented by McDermott’s ANA program [4]. ANA starts out with the ability to solve a few very specific problems, and attacks subsequent similar problems by using the methods it has analogically. The starting methods are hand-coded. Connection of a new goal to an existing method takes place via special method description rules that are designed to respond to the full class of goals that appear possible for a method to deal with by analogy. An analogy is set up by finding paths through a semantic network containing known objects and actlons. As a method operating by analogy executes, rules recognize points where an analogy breaks down. Then general analogy methods are able either to patch the method directly with specific mappings or to query the instructor for new means-ends rules. Each attempt to build an IPS has been based on the idea of an initial hand-coded kernel system, with enough structure in it to support all further growth by instruction, A kernel establishes the internal representations and the overall approach to instruction. The following are presented in roughly chronological order. These ideas were introduced by A. Newell in October, 1977. 229 Problem Soaces. Problem spaces [5] provide a novel basis for IPS by embedding all behavior and interactions in search. A problem space consists of a collection of knowledge elements that compose states in a space, plus a collection of operators that produce new states from known ones. A problem consists of an initial state, a desired state, and possibly path constraints. Newell’s Problem Space Hypothesis (ibid.) claims that all goal-oriented cognitive activity occurs in a problem space, not just activity that is sufficiently problematical. Interaction consists of giving IPS problems and search control knowledge (hints as to how to search specific spaces). Every Kernel component must be a problem space too, and thus subject to the same modification processes. The concrete proposal as it now stands concentrates on interaction, explanation (which involves sources of knowledge about the present state of the search), and organization. Schemas. The use of schemas as a basis for an IPS kerner* make slot-filling the primary information-gathering operation. A slot is implemented as a set of rules. The slots are: executable method; test of completion; assimilation (connects present WM with the schema for a goal); initialization (gathers operands for a method); model (records the instruction episode for later reference); accommodation (records patches to the method); status (records gaps in the knowledge); monitoring (allows careful execution); and organization (records method structure). Orientation towards instruction is active, as in ASA. Explanation consists of interpreting the model slot, and accommodation, of fitting additions into the model. Connection is via a discrimination network composed of the aggregated assimilation slots of all schemas. Compilation is needed here, to map model to method. Kernel Version 2. An approach with basic ideas similar to ASA and to Waterman’s Exemplary Programming [8], Kernel2 [7] focusses on the process of IPS interacting with the instructor to build rules in a dynamic execution context. The instructor essentially steps through the process of achieving a goal, with IPS noting what is done and marking elements for inclusion in the rules to be built when the goal is achieved. Kernel2 includes a semantic network of information about its methods, for use as a ‘help’ facility. Kernel2 is the basis from which the IPMSL system, below, is built. Semantic Network. Viewing accumulation of knowledge as additions to a semantic network is the approach taken by the IPMSL system [7]. Interaction consists of definition and modification of nodes in a net, where such nodes are represented completely as rules. Display and net search facilities are provided as aids to explanation and accommodation. The availability of traditional semantic network inferences makes it possible for IPMSL to develop an approach to connection and reformulation, since they provide a set of tools for relating and mapping knowledge into more tractable expressions. This 1978. approach was formulated by A. Newell and J. Laird in October of 4. Conclusions The IPS project has not yet succeeded in combining effective versions of components as discussed above, to produce an effective IPS. The components as presently understood and developed, in fact, probably fall short of complete adequacy for such a system. But we have explored and developed a number of approaches to instructability, an exploration that has added to the stock of techniques for exploiting the advantages of PSs. We are encouraged by the ability of the basic PS architecture to enable explorations in a variety of directions and to assume a variety of representations and organizations. Acknowledqments. Much of the work sketched has been done jointly over the course of several years. Other project members are (in approximate order of duration of commitment to it): Allen Newell, John McDermott, Charles L. Forgy, Kamesh Ramakrishna, Pat Langley, Paul Rosenbloom, and John Laird. Helpful comments on this paper were made by Allen Newell, Jaime Carbonell, David Neves and Robert Akscyn. References 1. Anderson, J. R., Kline, P. J., and Beasley, C. M. Jr. A Theory of the Acquisition of Cognitive Skills. Tech. Rept. 77-1, Yale University, Dept. of Psychology, January, 1978. 2. Forgy, C. L. OPS4 User’s Manual. Tech. Rept. CMU-CS-79-132, Carnegie-Mellon University, Dept. of Computer Science, July, 1979. 3. Forgy, C. L. On the Efficient Implementation of Production Systems. Ph.D. Th., Carnegie-Mellon University, Dept. of Computer Science, February 1979. 4. McDermott, J. ANA: An Assimilating and Accommodating Production System. Tech. Rept. CMU-CS-78-156, Carnegie-Mellon University, Dept. of Computer Science, December, 1978. Also appeared in IJCAI-79 5. Newell, A. Reasoning, problem solving and decision processes: the problem space as a fundamental category. In Attention and Performance VIII, Nickerson, R., Ed.,Erlbaum, Hillsdale, NJ, 1980. 6. Rychener, M. D. and Newell, A. An instructable production system: basic design issues. In Pattern-Directed Inference Systems, Waterman, D. A. and Hayes-Roth, F., Eds., Academic, New York, NY, 1978, pp. 135-153. 7. Rychener, M. D. A Semantic Network of Production Rules in a System for Describing Computer Structures. Tech. Rept. CMU-CS-79-130, Carnegie-Mellon University, Dept. of Computer Science, June, 1979. Also appeared in IJCAI-79 8. Waterman, D. A. Rule-Directed Interactive Transaction Agents: An Approach to Knowledge Acquisition. Tech. Rept. R-21 71 -ARPA, The Rand Corp., February, 1978. dSchemas were first proposed for IPS by Hychener, May, 1978 230	1980	19
11	SOME ALGORITHM DESIGN METHODS Steve Tappel Systems Control, Inc., 1801 Page Mill Road Palo Alto, California 94304 and Computer Science Department, Stanford University Abstract Algorithm design may be defined as the task of finding an efficient data and control structure that implements a given input- output specification. This paper describes a methodology for control structure design, applicable to combinatorial algorithms involving search or minimization. The methodology includes an abstract process representation based on generators, constraints, mappings and orderings, and a set of plans and transformations by which to obtain an efficient algorithm. As an example, the derivation of a shortest-path algorithm is shown, The methods have been developed with automatic programming systems in mind, but should also be useful to human programmers. 1. introduction The general goal of automatic programming research is to find methods for constructing efficient implementations of high- level program specifications. (Conventional compilers embody such methods to a very limited extent.) This paper describes some methods for the design of efficient control structures, within a stepwise refinement paradigm.. in stepwise refinement (see for instance Cl,Zl), we view the program specification itself as an algorithm, albeit a very inefficient one. Through a chain of transformation steps, we seek to obtain an efficient algorithm. Specification -+ Alg -+ . . . 3 Algorithm Each transformation step preserves input-output equivalence, so the final algorithm requires no additlonal verification. Algorithm design is a difficult artificial intelligence task involving representation and planning issues. First, in reasoning about a complicated object like an algorithm it is essential to divide it into parts that interact in relatively simple ways. We have chosen asynchronous processes, communicating via data channels, as an appropriate representation for algorithms. Details are in Section II. Second, to avoid blind search each design step must be clearly motivated, which in practice requires organization of the transformation sequence according to high-level plans. An outline of the plans and transformations we have developed is given .in Section III, followed in Section IV by the sample derivation of a shortest path algorithm. Sections V and VI discuss extensions and conclude. This methodology is intended for eventual implementation within the CHI program synthesis system 131, which is under development at Systems Control Inc. This research is supported in part by the Defense Advanced Research Projects Agency under DARPA Order 3687, Contract N00014-79-C-0127, which is monitored by the Office of Naval Research. The views and conclusions contained in this paper are fhe of the author and should not be interpreted as necessarily representing the oflicial policies, either expressed or implied, of Xl, DARPA, ONR or the US Government. II. Process graph representation of algorithms Our choice of representation is motivated largely by our concentration on the earlier phases of algorithm design, in which global restructurings of the algorithm take place. Most data structure decisions can be safely left for a later phase, so we consider only simple, abstract data types like sets, sequences and relations. More importantly, we observe that conventional high- level languages impose a linear order on computations which is irrelevant to the structure of many algorithms and in other cases forces a premature committment to a particular algorithm. To avoid this problem, we have chosen a communicating process representation in which each process is a node in a directed graph and processes communicate by sending data items along the edges which act as FIFO queues. Cycles are common and correspond to loops in a conventional language. The use of generators (or producers) in algorithm design was suggested by 141. Our representation is essentially a specialized version of the language for process networks described in I51 Rather than strive for a general programming language we use only a small set of process types, chosen so that: (I) the specifications and algorithms we wish to deal with are compactly represented, and (2) plans and transformations can be expressed in terms of adding, deleting or moving process nodes. The four process types are: Generator: produces elements one by one on its output edge. Constraint: acts as a filter; elements that satisfy the constraint pass through. Mapping: takes each input element and produces some function of it. If the function value is a set its elements are produced one by one. Ordering: permutes its input elements and produces them in the specified order. The representation is recursive, a very important property. There can be generators of constraints, constraints on constraints, mappings producing generators, etc. Most of the same design methods will apply to these “meta-processes”. To illustrate the representation, we encode the specification for our sample problem of finding the shortest path from a to b in a graph. Idotation and terminology for shortest path. A directed graph is defined by a finite vertex set V and a binary relation Edge(u,v). A path p is a sequence of vertices (pi . . . p,), in which Edge(pi.pbr) holds for each pair. The “.” operator is used to construct sequences: (u . . . v).w = (u . . . v w). Every edge of the graph is labelled with a positive weight W(u,v) and the weight of an entire path is then Weight(p) = W(pi,pz)+...+W(p,l,p,,). The shortest path from a to b is just the one that minimizes Weight. 64 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. A specification should be as simple as possible to ensure correctness. Shortest path can be simply specified as: generate ail paths from a to b, and select the shortest. We express selection of the shortest path in a rather odd way, feeding all the paths into an ordering process whose very first output will cause the algorithm to stop. The point is that by using a full ordering for this comparatively minor task, we can apply all the plans and transformations for orderings. As for the paths from a to b, they are defined as a certain kind of sequence of vertices, so we introduce a generator of all vertex sequences and place constraints after it to eliminate non-paths. This completes the specification. Constraint methods. The goal of constraint methods is to reduce the number of elements generated. The top level plan for constraints says to: I. propagate constraints through the process graph to bring them adjacent to a generator, 2. incorporate constraints into a generator whenever possible, and if the results are not satisfactory, 3. deduce new constraints beyond those given in the specification, and repeat. Each of the three subtasks is nontrivial in itself and is carried out according to a set of (roughly speaking) intermediate- level methods. For (2) an intermediate-level method that we use several times in the Shortest Path derivation is: Selection of an appropriate internal structure for the generator of Sequences(V) is actually part of the design process, but to simplify the example we will take as a default the usual recursive definition of sequences. The recursion in the definition corresponds to a cycle in the process graph. The constraint incorporation plan ConstrainComponent. ConstrainComponent applies when a constraint on composite objects x (sets, sequences, not numbers) is reducible to a constraint on a single component c of x, i.e. P(x) e P’(x,). ConstrainComponent then gives a plan: I. Inside the generator, find the sub-generator of values for component c. If necessary, manipulate the process graph to isolate this generator. Again, other methods must be called upon. 2. Remove constraint P and add constraint P’ to the sub- generator. 0 csequences (VI e Ordering methods. Another group of methods is concerned with the deduction, propagation and incorporation of orderings i f scSe 4 uences(VI I on a generated set. These methods are analogous to the methods r I and vc for constraints but more complicated. In the Shortest Path then s. vcsequences (VI Map: is.v 1 WV) derivation we use a powerful transformation, explained rather A sketchily here: The generation process starts when the empty sequence () is produced on the 3” edge. From the “s” edge it goes to the constraint and also to the mapping, which produces the set of all one-vertex sequences ().v, for vcV. These are fed back to generate two-vertex sequences, and so on. A mapping cycle like this is a very common kind of generator. 111. hods for algorithm design The program specification from which design starts is typically written as an exhaustive generate-and-test (or generate- and-minimize) process, and bears little resemblance to the algorithm it will become. The design methods all have the goal of incorporating constraints, orderings or mappings into the generator, or else the goal of planning or preparing to do so. To incorporate a constraint means to modify the generator so that it only generates items which already satisfy the constraint; to incorporate an ordering means to modify the generator so it generates elements directly in that order; and to incorporate a mapping f means to generate elements f(x) instead of elements x. Accordingly, the methods fall into three main classes, briefly described below. Sunerimnosed upon this class division is a heirarchy (not strict) &th multi-step’plans at the higher levels and a large number of specific syntactic transformations at the bottom. The heirarchy is organized according to goals and subgoals. Heuristics and deduction rules are required to support the planning activity. At the time of writing, a total of about 20 methods have been formulated not counting low-level syntactic transformations. The ordering incorporation transformation InterleaveOrder. InterleaveOrder applies when an ordering R is adjacent to a generator consisting of a mapping cycle, in which the mapping f has the property R(x f(x)) for all x. In other words, f(x) is greater than x under the ordering R. InterleaveOrder moves the ordering inside the mapping cycle and adds a synchronization signal to make the ordering and mapping operate as coroutines. The ordering produces an element x, the mapping receives it and produces its successors f(x) (there would be no need for the ordering at all if f were single-valued), then the ordering produces the next element and so on. Mapping methods. The methods for incorporating a mapping into a generator are mostly based upon recent work in the “formal differentiation of algorithms” [6] and are related to the well-known optimizing technique of reduction in operator strength. (They are not used in our sample design.) Some syntactic transformations and other methods described in this section will appear in the derivation. Example: Design of a shortest path ~~go~~t~rn not In the design which follows, the specification will be transformed from an inefficient generate-and-minimize scheme into a dynamic programming algorithm. The final algorithm grows paths out from vertex a, extending only the shortest path to each intermediate vertex, until reaching b. Of necessity we omit many details of the design. IV. 1. Constraint methods. Since the specification’s constraints are already next to the generator (step i), the overall plan for constraints says to try to incorporate them (step 2.) We will follow the heuristic of incorporating the strongest constraint first. Right now, the algorithm reads Incorporate the Edge constraint. More detail will be shown in this first step than in later derivation steps. ConstrainComponent applies because once a vertex Si has been added to a skquence, -the constraint Edge(Si,Sk,) reduces to a constraint on the single component Si+lm (This reasoning step is really the application of another method, not described here.) Step (I) in the Co ns raincomponent t plan says to find the generator of values for components Si+l* Though we have written it in linear form for convenience, the expression (s.v 1 vcV} is really a generator followed by a mapping. Unfortunately “vcV” generates SI as well as the desired siri values, so we have to unroll one cycle of the graph to isolate the generator of Si.1 values. (Agaln, we have applied methods not described in this paper.) Step (2) is now possible and consists in constraining v to satisfy Edge(s,,v). With the Edge constraint inzorporated, only paths are now being generated so we change s to p in the diagram. I Hap: (p.v 1 vcV A Edge(pn,vll I Incorporate the constraiut that pl=a. Since the pl=a constraint refers only to a component of p, ConstrainComponent applies again. We constrain v in the first “vcV” generator to be equal to a. After simplifying, we obtain Incorporate the constraint that pn=b. Once again ConstrainComponent applies. This time, however, we are unable to isolate a generator for the last vertex of paths. The last vertex of one path is the next-to-last vertex of another, and so on. ConstrainComponent fails, other methods fail too; we leave the pn=b constraint unincorporated. Deduce new constraint. In accordance with the general constraint plan (step 3) we now try to deduce more constraints. One method for deducing new constraints asks: do certain of the generated elements have no eflecf whatroever upon the result of the algorithm? If the answer is “yes”, try to find a predicate that is false on the useless elements, true on others. Motive: if we later succeed in incorporating this constraint into the generator, the useless elements will never be produced. NOW consider the Order + STOP combination. Because all it does is select the shortest path, any path which is not shortest will have no effect! The corresponding constraint says: p is a shortest path from a to b. A further deduction gives the even stronger constraint that every subpalh of p must be a shortest path (between its endpoints). Incorporation of this constraint is complex and is deferred till after incorporation of the Weight ordering. IV.2. Ordering methods. So far paths are generated according to the partial order of path inclusion; path p is generated before path q if q = p.u . . .V for some vertices u,..,v. We may generate a lot of paths to b before generating the shortest one - possibly an infinite number. However if the Weight ordering can be incorporated into the path generator, then only a single path to b (the shortest one) will ever be generated. Propagate Ordering. Before applying an incorporation method we need to bring the Weight ordering next to the generator. Constraints and orderings commute so this is easy. Incorporate the ordering into the generator. The InterleaveOrder method applies, because Weight(p.v) is greater than Weight(p). It moves the ordering from outside the generator cycle to inside and also causes the ordering to wait for the mapping to finish extending the previous path before it produces another. Incorporate new constraint. The “p is a shortest path” constraint is readily incorporated now: the shortest path to any vertex will be the jirsi path to that vertex. Any later path q, with the same last vertex q,=p,, can be eliminated by a new constraint C(p) = xq. q,=p,. We introduce a mapping to produce these new constraints C(p), and now we have a generator of consfrainfs. The result of the last three steps is The algorithm is now a breadth-first search for a path to b, with elimination of non-optimal paths at every vertex. Despite various inefficiencies that remain, the essential structure of a dynamic programming algorithm is present. One interesting improvement comes from incorporating the generated constraints C(p) Into the generator of paths, using ConstrainComponent. To complete the derivation would require data structure selection and finally a translation into some conventional programming language. 66 V. Other results and limitations Besides the Shortest Path algorithm shown here (and variants of it) the algorithm design methods have been used [71 to derive a simple maximum finding algorithm and several variants on prime finding including the Sieve of Eratosthenes and a more sophisticated linear-time algorithm. In these additional derivations, no new process types and only a few new methods had to be used. Single and multiple processor implementations have informally been obtained from process graph algorithms, for both prime finding and Shortest Path. More algorithms need to be tried before specific claims about generality can be made. The intended domain of application is combinatorial algorithms, especially those naturally specified as an exhaustive search (possibly over an infinite space) for objects meeting some stated criteria, which can include being minimal with respect to a defined ordering. Backtrack algorithms, sieves, many graph algorithms and others are of this kind 181. The methods described here are quite narrow in the sense that a practical automatic programming system would have to combine them with knowledge of: 1. Standard generators for different kinds of objects. Our methods can only modify an existing generator, not design one. 2. Data structure selection and basic operations such as searching a list. 3. Efficiency analysis to determine if an incorporation really gives a speedup. 4. Domain specific facts, e.g., about divisibility if designing a prime finding algorithm. 5. How to carry out the final mapping of process graph into a conventional programming (or multiprogramming) language. VI. Discussion and Conclusions The main lesson to be learned from this work is the importance of using an abstract and modular representation of programs during algorithm design. Details of data structure, low- level operations and computation sequencing should be avoided, if possible, until the basic algorithm has been obtained. (Since some algorithms depend crucial)y upon a well-chosen data structure, this will not always be possible.) Further, it is advantageous to represent algorithms in terms of a small number of standard kinds of process, for which a relatively large number of design methods will exist. The results so far indicate that just four standard processes suffice to encode a moderate range of different specifications and algorithms. Presumably others will be required as the range is extended, and it is an important question whether (or how long) the number can be kept small. A similar question can be asked about the design methods. One would not expect methods based upon such general constructs as generators, constraints, orderings and mappings to have much power for the derivation of worthwhile algorithms. For instance, if we had explicitly invoked the idea of dynamic programming, our derivation of a shortest path algorithm would have been shorter. For really difficult algorithms, the general methods may be of little use by themselves. We suggest that they should still serve as a useful complement to more specific methods, by finding speedups (based on incorporation of whatever constraints, orderings and mappings may be present) in an algorithm obtained by the specific methods. As a final issue, it is interesting to speculate how the stepwise refinement approach to programming might be used by human programmers. Use of a standard set of process types and correctness-preserving transformations would be analogous to the formal manipulations one performs in solving integrals or other equations. If that were too restrictive, perhaps one could use the methods as a guide, without attempting to maintain strict correctness. After obtaining a good algorithm, one could review and complete the design, checking correctness of each transformation step. The result would be a formally correct but also well-motivated derivation. Acknowledgements. Many helpful ideas and criticisms were provided by Cordell Green, Elaine Kant, Jorge Phillips, Bernard Mont-Reynaud, Steve Westfold, Tom Pressburger and Sue Angebranndt. Thanks also to Bob Floyd for sharing his insights on algorithm design. References I. Baiter, Robert; Goldman, Neil: and Wile. David. “On the transformational implementation approach to programming”, Proc. 2nd Int’l Conference on Software Engineering (1976) 337-344. 2 3 4. 5 6, 7. 8. 9. Barstow. David R. Knowledge Based Program Construction, Elsevier North-Holland, New York, 1979. Phillips, Jorge and Green, Cordell. “Towards Self-Described Programming Environments”, Technical Report, Computer Science Dept., Systems Control, Inc., Palo Alto, California, April 1980. Green, Cordell and Barstow, David R. “On Program Synthesis Knowledge”, Arti.cial Intelligence, 103 ( 1978) 24 I - 279. Kahn, Gilles and MacQueen, David B. “Coroutines and Networks of Parallel Processes”, lnformafion Processing 77, IFIP, North-Holland Publishing Company, Amsterdam, ( 1979) 993-998. Paige, Robert. “Expression Continuity and the Formal Differentiation of Algorithms”, Courant Computer Science Report x5, (1979) 269-658. Tappel, Steve. “Algorithm Design: a representation and methodology for control structure synthesis”, Technical Report, Computer Science Dept., Systems Control, Inc., Palo Alto, California, August 1980. Reingold. Edward M., Nievergelt. Jurg, and‘ Deo, Narsingh. Combinatorial Algorithms: Theory and Practice, Prentice-Hall Inc., Englewood Chffs, New Jersey, 1977. Elschlager, Robert and Phillips, Jorge. “Automatic Programming” (a section of the Handbook of Artificial Intelligence, edited by Avron Barr and Edward A. Feigenbaum), Stanford Computer Science Dept., Technical - --_ _--_ Report 758, 1979. 10. Floyd, R. W. “The Paradigms of Programming” (Turing Award Lecture), CACM 22:s (1979) 455-460. 67	1980	2
12	USING A MATCHER TO MAKE AN EXPERT CONSULTATION SYSTEM BEHAVE INTELLIGENTLY* Rene' Reboh Artificial Intelligence Center SRI International, Menlo Park, CA 94025 ABSTRACT This paper describes how even a simple matcher, if it can detect simple relationships between statements in the knowledge base, will support many features that will make a knowledge- based consultation system appear to behave more intelligently. We describe three features that are useful during the knowledge acquisition phase (involving the building and testing of the knowledge base), and four features that are of assistance during a consultation. Although these features are described in terms of the Prospector environment [2], it will be clear to the reader how these features can be transferred to other environments. I INTRODUCTION Partial-matching (also referred to as interference-matching, correspondence-mapping, . ..) touches many issues representation and efficiency in many AI syZ:ems [ 31. Its role in Prospector is significant because it is involved in many aspects of consultation and knowledge acquisition. Given any two statements Sl and S2 of the knowledge base, the Semantic Network Matcher of Prospector determines which of the following situations applies: Sl and S2 are identical (Sl = S2) Sl is a restriction of S2 (SlC S2), (or S2 is a generalization of Sl (s2 3 Sl)) Sl and S2 are disjoint statements (Sins2 = $ > Sl overlaps S2 (otherwise) For instance, suppose the knowledge base contains the following statements: Sl : "rhyolite is present" s2: "a rhyolite plug is present" s3: I(an igneous intrusive is present" s4: "rhyolite or dacite is present" s5: "pyrite is present" As these statements are being added to the knowledge base, the Matcher will conclude that: S2 is a restriction of Sl, S2 is a restriction of S3 (rhyolite is an igneous rock and a plug is a special kind of intrusive), Sl and S3 overlap (rhyolite is an igneous rock, but need not be an intrusive), Sl is a restriction of S4, S2 is a restriction of S4 (transitivity from the first and third conclusion), S3 and S4 overlap, S5 is disjoint from Sl, S2, S3 and S4. A detailed description of how the Matcher operates in the Prospector environment can be found in [6]. We mention briefly here that the Matcher views each statement as a set of constraints corresponding to a set of assertions about the existence of physical entities or processes and their attributes. semantic networks [:i l5-;spez;;;, partitioned to represent statements in the knowledge base whereby these assertions are expressed in terms of relations and entries in taxonomies of the domain of application (in this case geology). Let us examine some of the features of a knowledge-based system that can be supported by such a Matcher. --a----- * This work was supported by the Office of Resource Analysis of the U. S. Geological Survey under Contract No. 14-08-0001-15985. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the author, and do not necessarily reflect the views of the U. S. G. S. 231 II USE OF THE MATCHER IN KNOWLEDGE ACQUISITION ----- A. Aid in Maintaining Probabilistic Consistency -- of the Inference Network -- The knowledge bases of many expert systems are organized as explicit or implicit networks of statements connected by rules or logical constructs. Because such networks provide the framework for judgmental reasoning, various numerical values, such as probabilities or certainties, are often maintained in them. A major concern of expert systems is the difficulty of keeping the knowledge base error free and consistent in form and content as it grows. Let us examine how the Matcher assists Prospector in maintaining probabilistic consistency in the case where Sl is the most recently entered statement, and S2, which already exists in the knowledge base, is a restriction of Sl. (a) Because S2 is a restriction of Sl, the probability of S2 can never exceed that of Sl. In particular, if the prior probabilities supplied by the domain specialist (DS) do not satisfy this constraint, the Matcher will detect the violation and a correction will be required. Thus, before a consultation begins, we can assume that P(S2) < P(S1). (b) Unfortunately, even though all the probabilistic constraints are initially satisfied, the probability changes .that follow from the use of inference rules may not maintain them. For example, if Sl and S2 are the hypotheses (right- hand sides) of two rules El --> Sl and E2 --> S2, and if the evidence (left-hand side) El is sufficiently unfavorable for Sl, we may have P(Sl[El) < P(S2). Similarly, if the evidence E2 is sufficiently favorable for S2, we may have P(S1) < P(S21E2). In essence, the problem is that when the DS provided the rule saying that the evidence El is unfavorable for Sl (rhyolite), he overlooked the fact that El is also unfavorable for S2 (a rhyolite plug), and did not supply a rule of the form El --> s2. Similarly, when he supplied the rule saying that the evidence E2 is favorable for a rhyolite plug, he overlooked the fact that E2 is also favorable for rhyolite, and did not supply a rule of the form E2 --> Sl. Indeed, the DS should not be asked to supply rules for such obvious deductions. The Matcher helps to detect these situations. It is the responsibility of the consultation system to take the appropriate actions to maintain the probabilistic consistency in the inference networks. In [l] it is shown how Prospector uses results from the Matcher to create additional rules ensuring that the probabilistic constraints will be maintained at run time when inference rules are applied. B. Aid in Designing the Semantic Representation -- Because statements in the knowledge base may be arbitrarily complex, their semantic encoding is often entered manually during the knowledge acquisition phase. During a consultation, however, the user is allowed to volunteer information to the system, and semantic re;r;;~;;:t,~~] is used to create the corresponding to the volunteered statements. The kinds of statements that can be translated by the parser depend upon taxonomy contents and an external grammar. Whether the semantic representation of statements is entered manually or constructed by a parser, the knowledge engineer needs to determine if the resulting representation is adequate. He must ensure that it reflects the intentions of the DS in all situations that could occur during of a consultation. Statements can be combined to form larger statements or broken into smaller units, and their semantic representation need not always be elaborate. Which representation is finally chosen depends upon what other statements are in the knowledge base and how they are related, as well as what the DS thinks is the most appropriate for each particular situation. Because the Matcher can be used to analyze how statements are related, it can assist in choosing an appropriate representation. In particular, no elaborate semantic representation may be needed for a statement (or a portion of a statement) that is unrelated to any other statement in the knowledge base. Because such a statement is unlikely to have a major effect on the consultation, a simple text-string representation would be adequate for most purposes. In addition to determining if a restriction/generalization relation exists between two statements, the Semantic Network Matcher in Prospector can identify corresponding elements of the statements and point out the nature of their differences. This feature has been exploited to some extent in the knowledge acquisition module of Prospector [2] where it was used to choose a representation for a statement from possible alternative representations. For instance, a conjunction "X and Y" can be encoded either as a single statement or as two statements, "X" and "Y," connected by a logical AND in the inference network. The first alternative is chosen if a statement already exists in the knowledge base that is equal to or is a restriction of "X and Y," or if "X and Y" is not related to any existing statement. The second alternative is chosen otherwise. We believe this approch can be generalized, and that an automatic procedure using the Matcher can be devised to assist in the uniform, and perhaps optimal, encoding of all statements in the knowledge base. C. As a Search Feature--Accessing by Contents --- - Development and testing of a knowledge base typically extend over long periods. The knowledge engineer cannot be expected to remember all the statements (or any labels assigned to them) that he 232 or another knowledge engineer developing the same knowledge base has already entered. The Matcher can be used as a search feature allowing the knowledge engineer to access statements by specifying a partial description of their (semantic) contents. In effect, the Matcher-search (a command of the knowledge acquisition system) will allow the knowledge engineer to say something like: "Now I want to work on that portion of the inference network that deals with sulfide mineralization." The search-by-content feature is accomplished by matching the partial description specified by the knowledge engineer with statements currently in the knowledge base. III USE OF THE MATCHER IN CONSULTATION ----- A. As a Tool for Maintaining Consistency of the -- User'-s Answers: -- 1 . Discovering Inconsistencies If a user is allowed to volunteer information to an expert system, logical inconsistencies in the input could result. For example, suppose the user volunteers the two statements Sl and S2 concerning rhyolite and a rhyolite plug. Because S2 is a restriction of Sl, if his statements imply that P(S2) > P(Sl), he will be reminded that his answers are contradictory. This is the case, for instance, if the user says that "there are igneous rocks" with some degree of certainty, but later volunteers that "there is a rhyolite plug" with a higher degree of certainty. The contradictions occurring in a consultation often involve several levels of inference and long chains of restriction/generalization links, which sometimes have embarrassed our expert users while being impressed by Prospector's ability to detect the inconsistencies. 2. Changing Answers A significant advantage of the Bayesian method used in Prospector for computing probabilities is the ease with which answers to questions can be changed without having to repeat all previous calculations. Basically, all that is required in changing an answer to a question about any evidence E is to change the probability for E and to propagate the results through the inference network. The possibility of violating the restriction/generalization probabilistic constraints causes the only complication in this process. However, by keeping a record of how statements are related as computed by the Matcher , the answer-changing program knows that it may also have to change the probabilities of some of the related statements in order to maintain consistency. For instance, if the inference network contains the two rules Sl --> Hl and S2 --> H2, and the user gives a negative answer to a question about Sl, the probability of Hl will be updated (in accordance with the rule strengths associated with the first rule). In addition, because S2 is a restriction of Sl, the probability of H2 must also be updated (in accordance with the rule strengths of the second rule) as if a negative answer had been given for S2 as well. When the user then changes his answer for Sl, the probabilities of both Hl and H2 will be automatically updated and propagated through the inference network. By changing an answer, the user may contradict some of his earlier assertions, and changing these assertions may give rise to still further contradictions. This can confuse the user, but poses no problem for the answer-changing program, which is recursive and will make sure no contradictions are introduced before resuming a consultation. B. Use of the Matcher as a Dialog Management Tool ---____-- 1. Mixed Initiative, Volunteering Information Prospector can work in either of two modes -- the consequent mode or the antecedent mode. In the consequent mode Prospector attempts to establish (or rule out) an hypothesis, and it traces backward through the inference network and searches for appropriate evidences to ask about. In the antecedent mode, the inference network is used in a forward direction to propagate the consequences of input information. Prospector is a mixed initiative system whereby the user has the option of taking control any time during the consultation session to inform Prospector about facts he believes to be relevant. The Matcher makes this possible by relating the volunteered information to the current knowledge base in the same fashion as it did for the knowledge acquisition phase. 2. Control Strategy and Goal Selection -- The information volunteered by the user is often relevant to several hypotheses; In Prospector, a simple scoring criterion is used to select the goal hypothesis. Among other things, this criterion takes into account the volunteered statements (whose effect on the hypotheses may be encouraging or discouraging) that are linked to each hypothesis as recorded by the Matcher. 7. Interaction Psychology Before the user is asked about the evidence E selected by the control strategy, Prospector 233 reminds him about any facts it thinks relevant. The information needed to recognize these facts are the links relating E to other statements in the knowledge base computed by the Matcher and recorded at some earlier phase of the consultation or during knowledge acquisition. How these facts are presented to the user depends upon the current llstatew of the statements involved. The state of a space is determined by its certainty value and how that certainty was established --whether it was inferred by using rules, volunteered by the user, Or inferred through restriction/generalization links through the Matcher. Depending upon the actual situation, one of several standard phrases is displayed before the question is asked, and an appropriate phrase is selected to ask the question. used The fol lowing are some of the in a Pr ospector consultation: standard phrases - You told me about . . . - You suspected . . . - I know you doubted . . . - Your statements imply . . . - I know there is reason to doubt . . . - I have reason to suspect . . . - I have reason to doubt . . . Thus, the program might preface a question about a piece of evidence E by saying: "I have reason to doubt E. What is your degree of belief about that?" Clearly, these stock phrases are simple attempts to inform the user about the implications of his earlier statements. Although they have no effect on the function of Prospector and are not necessary in any logical sense, they enhance communication between the user and the consultation system and often serve to make the logical processes of the consultation system more evident. The Matcher has been an important tool for the design of the interaction environment in all phases of development and use of the Prospector knowledge- based system. It is particularly important in the "psychology" of man-machine interaction in consultation systems that the user does not feel ignored and that the dialogs are not totally dictated by the system. Whenever possible, the user should be shown evidence that the system listens to him, understands what he says, and sometimes can even use the information he supplied! IV CONCLUSION w M [31 [41 [53 b1 L-71 REFERENCES Duda, R.O., P.E. Hart, N.J. Nilsson, R. Reboh, J. Slocum and G.L. Sutherland, "Development of a Computer-Based Consultant for Mineral Exploration," Annual Report, SRI Projects 5821 and 6415, SRI International, Menlo Park, California (October 1977). Duda, R.O., P.E. Hart, K. Konolige, R. Reboh, "A Computer-Based Consultant for Mineral Exploration," Final Report SRI Project 6415 SRI International, Menlo Park, California (September 1979). Hayes-Roth, F., "The Role Of Partial and Best Matches in Knowledge Systems," in Pattern- Directed Inference Systems, D.A-Waterman and F. Hayes-Roth, eds., pp. 557-574 (Academic Press, New York, 1978). Hendrix, G.G., "LIFER: A Natural Language Interface Facility," SIGART Newsletter, No. 61, pp 25-26 (Februarym. Hendrix, G.G., "Encoding Know1 edge in Partitioned Ne tworks." in Associative - The Representation and Use Networks of Knowledge in -- - Computers, N. V. Findler, ed., Academic Press, New York (1979). Reboh, R., "A Knowledge Acquisition Environment for Expert Consultation Systems," Ph.D. Dissertation, Department of Mathematics, Linkoping University, Sweden (to appear 1980). Waterman, D.A. and F. Hayes-Roth, eds., Pattern-Directed Inference Systems (Academic -New York, 1978). - By providing means to relate the statements in the knowledge base to each other, the semantic network Matcher in Prospector has been an important instrument in supporting many of the features that constitute the AI contents of the system. We believe that the approach is a general one, and can enhance the intelligent behaviour of any knowledge- based system. 234	1980	20
13	AN APPROACH TO ACQUIRING AND APPLYING KNOWLEDGE Norman Haas and Gary G. Hendrix SRI International 333 Ravenswood Avenue Menlo Park, California 94025 ABSTRACT The problem addressed in this paper is how to enable a computer system to acquire facts about new domains from tutors who are experts in their respective fields, but who have little or no training in computer science. The information to be acquired is that needed to support question- answering activities. The basic acquisition approach is "learning by being told." We have been especially interested in exploring the notion of simultaneously learning not only new concepts, but also the linguistic constructions used to express those concepts. As a research vehicle we have developed a system that is preprogrammed with deductive algorithms and a fixed set of syntactic/semantic rules covering a small subset of English. It has been endowed with sufficient seed concepts and seed vocabulary to support effective tutorial interaction. Furthermore, the system is capable of learning new concepts and vocabulary, and can apply its acquired knowledge in a prescribed range of problem-solving situations. I INTRODUCTION Virtually any nontrivial artificial intelligence (AI) system requires a large body of machine-usabge knowledge about its domain of application. Construction of a knowledge base is currently a tedious and time-consuming operation that must be performed by people familiar with knowledge representation techniques. The problem addressed in this paper is how to enable computer systems to acquire sets of facts about totally new domains from tutors who are experts in their own fields, but have little or no training in computer science. In an attempt to find a practical solution to this problem, we have developed a pilot system for knowledge acquisition, which, along with -------- * This research was supported by the Defense Advance Research Projects Agency under contract N00039-79-C-0118 with the Naval Electronic Systems Command. The views and conclusions contained in this document are those of the-authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the Defense Advanced Research Projects Agency of the United States Government. several related research issues, is discussed below. The kinds of information we are most interested in acquiring are those needed to support what have been called "question-answering" or "fact-retrieval" systems. In particular, our interest is in collecting and organizing relatively large aggregations of individual facts about new domains, rather than in acquiring rules for judgmental reasoning. This is in contrast to previous work on such systems as those of Davis [I] and Dietterich and Michalski [2], that treat knowledge not so much as a collection of facts, but as a set of instructions for controlling the behavior of an engine. The type of acquisition process we are exploring is "learning by being told," in contrast to the idea of "learning by example." It is this latter concept which has formed the basis of research by other investigators in this area, such as Winston [ll] and Mitchell [8]. Our interest in knowledge acquisition is motivated by the desire to create computer-based systems that can aid their users in managing information. The core idea is that of a system that can talk to a user about his problems and subsequently apply other types of software to meet his needs. Such software would include data base management systems, report generators, planners, simulators, statistical packages, and the like. Interactive dialogs in natural language appear the most convenient means for obtaining most of the application-specific knowledge needed by such intelligent systems. II KNOWLEDGE ACQUISITION THROUGH ENGLISH DIALOGS Systems that acquire knowledge about new domains through natural-language dialogs must have two kinds of special capabilities. First, they must be capable of simultaneously learning both new concepts and the linguistic constructions used to express those concepts. (This need for simultaneous acquisition of concepts and language reflects the integral connection between language and reasoning.) Second, such systems must support interactive, mixed-initiative dialogs. Because a tutor may provide new knowledge in an incremental 235 and incomplete manner, the system must keep track of what it has already been told so that it can deduce the existence of missing information and explicitly ask the tutor to supply it. We are exploring the feasibility of such ideas by developing a series of Knowledge-Learning and -Using Systems (KLAUS). A KLAUS is an interactive computer system that possesses a basic knowledge of the English language, is capable of learning the concepts and vocabulary of new subject domains, and has sufficient expertise to apply its acquired knowledge effectively in problem-solving situations. III RESEARCH ISSUES FOR KNOWLEDGE ACQUISITION -- To create systems capable of acquiring knowledge through tutorial dialogs in English, several fundamental research problems must be resolved: A powerful natural-language processing capability is required. Although much progress has been made in recent years, previous work has assumed a complete knowledge base. Knowledge-acquisition dialogs require several adaptations and extensions. Seed concepts and seed vocabulary must be identified for inclusion in a core system. It is not at all obvious which words and concepts will be most useful in helping tutors describe the concepts of new domains. A structure for lexical entries must be specified so that the system can acquire new lexical information. Because such information provides a key link between surface linguistic form and underlying meaning, structural specification is a very challenging task for certain categories of words, particularly verbs. The linguistic constructions that people use in introducing new concepts must be identified and analyzed so they can be interpreted correctly by the natural- language processing system. Such constructions range from simple syntactic patterns to complex uses of analogy. A flexible scheme of knowledge representation is necessary. The representation must have general expressive power, since it may be applied to many different domains and must support the addition of new information. It should include inherent features that can aid in organizing knowledge and in supporting the incremental acquisition of knowledge. An efficient problem-solving capability is needed to answer questi6ns and to draw inferences for integrating newly acquired information. This capability must be based on general principles, because no application-specific problem-solving procedures will be included in the system. (How to acquire such procedures is a separate and interesting research question.) methodology is needed for integrating new concepts into the system's knowledge base. Because tutors will often provide only partial descriptions of new concepts, methods have to be devised for ascertaining what additional facts must be sought from the tutor to insure proper linkage between the new concepts and those previously acquired. A set of readily understandable questions is needed for eliciting information from tutors. The length and number of questions should be minimized to impose as small a burden on tutors as possible. These problems must be dealt with in an integrated manner, balancing the requirements of one facet of the system against those of others. Our initial attempts to cope with this complex of issues are presented below. OK, now I have assimilated LENGTH. 'i'4> A foot is a unit of length ---____- FOOTS are the first units of LENGTH you've defined. Now I have assimilated FOOT. 75> The plural of foot is feet ----- FEET - got it. Thanks. 76> A meter is a unit of length ----__-- How is it related to FOOT? >> There are 3.3 --- feet in a meter ---- Now I have assimilated METER. 77> A physical object has a length So PHYSICAL OBJ~a~L~N~ Exactly what to use as seed concepts is problematical for a system that must bootstrap its entry into a new domain. Most of the seed concepts in NANOKLAUS are classes of THINGS and RELATIONS. They have been included in the system either simply to avoid forcing tutors to relate everything to the most general concept (THING), or because they have a special status in English. For example, because pronominal usage depends on gender, the class MALE- BEING is defined and associated with the pronoun "he." One might consider defining as seed concepts a set of primitives, to which all other concepts must be reduced, but such a reductionist approach is probably unworkable [3]. NANOKLAUS uses five principles of knowledge organization to integrate new knowledge: (1) there are things; (2) th ere are subclasses of things (i.e., things can be subclassified); (3) there are relations among things; (4) there are subclasses of relations; (5) some of the relations are functions. The concepts of uniqueness and equality also play important roles. NANOKLAUS is not programmed to hold explicit conversations about these concepts, but rather to use them in its internal operations. C. The Natural-Language Component The natural-language component of NANOKLAUS uses a pragmatic grammar in the style of LADDER b1. Although most of the linguistic processing performed by the system follows fairly standard practice, the pragmatic grammar is distinguished by its explicit identification of a number of syntactic structures used principally to define new concepts. As an oversimplified example, NANOKLAUS might be thought of as looking for the syntactic pattern <s> => <A> <NEW-WORD> <BE> <A> <KNOWN-COUNT-NOUN> to account for such inputs as A CARRIER IS A SHIP. When one of these concept-defining patterns is recognized, an acquisition procedure associated with the pattern is invoked. 'This procedure generally adds new facts to the system's set of wffs and generates new entries in its lexicon. The various acquisition procedures also have provi sions for responding to the tutor. Response generation is accomplished through the use of preprogrammed phrases and templates. D. Using Dialog to Aid Acquisition and ---- Assimilation By and large, it is unreasonable to expect tutors to volunteer all the information NANOKLAUS needs to assimilate new concepts. In particular, tutors cannot be expected to know what conclusions NANOKLAUS will draw about a newly taught concept from its previous knowledge, since they know neither the details of its knowledge nor the even more esoteric details of its assimilation procedures. NANOKLAUS must ask the tutor for the information it needs. But wording the system's questions in an understandable form involves a number of human engineering difficulties. The acquisition of new verbs offers a cogent illustration of the problem. English verbs are highly idiosyncratic; making proper entries for them in a lexicon can be a formidable task. Among other considerations, one must ascertain a verb's passivity and transitivity, determine whether its indirect object can become the object of a FOR or TO prepositional phrase, whether it is reflexive or nonreflexive, and how the syntactic cases of its arguments may be "normalized" when the verb appears in different syntactic constructions. NANOKLAUS's tutors, of course, cannot be expected to describe verbs in linguistic terms, so the system asks a series of questions tutors can understand to elicit the same information. The following is a typical verb-acquisition exchange: 92>An officer can assign a unit ------ Give the -ed and -en forms of ASSIGN (for example, for GO: WENT GONE): >>assigned assigned If an OFFICER can ASSIGNanUNIT something, type the most general description of that something, otherwise <return> >> (user typed a carriage return) If an OFFICER can ASSIGN SomethinganUNIT, type the most general description of that something, otherwise <return> >>a task group --- Can an OFFICER ASSIGN an UNIT TO a TASK GROUP ? >>yes Could an UNIT be ASSIGNED a TASK GROUP by an OFFICER ? >>yes Can an OFFICER ASSIGN an UNIT with a TASK GROUP ? >>no Can an OFFICER ASSIGN a TASK GROUP ? >>no - Can an OFFICER ASSIGN ? >>no Verbs that produce rather differenTanswers to the above questions include GIVg, BUY, RUN (run a machine), SERVE and SUPPLY. E. The Knowledge Representation System NANOKLAUS's knowledge representation system uses a many-sorted first-order logic, that combines features from [g] and [5]. The backbone of the system is a treelike data structure reflecting the NANOKLAUS does not consider modal verbs (e.g., "want" and "know"), or verbs with particles (e.g., "pick up", "preside over") or sentential objects (e.g., "The captain requested that the ship change ---- coursell). 237 hierarchy of sorts (or classes) used by the system. This tree is "tangled" in that any given sort may if any, together with CARRIER, span the class of all SHIPS (with or without have multiple ancestors. It is no accident that a sort hierarchy should serve as the primary data structure for an English-speaking acquisition system. "ISA" hierarchies are used by many natural-language processing systems; it appears that something very similar plays a central role in the their knowledge [7]. to a sort hierarchy way humans organize to determine whether or not a given entity to a particular sort . To a id in this task A fundamental task of the deductive system is belongs NANOKLAUS maintains information about the direct descendants of each sort, such as whether some or all of the descendants of an ancestor exhaust (span) it, and whether are mutually exclusive. two or more sibling sorts The utilization of sort information is clearly illustrated in the following dialog with NANOKLAUS: 80>A ship is a unit YouTrGyGg-that anything that is a SHIP is also an UNIT. What is an UNIT? >>a thing You're saying that anything that is znm is also a THING. Which of the following, if any, are proper subclasses of UNIT: PHYSICAL OBJECT, LEGAL PERSON, or MEASURE? >>none of them --- Which of the following classes, if any, could never have any members in common with UNIT: PHYSICAL OBJECT, LEGAL PERSON, or MEASURE ? >>none OK, now I have assimilated UNIT. OK, now I have assimilated SHIP. 81>A ship is a physical object YouTrEyzg-that anything that is a SHIP is also a PHYSICAL OBJECT. Is ANIMATE OBJECT a proper subclass of SHIP? >>no IS ANIMATE OBJECT necessarily<omposed of entirely different members from SHIP? >>yes Do ANIMATE OBJECT and SHIP span the set of all PHYSICAL OBJECTS >>no OK, now I have assimicted SHIP. 82>A submarine is a ship --- YouTre saying that anything that is a SUBMARINE is also a SHIP. OK, now I have assimilated SUBMARINE. 83>A destroyer is a ship --- YouTre saying that anything that is a DESTROYER is also a SHIP. Is SUBMARINE a proper subclass of DESTROYER? >>no Is SUBMARINE necessarily cozosed of entirely different members from DESTROYER? Do SUBMARINE and DESTROYER span the set- all SHIPS? >>no OK, now I ha= assimilated DESTROYER. 84>A carrier is a ship Yo?rwgThgt anything that is a CARRIER is also a SHIP. Which of the following, if any, are proper subclasses of CARRIERI SUBMARINE or DESTROYER? >>neither of them _I__-- Which of the following classes, if any, could never have any members in common with CARRIER: SUBMARINE or DESTROYER ? >>both of them --- What combination of the following subclasses, overlapping): SUBMARINE or DESTROYER ? >>neither of them --- OK, I have assimilated CARRIER. 85>A Kittyhawk is a carrier YouTre saying that-anything that is a KITTYHAWK is also a CARRIER. OK, now I have assimilated KITTYHAWK. . 98>A carrier has a doctor who is a person --------- So CARRIERS have PEOPLE for DOCTORS. Must a CARRIER have exactly one DOCTOR? >>no - 99>= Constellation is a Kittyhawk -- You're saying that CONSTELLATION is one instance of a KITTYHAWK. OK, now I know about CONSTELLATION. lOO>The JFK is a Kittyhawk --- You'resaying that JFK is one instance of a KITTYHAWK. OK, now I know about JFK. l>The JFK is commanded by Moffet AS-T= THA-(MOFFET C~MMANDFK) 2>The length of the JFK is 1072 feet --------- V FUTURE PROSPECTS At this time NANOKLAUS can be best described as a fragile proof-of-concept system still in its early developmental stage. During this coming year, we plan to greatly expand its linguistic coverage by replacing our current pragmatic grammar with Robinson's [lo] DIAGRAM grammar. Once this has been accomplished and NANOKLAUS's verb acquisition package extended to accept particles and prepositional phrases, we believe NANOKLAUS can serve as a useful tool for aiding AI researchers in the construction of knowledge bases for other AI systems --a task that currently consumes an inordinate proportion of research effort. As suggested in the introduction, one of our long-term objectives is the extension of KLAUS to knowing about diverse types of external software packages. Given knowledge of such packages, a KLAUS could serve as an agent that interacts with them on a user's behalf. To explore these possibilities, we plan in the near future to provide NANOKLAUS with the capability of using a conventional data base management system. In this configuration, a user should be able to tell NANOKLAUS about a new domain, about a data base containing information pertaining to that domain, and about the interrelationship of the two. The new system would then be able to use the data base in answering questions regarding the domain. Our work in the area of knowledge acquisition per se has really just begun. As development proceeds, we plan to turn our attention to making provisions for learning by analogy, for acquiring and reasoning about the internal structures of processes, for dealing with causality, and for dealing with mass terms. 238 ACKNOWLEDGMENTS The deduction system supporting NANOKLAUS was developed in large part by Mabry Tyson, with Robert Moore, Nils Nilsson and Richard Waldinger acting as advisors. Beth Levin made major contributions to NANOKLAUS's verb-acquisition algorithm. Paul Asente assisted in the testing of the demonstration system. Barbara Grosz, Earl Sacerdoti, and Daniel Sagalowicz provided very useful criticisms of early drafts of this paper. 1. 2. 3* 4. 5. 6. 7. 8. 9. REFERENCES R. Davis, "Interactive Transfer of Expertise: Acquisition of New Inference Rules," Proc. 5th International Joint Conference on Artmar Intelligence, Cambridge, Massaczsetts, pp. 321-328 (August 1977). T. G. Dietterich and R. S. Michalski, "Learning and Generalization of Characteristic Descriptions: Evaluation Criteria and Comparative Review of Selected Methods," Proc. 6th International Joint Conference on Artificial Intelligence, Tokyo, Japan, pp. 223-231 (August 1979). J. A. Fodor, The Language Of Thought, pp. 124- 156, (Thomas Y. Crowell CoFaz York, New York 1975). G. G. Hendrix, "The LIFER Manual: A Guide to Building Practical Natural Language Interfaces," Technical Note 138, Artificial Intelligence Center, Stanford Research Institute, Menlo Park, California (February 1977). G. G. Hendrix, "Encoding Knowledge in Partitioned Networks." in Associative Networks I - The Representation and Use of Knowledge in -- Computers, N. V. Findler, ed.-(Academic Press, New York, New York 1979). G. G. Hendrix, E. D. Sacerdoti, D. Sagalowicz and J. Slocum, "Developing a Natural Language Interface to Complex Data," ACM Transactions on Database Systems, Vol. 3, No. 2 (June 1978).------ P. H . Lindsay and D. A. Norman, Human Information Processing. (Academicxs, New York, New York, 1972.) T. M. Mitchell, "Version Spaces: a Cand Elimination Approach to Rule Learning," 5th International Joint Conference on - idate Proc. Artificial Intelligence, Cambridge, Massachusetts, pp. 305-310 (August 1977). R. Moore, "Reasoning from Incomplete Knowledge in a Procedural Deduction System," MIT Artificial Intelligence Laboratory, AI-TR-347, Massechusetts Institute of Technology, Cambridge, Massechusetts (1975). IO. 11. J. J. Robinson, "DIAGRAM: an Extendable Grammar for Natural Language Dialogue," Technical Note 205, Artificial Intelligence Center, SRI International, Menlo Park, California (February 1980). P. H. Winston, "Learning Structural Descriptions from Examples," Chapter 5 in The Psychology of Computer Vision, P. H. Wins%%, ed. (McGrawxi.11 Book Company, New York, New York (1975). 239	1980	21
14	SELF-CORRECTING GENERALIZATION Stephell B. Wilitehill Deyartment of Irlformation and Computer Science University of California at Irvine Irvine, Ca 92717 ABSTRACT A system is described which creates and generalizes rules from examples. The system can recover from an initially misleading input sequence bY keeping evidence which supports (or doesn't support) a given generalization. By undoing over-generalizations, the system maintains a minimal set of rules for a given set of inputs. I GENERALIZATION Many programs have been written which generalize examples into rules. Soloway[5] generalizes the rules of baseball from examples. Hayes-Roth[3] and Winston[8] yeneralize common properties in structural descriptions. Vere[G] has formalized yeneralization for several applications. If a program maintains a current hypothesis about the rule set as it sees new examples it is said to generalize incrementally. A program that incrementally forms generalizations may be sensitive to the order in which examples are presented. If exceptional examples are encountered first, the program may over-generalize. If the program is to recover and undo the over-generalization it must have a certain amount of knowledge about why the over-generalization was made. The system to be described here has this type of self-knowledge. BY associating positive and negative evidence with each generalization, the system is able to reorganize its rules to recover from over-generalizations. Even if it is initially misled by an unusual sequence of inputs it still discovers the most reasonable set of rules. II THE PROBLEM DOMAIN The problem domain chosen for the system is learning language rnorphology from examples. For example, from the words " j urnped" , "walked" and "kissed" we can deduce the rule that the English past tense is formed by adding "ed". The Concept Description Language consists of a set of rules. Each rule is a production consisting of a pair of character strings. When the left-hand side is matched the right-hand side is returned. The left-hand string may optionally contain a I# which will match zero or more cllaracters. In this case the right hand cklaracter string may contain a 'I and the value the star on the left hand side which was matched is suhstltuted for the star on the riyht hana side. For example the protiuction for t11e example above looks like: ->Nl. 'l'his ->ED rule does not always work. From "baked", "related" and "hoped" we see that for words ending in "e" we need only to add an 'Id". This rule is written as E->ED. For this domain, the problem bears resemblence Grammatical Inferznce Problem[4]. to the III RELATIONSHIPS BETWEEN RULES Rule Containment. Given rules Pl and p2 I let Sl be the set of strings which match the left-hand side of Pl and S2 the set of strings which match left-hand side of P2. If Sl is a subset of S2 then we say that Pl is contained by P2. This forms a partial ordering of the rules. The Is-a-generalization-of Operator. Given rules Pl and P2, let Sl be the set of strinys which match the left-hand side of Pl and S2 the set which match the left-hand side of P2. If Pl contains P2 and if Pl and P2 produce the same result for every element in S2 trlen Pl is d yeneralization of P2. This is also a partial ordering. Note the distinction between the containment operator and the is-a-qeneralization-of operator. Basically, containment deals with the left-hand side of a rule. Is-a-generalization-of deals with both sides. An example will clarify this: ->S contains K->KS ->S is a generalization of K->KS ->S contains CH->CHES ->S and CH->CHES are unrelated by generalization By definition, if Pl is a generalization of P2, Pl contains P2. The converse is not necessarily true. is If Pl is a generalization of P2 and Pl a generalization of P3 then Pl is a a generalization of no other common generalization. Roughly, the maximal common generalization is the one which captures all common features of the rules being generalized. E'o r example, given WALK->WALKED and TALK->TALKED, possible generalizations are: ->ED, K->KED, LK->LKED and ALK->ALKED. The last one, ALK->ALKED is the maximal one. In the concept description language we are using all common generalizations are related on the is-a-generalization-of operator. Therefore in our domain the maximal common yeneralization is unique. IV ORGANIZATION OF - RULES The rules and their evidence are organized in a tree structure. At the top level the rules are organized as a rule list. A rule list is a list of rules partially ordered on the containment operator. No rule may be contained by a rule which precedes it in the list. Associated with most rules is some evidence which is itself in the form of another rule list. The only rules without evidence are the example pairs whose evidence is fact. These correspond to terminal nodes in the evidence tree. If a rule Rl would contain a rule R2 which follows it then Rl is marked as being blocked by R2. If Rl blocks R2 then evidence for Rl is negative evidence for H2. The positive evidence consists of those rules which were generalized into the current generalization. Negative evidence for a generalization G is all the evidence of generalizations that are blocked by G. .Thus when ->ES blocks (N->NS + K->KS == > ->S) that is negative evidence for ->ES. The evidence described here is much like evidence used for explanation in [ll or to maintain beliefs in [2]. In our system the evidence is used for reorganization, but it could be used for these other purposes as well. Hule Application and Conflict Hesolution. When the rule interpreter produces a response, it is as if it finds a11 rules which match the given input and tnen uses the one which doesn't contain any of the others (informally, the one with the least general left-hand side). In reality the rules and rule interpreter are organized so that the first rule that matches is the desired one. Inserting New Rules. If a rule has produced the ----EOrrect result, the new example pair is inserted into the evidence list for the rule. If the rule system has not produced the correct result the rule is inserted in the main rule list before the first rule witn which it will generalize. If it will not generalize with any rule, it is inserted before the first rule that contains it. The same rule insertion algorithm is used to insert new rules or evidence. This means that generalizations take place in an evidence list in the same way that they do in the main rule list. V SYSTEM REORGANIZATION Each blocked generalization has knowledge about which generalization is blocking it. Whenever evidence for a blocked generalization Gl is entered into the rule structure, we check to see if there is now more evidence for Gl than for the blocking generalization G2. If so, G2 is moved to the position in the rule list immediately preceding Gl, G2 is marked as being blocked by Gl and Gl is no longer marked as being blocked. There are several choices on how to compare positive and negative evidence. The one chosen is to count how much direct evidence there is for a rule. Direct evidence is that evidence found in the top level rule list in the evidence tree. Another metnod which was reJected for this application is to count the total number of pieces of evidence in the evidence tree. The first method was chosell because Ch->CHES and X->XES are exceptions to -> s (rather than "A->AS, B->BS, C->CS, etc. being exceptions to ->ES) because there is more direct evidence for ->S (rules like A->AS) than for ->ES. Even if half the words in English used CH->CHES this would still be an exception to ->s. This method produces the most reasonable set of rules. The system has been tested on pluralizing French adjectives. French has a much more complicated morphology than English, having not only exceptions to rules but also exceptions to exceptions. The system was found to produce the same rules for pluralizing French adjectives as those found in a French-English dictionary . A detailed example of this appears in [7]. VI UNDOING GENERALIZATIONS - AN EXAMPLE - The system is written in MLISP and runs on UC1 LISP. The following example was annotated from a trace of the system in operation. The rules and evidence are printed as a tree. The evidence for a node is indented from the parent node in the printout. INPUT WORD? church WHAT IS RESPONSE? churches INPUT WORD? match WHAT IS RESPONSE? matches 241 INPUT WORD? bus WHAT IS RESPONSE? buses RULES: ->ES BUS->BUSCS CH->CHES MATCH->MATCHES ChUHCH->CHURCHES At this point we have over-generalized. We will find this out later. The only rule seen by tile rule interpreter is -> ES. BUS->BUSES arid CH->CdES are evidence for ->ES. MATCH->MATCHES and CHURCH->CHURCHhS are evidence for the CH->CHES rule (which is itself evidence). INPUT WORD? book IS RESPONSE BOOKES? n WHAT IS RESPONSE? books INPUT WORD? back IS RESPONSE BACKES? n WHAT IS RESPONSE? backs RULES: BK->BKS BACK->BACKS BOOK->BOOKS ->ES BUS->BUSES "CH->CHES MATCH->MATCHES CHURCH->CHURCHES What should be regular cases are treated as exceptions. INPUT WORD? car IS RESPONSE CARES? n WHAT IS RESPONSE? cars RULES: (->s> CAR->CAKS BK->BKS BACK->BACKS BOOK->BOOKS ->ES Bus->i3usLs CH->CfiES MATCH->MATCHES CHURCH->ChURCHES At this point we want to make the generalization ->s but this generalization is blocked by ->ES. We make the generalization but mark it as blocked. The parentheses indicated that the rule is blocked. The only productions seen by the production system are: CAR->CARS, BK->BKS and ->ES. The blockage of ->S is negative evidence for ->lzs. The system will detect that the rules are in the wrong order when there is more evidence for ->S (and hence against ->ES) than there is for ->ES. At this point there is just as much negative evidence as positive (looking down one level in the evidence tree). INPUT WORD? bat IS HESPGNSE BATES? n WHAT Is RESPONSE? bats RULES: (->ES) BUS->BUSES CH->ChES MATCH->kATCHES CHURCH->CHUHCHES -> s CAR->CAhS BK->BKS BACK->BACKS BOOK->BOOKS BAT->BATS This addition negative evidence for ->ES has caused a reordering of the rules. ->ES is now blocked by ->s (as it should be). INPUT WORD? house IS RESPONSE HOUSES? y INPUT WORD? bunch IS RESPONSE BUNCHES? y The system now has a properly ordered rule set and can handle both regular and irregular cases. VII CONCLUSIONS BY giving a generalization program some self-knowledge it can recover from initially misleading input sequences. This introspection can be achieved by associating positive and negative evidence with generalizations. Without knowledye about what led to a generalization, it is not possible to undo the generalization. The system described here discovers the most reasonable set of morphological rules for a yiven language construct (the set found in a dictionary) reyardless of the input sequence. Tne choice of language morpllology as the problem domain was arbitrary. Any domain with a concept descriptiorl language whose maximal common generalization is unique would serve just as well. Furtlier work is needed for concept description languages whose maximal common generalization is not necessarily unique. Any incremental generalization program could irnprove its ability to recover from misleading input by applying the techniques described. [II [21 c31 [41 c51 [61 [71 [al REFERENCES Bechtel, R., Morris, P. and Kibler, D., "Incremental Deduction in a Real-time Environment", Canadian Society for the Computational Studies of -- Intelligence (May 1980). Doyle, J., "A Glimpse of Truth Main- terlance", 6-IJCHI (1979), 232-237. Hayes-Rotll, F . and McDermott, J . , “Knowledye Acquisition from Structurai Descriptions", Department of Computer Sciellce, Carnegie-Mellon Univ. (1976). Hunt, E. "Artificial Intelligence" 1975 Soloway, E. and Riseman, E., "Levels of Pattern Description in Learning", 5-IJCAI (1977), 801-811. Vere, S., "Induction of Relational Productions in the Presence of Background Information", 5-IJCAI (19771, 349-355. Whitehill, S., "Self-correcting Gener- alization". U.C. Irvine C.S. Tech Report no. 149. (June 1980). Winst on, P - , " Le Descr iptio ns From Techn ical Report 231 ar EX ning ample (1970 Structural s " , MIT-AI 242	1980	22
15	intelligent Retrieval Planning Jonathan J. King Computer Science Department Stanford University A. introduction artificial intelligent retrieval planning is the application of intelligence techniques to the task of efficient retrieval of information from very large databases. ’ Using such techniques, significant increases in efficiency can be obtained. Some of these improvements are not available through standard methods of database query optimization. lntelligent retrieval planning presents interesting issues related to other artificial intelligence planning research: planning with limited resources[2], optimizing the combined planning and execution process[9], and pursuing plans whose success depends upon the current contents of the database[S]. An experimental system has been implemented to demonstrate the novel kinds of query optimizations and to test strategies for controlling the inference of constraints. The problem of query optimization has arisen wlth the development of high level logical data models and nonprocedural query languages ([I], [3]). These free a user from the need to understand the physical organlzation of the database when posing a query. However, the user’s statement of the query may lead to very inefficient processing. Standard techniques of query optimization WI, Cl 11, Cl211 manipulate the set of retrieval operations contained in the query to find a relatively inexpensive sequence. The manipulations are independent of the meaning of the query, depending entirely on such factors as the size of the referenced files. The essential advance of intelligent retrieval planning over standard techniques of database query optimization is to combine knowledge about the semantics of the application domain with knowledge about the physical organization of the database. Domain knowledge makes it possible to use the constraints in a database query to infer additional constraints which the retrieved data *must satisfy. These additional constraints may make it possible to use more efficient retrieval operations or permit the execution of a sequence of operations that has a lower cost. Knowledge of the physical organization of the database can be used to limit the attempts to make such inferences so that the combined process of retrieval and inference is cost effective. ” The research described here is part of the Knowledge Base Management System Project at Stanford and SRI, supported by the Advanced Research Projects Agency of the Department of Defense under contract MDA908-77-C-0822. 8. Findino semantic eauivaients of a database auerv The techniques of intelligent retrieval planning will be illustrated with a simple example relational database with data about the deliveries of cargoes by ships to ports. The database contains three files, SHIPS, PORTS, and VISITS, with the attributes indicated: SHIPS: (Shipname Type Length Draft Capacity) PORTS: (Portname Country Depth Facilities) VISITS: (Ship Port Date Cargo Quantity) Semantic knowledge of the application domain is represented as a set of rules. The database is forced, via update restrictions, to conform to this set of rules. The general semantic knowledge for our sample database consists of these rules: Rule RI. “A ship can visit a port only if the ship’s draft is less than the channel depth of the port.” Rule R2. “A ship can deliver no more cargo than its rated capacity.” Rule R3. “Only liquefied natural gas (LNG) is delivered to ports that are specialized LNG terminals.” Rule R4. “Only tankers deliver oil”. Rule R5. “Only tankers can be over 600 feet long.” During intelligent retrieval planning, the use of the rules is shifted from checking updates to inferring constraints. That is, given certain query constraints, it is Possible to infer new constraints that the desired items must meet. For example, suppose a query requests the names of all ships that are longer than 650 feet. By rule R5, it can be inferred that a semantically equivalent retrieval request is for the names of tankers that are longer than 860 feet. This inferred description of the items to be retrieved may permit more efficient processing than the origlnal description. 243 C. The physical orqanization of a database Inferred semantically equivalent sets of constraints can be exploited for intelligent retrieval only if the physical organization of the database, and hence the cost of processing queries, is taken into account. Often, the physical organization has been arranged so that the cost of retrieving a restricted subset of data depends upon the data attributes that have been restricted. For instance, a file may have an auxiliary “index” on one of its attributes. If such an index exists, then the data pages that contain items that meet a constraint on that attribute can be identified directly and only those pages will be fetched. An indexed scan will be much less expensive than a scan through an entire file, measured in terms of pages fetched from disk. A discussion of retrieval costs for-different physical database organizations is contained in [4]. Thus, given a query that constrains only unindexed attributes. a reasonable semantic retrieval strategy (subject to qualifications discussed in [4]) is to attemdt to infer constraints on indexed attributes. Suppose that the SHIPS file has an index on the Type atf%bute. In that case. the best way to retrieve all the ships longer than 650. feet would be to fetch all the tankers by means of an indexed scan on Type, and then to check the Length value of each record fetched into main memory by that- scan. D. Novel auery optimization based on the use of domain semantics A query optimization method that uses domain semantics is interesting to the extent that it achieves significant increases in efficiency that are not available by other methods. One unique strategy that can arise when semantics are considered is the inclusion of an extra file in the set of files examined when a query is processed. For example, suppose a query requests the quantity of liquefied natural gas delivered for each known visit to ports with a channel depth of less than 20 feet. With no inference, a typical query processor would retrieve all PORTS records with a Depth value of less than 20. For each one, it would retrieve all VISITS whose Port attribute was the same as the Portname for the PORTS record and whose Cargo attribute was liquefied natural gas. The cost of the retrieval varies as the product of the sizes of the PORTS and VISITS files. However, with appropriate rules and indexes, intelligent retrieval planning can provide a much faster retrieval method. Suppose that the VISITS file has an index on the Ship attribute. In effect, this means that the database has been set up to provide inexpensive access from each ship to the set of its visits, while the set of visits to a specific port is much costlier to retrieve. Using rule Rl, it can be inferred that the visits requested by the query could have been made only by ships with a draft of less than 20 feet. It is now possible to retrieve SHIPS with Draft less than 20, then retrieve their associated VISITS (using the index), and finally, for each VISITS record with a Cargo value of liquefied natural gas, retrieve the associated PORTS record to check the value of Depth. If the Draft constraint substantially restricts SHIPS (and therefore the associated VISITS as well), then the overall cost will be much lower than that of the straightforward method, despite the fact that an extra file and an extra retrieval operation have been added. In a simulation test of this method using a cost model based on the System R relational database system[7] in which the VISITS file Is much larger than the PORTS and SHIPS files, the simulated retrieval cost was reduced by more than order of magnitude. E. Controllinq the inference of additional constraints Intelligent retrieval planning is complicated by the need to weigh possible gains in retrieval efficiency against the cost of performing inferences. The amount of planning done in the intelligent retrieval planning system in processing a particular query is determined by the cost of answering the unimproved query, and the possible improvements. The inference control mechanism has these key features: (1) The specific retrieval problem determines which constraints to try to infer (for example, an attempt is made to add constraints to indexed fields). (2) Knowledge about both the structure and the content of the database determines the effort to devote to attempting some inference. (3) Retrieval from the database is an inherent part of the inference process. The ability to carry out an inference (and hence the shape of the whole retrieval plan) may depend upon the current contents of the database. These features can be illustrated briefly in another example. Suppose the VISITS file is indexed only on Cargo, and a query requests data on visits to the port of Zamboanga. The retrieval strategy mentioned in section 3 suggests an attempt to infer a constraint on Cargo from the given constraint on Port. Given the number of records in the VISITS file, it is possible to compute the effort needed to perform a sequential scan. The effort alloted to inference will be a function of this. There is no guarantee that a helpful constraint can be found for any particular query. This suggests a policy to allot to the inference process a fixed small fraction of the effort which the original retrieval would take. With such a policy, the effort to plan the retrieval will result in a minor increase in response time if the inference attempt fails, but may provide a major improvement if It succeeds. Although the policy is intuitively plausible, other strategies for alloting effort during problem solving under uncertainty, such as those discussed in [S], are being investigated. Control of the inference process can be viewed as control of the moves in a space of constraints on attributes. Constraints can be moved either by applying a rule, by retrieving items restricted on one attribute and observing their values on other attributes, or by matching constraints on attributes defined on the same underlying set of entities. Continuing the example, starting with a constraint on the Port attribute of VISITS, new constraints can be found by retrieving from VISITS or by assigning the 244 value “Zamboanga” to the Portname field of PORTS. The first choice is rejected because the objective is to reduce the cost of that very retrieval. With a constraint on Portname in PORTS, a retrieval from PORTS can be performed. In this case, just a single record will be obtained because Portname is the unique identifier in that file. With appropriate access methods, such as hashing, the retrieval will be very inexpensive. When the PORTS record for l@Zamboanga” has been obtained, rules Rl and R3 may apply. if rule R3 applies, that is, if Zamboanga is a specialized liquefied natural gas terminal, then a strong constraint will be obtained on the goal attribute Cargo, and retrieval from VISITS will take place by means of an indexed scan rather than by means of a more expensive complete scan. If the data on Zamboanga does not support that inference, then other inference paths will have to be considered. This illustrates the possible dependence of retrieval planning on the current contents of the database. References 1. Codd, E.F., A relational model for large shared dala banks Commun ACM 13:6 (19701, 377-387. IA- 2. Garvey, Thomas D., Perceptual strategies for purposive vision, Technical Note 117, SRI International, Menlo Park, California, September 1976. 3. Kim, Won, Relational database systems, ACM Computinq Surveys 11:3 (19791, 186-212. 4. King, Jonathan J., Exploring the use of domain knowledge for query processing efficiency, Technical Report HPP-79-30, Heuristic Programming Project, Computer Science Department, Stanford University, December 1979. The cost of each inference step: generating new inference path nodes, testing rules, and retrieving from the database itself, is taken from the allotment of planning resources. Planning terminates if a strong goal constraint is found, if no potential inference path can be extended, or if planning resources are exhausted. F. Conclusion 6. Klahr, Philip, Planning techniques for rule selection in deductive question-answering, In Pattern Directed Inference Systems, D.A.Waterman and F. Hayes-Roth (Eds.), Academic Press, 1978. 6. Moore, Robert C., Handling complex queries in a distributed data base, Technical Note 170, SRI International, Menlo Park, California, October, 1979. Intelligent retrieval planning can provide novel and significant improvements in query processing efficiency. it draws on a knowledge of the physical organization of the database and on semantic knowledge of the application modelled in the database. The outcome of retrieval planning, both the retrieval method chosen and its cost, can depend upon the current contents of the database. An experimental system exists that performs inferences on queries stated in a subset of the SODA relational database query language[6]. The system uses a simple retrieval cost model to select the least expensive semantically equivalent expression of the retrieval request. The cost model is used in conjunction with a planning executive to limit the inference of additional constraints. Work Is under way to codify intelligent retrieval strategies which, though they are specific to a given class of physical database organizations, are independent of the application domain. The eventual aim of this work is to develop a system which, given the set of domain rules and the description of the physical organization for a database, can provide the functions of intelligent retrieval planning described in this paper, much as the EMYCiN system[lO] provides knowledge acquisition functions independently of the knowledge base to be built. 7. Selinger, P. Griffiths et. al. Access path selection in a relational database management system, In Proc. ACM-SIGMOD 1979 Boston, Mass., pp. 23-34. -9 8. Smith J.M. and P. Chang, Optimizing the performance of a relational algebra data base interface, Commun ACM 18:lO (1975), 568-579. -- 9. Sproull, Robert F., Strategy construction using a synthesis of heuristic and decision-theoretic methods, Report CSL-77-2, Xerox Palo Alto Research Center, Palo Alto, California, July 1977. 10. Van Melle, William, A domain-independent production- rule system for consulation programs, IN Proc. IJCAI-79 Tokyo, Japan, 1979, pp. 923-926. 11. Yao, S. Bing, Optimization of query evaluation algorithms, ACM Transactions on Database Systems 4:2 (1979) 133-l 55. 12. Youssefi, Karel A. Allen, Query processing for a relational database system, Memorandum UCB/ERL M78/3, Electronics Research Laboratory, University of California, Berkeley, California, January 1978. Acknowledgments Many thanks for perceptive comments by Jim Bennett, Jim Davidson, Larry Fagan and Jerry Kaplan of Stanford University, and Barbara Grosz of SRI International. 245	1980	23
16	A THEORY OF METRIC SPATIAL INFERENCE Drew McDermott Yale University Department of Computer Science New Haven, Connecticut ABSTRACT* Efficient and robust spatial reasoning requires that the properties of real space be taken seriously. One approach to doing this is to assimilate facts into a "fuzzy map" of the positions and orientations of the objects involved in those facts. Then many inferences about distances and directions may be made by "just looking" at the map, to determine bounds on quantities of interest. For flexibility, there must be many frames of reference with respect to which coordinates are measure. The resulting representation supports many tasks, including finding routes from one place to another. **Jr In the past, AI researchers have often sought to reduce spatial reasoning to topological reasoning. [4, 6, 81 For example, the important problem of finding routes was analyzed as the problem of finding a path through a network or tree of known places. This sort of formulation throws away the basic fact that a route exists in real physical space regardless of our knowledge of any of the places along the way. So a network-based algorithm will fail to exhibit two important phenomena of route-finding: > Often you know roughly what direction to go in without having any idea of the details of the path, or even if the path is physically possible. > You can tell immediately that you don't know how to get to a place, just by verifying that you don't know the direction to that place. There are many other problems that a topological approach fails to treat adequately. Here are some of the problems we (Ernie Davis, Mark Zbikowski and I> have worked on: > How are metric facts, such as "A is about 5 miles from B" or “The direction from A to B is north" to be stored? > How are queries such as "Is it farther A to B than from A to C?" to be answered? from > Given a large set of objects and facts relating them, how do you find the objects that might be near some position? or with some orientation? Some of these problems have received more of our attention than others. In what follows, I will sketch our approach, the details of various algorithms and data structures, and the results we have so far. All of our solutions revolve around keeping track of the fuzzy coordinates of objects in various frames of reference. That is, to store -- metric facts about objects, the system tries to find, for each object, the ranges in which quantities like its X and Y coordinates, orientation and dimensions lie, with respect to convenient coordinate systems. The set of all the frames and coordinates is called a fuzzy map. We represent shapes as prototypes plus modifications. 13, 51 The domain we have used is the map of Yale University, from which most of my examples will be taken. To tasks: date we have written programs to do (1) Given a stream of metric relationships, create a fuzzy map of the objects involved. Research supported by NSF under contract MCS7803599 246 (2) Given a fuzzy map, test the consistency of a relationship or find the value of a term. (3) Given a fuzzy map, find objects with a position and orientation near some given value. (4) Plot a course around objects or through conduits discovered using (3). So far we have invested most of our effort in the study of task (21, what I described as "just looking" at the map to see what's true. This actually involves using hill climbing to see if a relationship can be satisfied, or to find the possible range of values of a term. So, in Figure 1, to answer the query 'What's the distance from Kline to Sterling in meters?" the system plunks down two points in the fuzz boxes of Kline and Sterling, and moves them as close together, then as far apart, as it can. To answer the query "Is Kline closer to Dunham than to Sterling?" it looks for a configuration of points from the fuzz boxes of Kline, Dunham and Sterling in which Kline is further from Dunham than Sterling. (Since it fails to find it, the answer to the query is "Yes.") The same hill-climbing algorithm is used for task (11, the assimilation of facts into a map. In this case, the object is to find the smallest and largest possible values of each "primitive term" involved in a relationship. (A primitive term is a quantity like (X A) or (LENGTH A) that characterizes an object's position, orientation or dimensions. More complicated terms, like (DIST A B), are functions of primitive terms.) The new, smaller range of a primitive term is then stored for future reference. This device, called fuzz constriction, suffices to capture many spatial facts However, it can happen that the range of a primitive term does not change at all after a new fact is assimilated, especially when the fact relates objects about which little is known. For example, if we tell the system that the orientation of Sterling Library is the same as the orientation of Becton Library, when it knows nothing about their orientations with respect to Yale, this new fact doesn't constrain them any further. In this case, mere fuzz constriction has failed to capture the new fact. The solution is to introduce a new frame of reference F, and to indicate that (ORIENT STERLING) = (ORIENT BECTON) = 0.0 in this frame, while the orientation of F is completely fuzzy (from 0 to 2pi) with respect to Yale. The machinery introduced so far enables us to retrieve characteristics of given objects. It is also important to be able to retrieve objects given their spatial characteristics (task (3)). For example, if you are trying to get from one place to another in a city, YOU will want to know what streets to use, i.e., how to find the nearest 'long porous objects" with approximately the right position and orientation. This is a job for k-d trees of the kind devised by Bentley. [l, 21 In these trees, a large set of objects is broken down into manageable chunks by an obvious generalization of binary search: repeatedly discriminate on each coordinate. An example is shown in Figure 2. The original version of k-d trees was designed to work on data bases in which all primitive terms have known values. In our application, most primitive terms can only be assigned ranges. To deal with this problem, we take the following tack: if a given attribute of an object is "very fuzzy" (e.g., its orientation is known only to lie between 0 and 2 pi>, then we do not index it on that attribute. But if it is only moderately fuzzy, 247 I I then we index it as though its value were the midpoint of its range. This requires that on retrieval we be willing to look around a little for objects fitting our requirements. That is, if we need to find a street with a given orientation, we keep relaxing the requirement until we find one. Obviously, a street found after many relaxations is only a plausible candidate, and must be proven to actually work; and the process must be terminated when it is unlikely to give worthwhile results. 02 route means depends on the density of the region to be traversed. If it is mostly open, then the problem is to plan to avoid obstacles; if it is mostly obstacle, then the problem is to plan to use conduits. Either way, the system must find objects with known properties (e.g., "open and pointing in the right direction" or "filled and lying in my way"). To summarize, our main results so far are these: representing space as a structure of multiple frames of reference, within which objects have fuzzy positions, is efficient and robust. In this context, assimilation is the process of constricting fuzz or creating new frames to capture a new fact. Route finding involves computing a fuzzy vector from where you are to where you want to be, then finding objects which can help or hinder your progress, and altering the plan to take them into account. Many problems still remain. The assimilation algorithm needs improvement. The route finder has not yet been completed or connected with the assimilation and retrieval algorithms. As yet we have not implemented a (simulated) route executor, although this is a high priority. AcknowledPements: Ernie Davis and Mark Zbikowski have helped develop many of the ideas in this paper, and made suggestions for improving the exposition. [ll 121 [31 [41 [51 [61 [71 181 REFERENCES Jon Bentley 1975 Multidimensional binary search trees used for associative searching, Comm. ACM 18, no. 9, PP. 509-517 Jon Bentley and Jerome Friedman 1979 Data I structures for range searching, Comnut. Surveys ll, no. 4, pp. 397-409 John Hollerbach 1975 Hierarchical shape description of objects by selection and modification of prototypes, Cambridge: MIT AI Laboratory Technical Report 346 Benjamin Kuipers 1978 Modeling spatial knowledge, Cognitive Science 2_, no. 2, p. 129 David Marr and H. Keith Nishihara 1977 Representation and recognition of the spatial organization of three dimensional shapes, Cambridge: MIT AI Laboratory Memo 416 Drew McDermott 1974 Assimilation of new information by a natural language-understanding w-w=-, Cambridge: MIT AI Laboratory Technical Report 291 Drew McDermott 1980 Spatial inferences with ground, metric formulas on simple objects, New Haven: Yale Computer Science Research Report 173 James Meehan 1976 The metanovel: writing stories by computer, New Haven: Yale Computer Science Research Report 74 This algorithm for finding objects that might have given characteristics is used by our route-finding programs. Exactly what finding a 248	1980	24
17	DESIGN SKETCH FOR A MILLION-ELEMENT NETL MACHINE Scott E. Fahlman Carnegie-Mellon University Department of Computer Science Pittsburgh, Pennsylvania 152 13 Abstract This paper describes (very briefly) a parallel hardware implementation for NETL-type semantic network memories. A million-element system can be built with about 7000 IC chips, including 4000 64K RAM chips. This compares favorably with the hardware cost of holding the same body of knowledge in a standard computer memory, and offers significant advantages in flexibility of access and the speed of performing certain searches and deductions. 1. Introduction In [l] I presented a scheme for representing real-world knowledge in the form of a hardware semantic network. In this scheme, called NETL, each node and link in the network is a very simple hardware processing element capable of passing single-bit markers through the network in parallel. This marker-passing is under the overall control of an external serial processor. By exploiting the parallelism of this marker-passing operation, we can perform searches, set intersections, inheritance of properties and descriptions, multiple-context operations, and certain other important operations much faster than is possible on a serial machine. These new abilities make it possible to dispense with hand-crafted search-guiding heuristics for each domain and with many of the other procedural attachments found in the standard Al approaches to representing knowledge. In addition to the ditficuity of creating such procedures, and the very great difficulty of getting the machine to create them automatically, I argue that the heuristic systems are brittle because they gain their efficiency by ignoring much of the search space. NETL, on the other hand, looks at every piece of information that might be relevant to the problem at hand and can afford to do so because it does not have to look at each piece of information serially. NETL has been viewed by many in the Al community as an interesting metaphor and a promising direction ‘for future research, but not as a practical solution to current Al problems because of the apparently impossible cost of implementing a large NETL system with current technology. The problem is not that the hardware for the nodes and links is too costly -- hundreds or even thousands of these elements can be packed onto a single VLSI chip. Rather, the problem is in forming new private-line connections (wires) between particular nodes and links as new information is added to the network. These connections cannot be implemented as signals on a single shared bus, since then all of 1 This research was sponsored by the Defense Advanced Research Projects Agency (DOD), ARPA Order No. 3507, monitored by the Air Force Avionics Laboratory Under Contract F33615-78-C-1551. The views and conclusions contained in this document are those of the author and should not be interpreted as representing the official policies, either expressed or implied, of the Defense Advanced Research Projects Agency or the US Government. the parallelism would be lost. Indeed, it is in the pattern of connecting wires, and not in the hardware nodes and links, that most of the information in the semantic network memory resides. A large switching network, similar to the telephone switching network, can be used in place of physical wires, but for a network of a million elements one would need the functional equivalent of a crossbar switch with 4 x IO’* crosspoints. Such a switch would be impossibly expensive to build by conventional means. In the past year I have developed a multi-level time-shared organization for switching networks which makes it possible to implement large NETL systems very cheaply. This interconnection scheme, which I tail a hashnet because some of its internal connections are wired up in a random pattern, has many possible uses in non-Al applications; it is described in its general form in another paper [2]. In this paper I will briefly describe a preliminary design, based on the hashnet scheme, for a semantic network memory with lo6 NETL elements. (An “element” in NETL is the combination of a single node and a four-wire link.) A million-element NETL system is lo-20 times larger than the largest Al knowledge bases in current use, and it offers substantial advantages in speed and flexibility of access. It is an open question whether a knowledge-base of this size will be adequate for common-sense story understanding, but a system of lo6 NETL elements should hold enough knowledge for substantial expertise in a variety of more specialized domains. In a paper of this length I will be able to sketch only the broad outlines of the design -- for a more complete account see [3]. The NETL machine itself, excluding the serial control computer, requires about 7000 VLSI chips, 4000 of which are commercial 64K dynamic RAM chips. (See the parts list, table 1.) As we will see later, with the same 64K memory technology it would require a comparable number of chips to store the same body of information in a conventional Planner-style data base, assuming that the entire data base is kept in a computer’s main memory and not on disk. So, far from being impossibly expensive, this scheme is quite competititve with standard random-access memory organizations. I am about to seek funding to build a million-element prototype machine within the next two or three years. The 64K RAM chips are not available today in sufficient quantities, and may be quite expensive for the next couple of years. The prototype machine will be designed so that 16K RAMS can be substituted if necessary, giving us a 256K element machine to use until the 64K RAM chips are obtained. 2. Requirements of NETL Figure 1 shows a basic NETL element as it was originally conceived. Commands from the control computer are received over the common party-line bus. The applicability of any command to a given element depends on the element’s unique serial number, on the state of 16 write-once flag bits which indicate what type of node or link the element represents, and on 249 the state of 16 read-write marker bits which indicate the current state of the element. These marker bits represent the short-term memory in the system. Also present are some number (4 in this design) of distinct link wires, and a node terminal to which link wires from other elements can be connected. Commands typically specify that all elements with a certain bit-pattern should send a one-bit signal across incoming or outgoing link wires, and that any element receiving such a signal should set or clear certain marker bits. It is also possible to address a command to a specific element, or to get any element with a certain marker pattern to report its serial number over the common bus. Using these commands, it is possible to propagate markers through the network Quillian-style or to control the marker propagation in any number of more precise ways. For details, see [l], especially section 2.3 and appendix A.1. In a million-element design, then, we have 4 sets of lo6 link wires to connect to IO6 node terminals, each by a private line. A link wire is connected to only one node terminal, but a node terminal may have any number of link wires attached to it. Unlike the telephone system, this network must support all 4 million connections simultaneously; once a connection is made, it becomes part of the system’s long-term memory, and a connection is seldom, if ever, released. As the system learns new information, new links are wired up one at a time, and this must be done without disturbing the connections already in use. If the same network hardware is to be used for different knowledge bases at different times, it must be possible to drop one set of connections and replace them with a new set. A few additional constraints will help us to separate interesting designs from uninteresting ones. If we want to obtain a roughly human-like level of performance in our knowledge base, the transmission of a bit from one element to another can take as long as a few milliseconds. Since answering a simple question -- for example, determining whether an elephant can also be a cabbage -w takes something like 20 to 60 basic marker-propagation cycles, a propagation time of 5 milliseconds gives a response time of .l to .3 seconds. This figure is independent of the total size of the network. This means that some degree of parallelism is essential, but that with microsecond-speed technologies there is room for some time-sharing of the hardware as well. New connections are established individually, and setting them up can take somewhat more time than simple propagations: humans are able to add only a few items to long-term memory per second. If an attempt to create a new connection should fail occasionally, nothing disastrous occurs -- the system simply skips over the link it is trying to wire up and goes on to the next free one. 3. The NETL Machine Design As mentioned earlier, the key problem here is to build a switching network for connecting link-wires to nodes. Since the four wires of a link are used at different times, we can think of this switch as four separate sub-netowrks, each with lo6 inputs, lo6 outputs, and with all lo6 connections operating at once. This network must be, in the jargon of network theory, a seldom- blocking network. It must be possible to add new connections one by one, without disturbing any connections that are already present, but some small chance of failure in establishing new connections can be tolerated. Once established, a connection must work reliably, and must be able to transmit one-bit signals in either direction. Note that by “connection” here I mean a setting of the network switches that establishes a path from a given input to a given output; the pysical wiring of the network is of course not altered during use. The basic concept used in this design is to build a 960 x 960 seldom-blocking switching network, then to time-share this network 1024 ways. The number 960 arises from packaging considerations; this gives us a total of 983,040 virtual connections, close enough to one million for our purposes. The 1024 time slices roll by in a regular cycle; a different set of switch settings is used during each slice. There are four sets of 1024 switch settings, corresponding to the four link-wire sub-netowrks. The bits describing the 4096 settings for each switch are stored in random access memory chips between uses. The NETL elements are likewise implemented in a time-shared fashion: 960 element units are implemented in hardware (four to a chip with shared bus decoders), and each of these element devices is time shared 1024 ways. Each NETL element exists in hardware only during its assigned time-slice: most of the time, it exists only as 32 bits of state in a memory chip. Let us begin by considering the construction of a 960 x 960 hashnet without time-sharing. The basic unit of construction is the 15way selector cell shown in figure 2a. This cell connects its input to any of its 15 outputs, according to the contents of a four-bit state register. A value of 0 indicates that the cell is currently unused and that its input is not connected to any output. A single integrated circuit chip can easily hold 15 of these selector cells; the corresponding outputs from each cell are wired together internally, as shown in figure 2b. With assorted power and control lines, this 15 x 15 switching element requires a 48.pin package. To build a 960 x 960 seldom-blocking network out of these elements, we arrange them in four layers with 1920 selector cells (128 chips) in each. (See figure 3.) The outputs of each layer are wired to the inputs of the next layer with a fixed but randomly chosen pattern of wires. Each of the input terminals of the hashnet is wired to 2 selector cells in the first layer; each of the outputs of the hashnet is wired to 2 outputs lines from the last layer of cells. Initially all of the selector cells are in the non-busy slate. As paths through the network are set up, each one uses up one of the selector cells in each layer. Note, however, that half of the selector cells remain unused even when all 960 connections are in place; this spare capacity ensures that the network will have a low chance of blocking even for the last few connections. To set up a new connection from a given input to a given output, we first broadcast a marking signal through the network from the input to all of the selector cells and outputs that can be reached. Only non-busy cells play a role in this process. If this signal reaches the desired output, one of the marked paths is traced back toward the source, with the selector cells along the way being set up appropriately. These cells become busy and will not participate in any other connections. Since the inter-layer wiring of the network is random, and since we are using many fewer switches than are needed for a strictly non-blocking network, we cannot guarantee that a desired connection can be found. We can guarantee that the probability of being unable to find a desired connection is very small. In simulation tests of this design, 100 complete sets of connections were attempted -- 10,000 connections in all -- with only 2 failures. As noted earlier, an occasional failure to find a connection is not disastrous in the NETL application; we just try again with a different link. Instead of using four layers of selector cells, it is possible to get the same effect using only one layer. The output wires of this layer are randomly shuffled and fed back to the inputs; these wires also go to the network’s ouput terminals. Four sets of switch settings are used, corresponding to the four layers of the original network. The signal bits are read into this layer of cells and latched. Using the first-layer switch settings, they are sent out over the appropriate output wires and shuffled back to the inputs where these bits are latched again. Then the second layer setup is used and the signal bits are shuffled again. After four such shuffles, the bits are in the right place and are read from the network’s outputs. We have traded more time for fewer switches and wires; the same number of setup bits are used in either case. To go from a thousand connections to a million, we time-share this network 1024 ways. This requires two modifications to the network. First, instead of having only 4 bits of state for each selector cell (or 16 bits if the shuffling scheme is in use), we need a 25CI different set of bits for each time slice. These bits are read from external memory chips; 4 bits of setup are read in through each selector cell input, followed by the signal bit. Second, since the NETL elements are permanently tied to their assigned time-slices, we need some way to move the signal bits from one time-slice to another. This operation is carried out by time-slice shifter chips. Each of these devices is essentially a 1024-bit shift register. During one cycle of time-slices this shift register is loaded: during each slice, a signal bit is received along with a IO-bit address indicating the slice that the signal bit is to go out on. The address governs where in the shift register the signal bit is loaded. On the next cycle of slices, the bits are shifted out in their new order. An entire layer of 1920 time-shifters is needed (packed 5 to a chip), along with memory chips to hold the lo-bit time-slice addresses. The chance of blocking is minimized if these are placed in the center of the network, between the second and third layers of cells. Some additional chance of blocking is introduced by addition of the shifters to the network, but not too much. In our simulations of an almost-full network with time sharing, we encountered 37 blocked connections in 110,000 attempts. 4. Cost and Performance As can be seen from the parts list, most of the cost of the NETL machine is in the memory chips. In order to keep the number of chips below 10,000 -- a larger machine would be very hard to build and maintain in a university research environment -- 64K dynamic RAM chips have been used wherever possible in the design. The memory associated with the element chips must be either 2K x 8 static RAMS or fast 16K dynamic RAMS for timing reasons. In fact, the limited bit-transfer rate of the memory chips is the principal factor limiting the speed of the network; if 16K x 4 chips were available, the system could be redesigned to run four times as fast. As it currently stands, the system has a basic propagation time of about 5 milliseconds. This is the time required to accept lo6 bits and to steer each of these to its independent destination. This assumes the use of 2K x 8 static RAMS for the element memories and allows for a rather conservative page-mode access time of 200 nanoseconds for the 64K RAM chips. (In page mode, half of the address bits remain constant from one reference to the next.) The 256K element version, using 16K RAM chips, should have a propagation time of 1.25 milliseconds. Since the million-element machine contains 4000 64K RAM chips, the parts cost of the machine is tied closely to the price of these chips. Currently, if you can get 64K RAMS at all, they cost over $100, but the price is likely to drop rapidly in the next two years. If we assume a price of $20 for the 64K RAMS and $10 for the 2Kx8 RAMS, we get a total of roughly $100,000 for the memory chips in the machine. It is also hard to know what price to assign to the three custom chips in the design. An initial layout cost of $150,000 for all three would probably be reasonable, but this cost would occur only once, and the chips themselves should be easy to fabricate. We have not yet thought hard about board-level layout, but rough calculations suggest that the machine would fit onto about 52 super-hex boards of two types. Two standard equipment racks ought to be sufficient for the whole machine. For comparison, it might be worthwhile to calculate the cost of storing the same information in a Planner-style data base in the memory of a serial machine. To make the comparison fair, let us assume that the entire data base is to be kept in main memory, and not paged out onto disk. To store the information in the network itself, in the simplest possible form, would require 80 million bits of memory: 4 million pointers of 20 bits each. This would require 1250 64K RAM chips. A simple table of pointers, of course, would be very slow to use. If we add back-pointers, the figure doubles. If we add even the most limited sort of indexing structure, or store the entries in a hash table or linked list, the arnount of memory doubles again. At this point, we have reached 4000 64K RAM chips, the same number used in the NETL machine. The moral of the story would seem to be that the greater power and flexibility of the NETL organization can be had at little, if any, extra cost. Acknowledgements Ed Frank and Hank Walker provided me with much of the technical information required by this design. The network simulations were programmed and run by Leonard Zubkoff. References ill Fahlman, S. E. NETL: A System for Representing and Using Real-World Knowledge. MIT Press, Cambridge, Mass., 1979. PI Fahlman, S. E. The Hashnet Interconnection Scheme. Technical Report, Carnegie-Mellon University, Department of Computer Science, 1980. [31 Fahlman, S. E. Preliminary Design for a Million-Element NETL Machine. Technical Report, Carnegie-Mellon University, Department of Computer Science, 1980. (forthcoming). Gable 1: Parts List Device Type Custom 15 x 15 Selector (48 pin) Custom 5x 1 K Time Shifter (16 pin) Custom 4x NETL Element Chip (48 pin) 64K Dynamic RAM (for selectors) 64K Dynamic RAM (for shifters) 2Kx8 Static RAM (for elements) Total Device Count Number 128 240 2176 1920 2160 7008 251 I I C%Ol Computer INPUT SELiCT BITS Marker bits (16) Node Terminal Figure 1: NETL Element IN 1 IN 2 IN 3 IN 4 OUT 1 OUT 2 OUT 3 OUT 4 Figure 2A: The Selector Cell Figure 2B: Selectors on 15x15 Chip (Drawn as 4x4 chip.) Link Wires Node Terminals Figure 3: The Basic Washnet A rrangement (simplified) 252	1980	25
18	PERCEPTUAL REASONING IN A HOSTILE ENVIRONMENT Thomas D. Garvey and Martin A. Fischler Artificial Intelligence Center SRI International, Menlo Park, CA 94025 ABSTRACT The thesis of this paper is that perception requires reasoning mechanisms beyond those typically employed in deductive systems. We briefly present some arguments to support this contention, and then offer a framework for a system capable of perceptual reasoning, using sensor- derived information, to survive in a hostile environment. Some of these ideas have been incorporated in a computer program and tested in a simulated environment; a summary of this work and current results are included. I INTRODUCTION Living organisms routinely satisfy critical needs such as recognizing threats, potential mates, food sources, and navigable areas, by extracting relevant information from huge quantities of data assimilated by their senses. How are such "relevant" data detected? We suggest that a reasoning approach that capitalizes on the goal-oriented nature of perception is necessary to define and recognize relevant data. Perception can be characterized as imposing an interpretation on sensory data, within a context defined by a set of loosely specified models. The ability to select appropriate models and match them to physical situations appears to require capabilities beyond those provided by such "standard" paradigms as logical deduction or probabilistic reasoning. The need for extended reasoning techniques for perception is due to certain critical aspects of the problem, several of which we summarize here: * The validity of a perceptual inference (interpretation) is determined solely by the adequacy of the interpretation for successfully carrying out some desired interaction with the environment (as opposed to verification within a "closed" formal axiomatic system). -------- * This work was supported by the Defense Advanced Research Projects Agency under Contracts MDAgOT-79- C-0588, F33615-77-C-1250, and F33615-80-C-1110. Since it is impossible to abstractly model the complete physical environment, the degree to which purely abstract reasoning will be satisfactory is limited. Instead, perception requires tight interaction between modeling/hypothesizing, experimenting (accessing information from the environment), and reasoning/verifying. Reasoning processes that embody concepts from physics, geometry, topology, causation, and temporal and spatial ordering are critical components of any attempt to "understand" an ongoing physical situation. Explicit representations appropriate to these concepts are necessary for a perceptual system that must provide this understanding. These representations are incommensurate and it is not reasonable to attempt to force them into a single monolithic model. There is typically no single, absolutely correct interpretation for sensory data. What is necessary is a "maximally consistent" interpretation, leading to the concept of perception as an optimization problem [l, 21 rather than a deductive problem. Research in perception and image processing at SRI and elsewhere has addressed many of these issues. An early effort focused upon the goal- directed aspect of perception to develop a program capable of planning and executing special-purpose strategies for locating objects in office scenes [31* Research addressing interpretation as an optimization problem includes [l, 2, 41. Current research on an expert system for image interpretation [ 5] h as also considered the strategy-related aspects of determining location in situations involving uncertainty. The most recent work (at SRI) on perceptual reasoning has addressed the problem of assessing the status of a hostile air-defense environment on the basis of information received from a variety of controllable sensors [6]. This work led us to attempt to formulate a theory of perceptual reasoning that highlighted explicit reasoning processes and that dealt with those aspects of perception just described. In the following section, we will use this work as a vehicle to illustrate a paradigm for perceptual reasoning. 253 II PERCEPTUAL REASONING IN A SURVIVAL SITUATION The specific problem addressed was to design a system able to interpret the disposition and operation (i.e., the order of battle or OB) of hostile air-defense units, based on information supplied by sensors carried aboard a penetrating aircraft [6]. The situation may be summarized as follows. A friendly aircraft is faced with the task of penetrating hostile airspace en route to a target behind enemy lines. Along the way, the aircraft will be threatened by a dense network of surface-to-air missiles (SAMs) and antiaircraft artillery (AAAs). The likelihood of safe penetration and return is directly related to the quality of acquired or deduced information about the defense systems. Partial information is furnished by an initial OB, listing known threats at, say, one hour before the flight. Additional knowledge is available in the form of descriptions of enemy equipment, typical deployments, and standard operating procedures. Since the prior OB will not be completely accurate, the information must be augmented with real-time sensory data. The OB forms the starting point for this augmentation. The explicit goal of the overall system is to produce and maintain an accurate OB, detecting and identifying each threat prior to entering its lethal envelope. The density of threats means that this goal will result in conflicting subgoals, from which selection must then be made to ensure that critical data will be received. This must be accomplished by integrating data from imperfect sensors with prior knowledge. The paradigm that was developed for this task is summarized below: (1) Available knowledge is used to create an hypothesized OB that anticipates the developing situation. (2) A plan that attempts to allocate sensors to detect or verify the presence of threats, in an optimal way, is constructed. Sensors are then allocated and operated. (3) Information returned from the sensors is interpreted in the context established by the anticipated situation. Interpretations modify the current OB, and the process is iterated. We will briefly discuss each of these steps. B. Experimentation (Accessing Information from the Environment) The goal of this step is to access information needed to detect or verify the presence of threats inferred in the anticipation step, but not available in the "internal" knowledge base of the system. In general, it might be necessary to define and execute one or more experiments to extract this needed information from the environment. In the more limited context of model instantiation by "passive" sensing, the problem reduces to that of allocating sensor resources to maximize the overall utility of the system; sensing is a specific instance of the more general process of experimentation. First the list of data-acquisition goals is ordered, based on the current state of information about each threat and its lethality. The allocator attempts to assign (a time sliced segment of) a sensor to satisfy each request based on the expected performance of the sensor for that task. Sensor detection capabilities are modeled by a matrix of conditional probabilities. These represent the likelihood that the sensor will correctly identify each threat type, given that at least one instance thereof is in the sensor's field of view. This matrix represents performance under optimal environmental conditions (for the sensor) and is modified for suboptimal conditions by means of a specialized procedure. This representation is compact and circumvents the need to store complete, explicit models describing sensor operation in all possible situations. Similar models describe each sensor's identification and location capabilities. The sensor models are used to compute the utility of allocating each sensor to each of the highest priority threats. These utilities form the basis for the final allocation, which is carried out by a straightforward optimization routine. At the same time, the program determines how the sensor should be directed (for example, by pointing or tuning). Appropriate control commands are then sent to the simulated sensors. C. Interpretation (Hypothesis Validation; Model Instantiation) In this phase, the program attempts to interpret sensor data in the context of threats that were anticipated earlier. It first tries to 254 determine whether sensor data are consistent with specifically anticipated threats, then with general weapon types expected in the area. Since sensor data are inherently ambiguous (particularly if environmental conditions are suboptimal), this step attempts to determine the most likely interpretation. Inference techniques used for interpretation include production rule procedures, probabilistic computations, and geometric reasoning. Production rules are used to infer probable weapon operation (e-g., target tracking, missile guidance), on the basis of such information as past status, environmental conditions, and distance from the aircraft. Probabilistic updating of identification likelihoods is based on the consistency of actual sensor data with expected data, and on agreement (or disagreement) among sensors with overlapping coverage. Geometric reasoning introduces a concept of global consistency to improve identification by comparing inferred identifications and locations of threat system components with geometric models of typical, known system deployments. The interpretation phase brings a great deal of a priori knowledge to bear on the problem of determining the most likely threats the sensors are responding to. This results in much better identifications than those produced by the sensors alone. Confident identifications are entered into the OB and the entire process is continued. D. Performance An experimental test of the system, using a simulated threat environment, allowed a comparison between two modes of operation--an "undirected" mode and one based on perceptual reasoning. A scoring technique that measured the effectiveness with which the system detected, identified, and located hostile systems in a timely fashion was used to grade performance. The ability of the perceptual reasoning system to use external knowledge sources effectively, and to integrate information from multiple sensors, produced superior capabilities under this measure. These capabilities showed themselves even more prominently in situations where environmental conditions tended to degrade sensor performance, rendering it critical that attention be focused sharply. III DISCUSSION Our approach to perceptual reasoning suggests_ that the problem of perception actually involves the solution of a variety of distinct types of subproblems, rather than repeated instances of the same general problem. The system we described utilizes a nonmonolithic collection of representations and reasoning techniques, tailored to specific subproblems. These techniques include both logical deduction and probabilistic reasoning approaches, as well as procedures capable of geometric reasoning and subjective inference. We have discussed several key aspects of the general problem of perceptual reasoning, including the assertion that perception is goal oriented, and inductive and interpretative rather than deductive and descriptive; that because complete modeling of the physical world is not practical, nexperimentationn is a critical aspect of perception; and finally, that multiple representations and corresponding reasoning techniques, rather than a single monolithic approach, are required. The specific system discussed above constitutes an attempt to address the reasoning requirements of perception in a systematic way and, to our knowledge, represents one of the few attempts to do so. While systems that truly interact with the physical world in an intelligent manner will certainly assume a variety of forms, we believe they will all ultimately have to resolve those aspects of the problem that have been described here. REFERENCES 1 . 2. 3. 4. 6. M. A. Fischler and R. A. Elschlager, "The Representation and Matching of Pictorial Structures," IEEE Transactions on Computers, Vol. C-22, No. 1, pp. 67-92 (January 1973) H. G. Barrow and J. M. Tenenbaum, "MSYS: A System for Reasoning About Scenes," Artificial Intelligence Center Technical Note No. 121, SRI International, Menlo Park, California (1976) T. D. Garvey, "Perceptual Strategies for Purposive Vision," Artificial Intelligence Center Technical Note No. 117, SRI International, Menlo Park, California (1976) A. Rosenfeld, "Iterative Methods in Image Analysis," Pattern Recognition, Vol. 10, No. 4, pp- 181-187 (1978) R. C. Bolles et al., "The SRI Road Expert: Image-to-Database Correspondence," Proc. DARPA Image Understanding Workshop, Scienr- Applications, Inc., Report No. SAI-79-814-WA, pp. 163-174 (November 1978) T. D. Garvey and M. A. Fischler, "Machine- Intelligence-Based Multisensor ESM System," ' Technical Report AFAL-TR-79-1162, Air Force Avionics Laboratory, Wright-Patterson AFB, OH 45433 (Final Report DARPA Project 3383) (October 1979) 255	1980	26
19	OVERVIEW OF AN EXAMPLE GENERATION SYSTE?l Edwina L. Rissland Elliot M. Soloway Department of Computer and Information Science University of Massachusetts Amherst, MA 01003 ABSTRACT This paper addresses the process of generating examples which meet specified criteria: we call this activity CONSTRAINED EXAMPLE GENERATION (CEG). We present the motivation for and architecture of an existing example generator which solves CEG problems in several domains of mathematics and computer science, e.g., the generation of LISP test data, simple recursive programs, and piecewise linear functions. II THE CEG MODEL --- From protocol analyses of experts and novices working CEG problems in mathematics and computer science, we developed the following description of the CEG task [lOI. When an example is called for, one can search through one's storehouse of known examples for an example that can be JUDGEd to satisfy the desiderata. If a satisfactory match is found, then the problem has been solved through SEARCH and RETRIEVAL. If a match is not found, one can MODIFY an existing example, judged to be close or having the potential for fulfilling the desiderata. If generation through modification fails, one often switches to another mode in which one CONSTRUCTS an example by instantiating certain models and principals or combining more elementary exemplars. The CEti model we use thus consists of processes that RETRIEVE, JUDGE, MODIFY and CONSTRUCT examples. III THE CEG -- SYSTEM The CEG system described here is written in LISP and runs on a VAX 11/780. In addition to solving CEG problems concerning data and simple programs in the LISP domain [I1 1, it is being used to solve CEG problems in a number of other domains 1121: the generation of descriptions of scenes for use in conjunction with the VISIONS scene interpretation system II21; the generation of action sequences in games; and the generation of piecewise linear functions. The flow of control in the Exanple Generator is directed by an EXECUTIVE process. In addition, there is: (1) the RETRIEVER which searches and retrieves examples from a data base of examples; (2) the MODIFIER which applies modification techniques to an example; (3) the CONS'ER which instantiates general "model" examples, such as code 'lte.mplates" (C91, 1161)~ (4) the JUDGE which determines how well an example satisfies the problem desiderata; (5) the AGENDA-KEEPER which maintains an agenda of examples to be modified, based for instance on the degree to which they meet the desiderata or possess flepistemological" attributes (C81, 191). The components use a common knowledge base consisting of two parts: a "permanent" knowledge base which has an Examples-space 191 containing known examples, and a temporary knowledge base which -contains information gathered in the solution of a specific CEG problem. In the Examples-space, an example is represented by a frame, which contains information describing various attributes of that example, e.g., epistemological class, worth-rating. Examples are linked together by the relation of nconstructional derivation," i.e., Example 1 ---> Example2 means that Exanplel is used in the construction of Example2. The temporary knowledge 256 base contains such,information as evaluation data generated by the JUDGE and proposed candidate examples created from known examples by the MODIFIER. IV AN EXAMPLE OF A CEG PROBLEM ----VP In the context of examples of LISP data elements, an example of a simple CEG problem would be the followirg: Give an example of a LISP list of length 3 with the depth of its first atom equal to 2. (Examples teaching.) such as this are needed when debugging and Suppose the permanent knowledge base only contains the lists RETRIEVAL phas;hasf;ils), the system enters the MODIFICATION (A B C), (0 11, (A), ( ); Since the list (A B C) the first two lists are satisfies two of the'three constraints, and thus has standard ttreferencelt examples, the third, a "start-up" example, and the fourth, a known "counter example," i.e., an a higher "constraint-satisfaction-count" example which often than the is handled other examples, it is incorrectly by programs. Since the knowledge base does not contain an example placed which completely satisfies the desiderata (i.e., the as the top-ranking candidate for MODIFICATION by the AGENDA-KEEPER and modifications are tried on it first. existing example is modified to create a new example that is no longer deficient with respect to the constraints. For example, if there were more than one unsatisfied constraint, one could attempt to rectify a candidate example along each of the unsatisfied dimensions. Consider the following problem: Give an example of a list of O’s and l's, which is longer than 2, and which has at least one element deeper than depth 1. One could modify the list (0 1) to meet the unsatisfied second and third constraints by adding at least one more element, say another 1 to give the list (0 1 11, and then modify this new list by adding parens around 0r.e of the elements. Alternatively, one could add parens and then fix the length, or even do both 'at once' by appending on an element such as (1) to the original list (0 1). In this example, there are many ways to modify the original list to meet the constraints, and the order of modification does not much matter. One can 'ldivide-and-conquer'l the work to be done in a GPS-like manner of difference-assessment followed by However, difference-reduction in other cases the order of difference reduction since matters greatly. For instance, consider the constraints are independent [51. the problem: Give an example of a list of length 5 with an embedded sublist of length 2. The candidate example, (A B C) is analysed and found to be lacking in one respect, namely, the depth of its first atom, A, must be made deeper by 1. The sys tern accomplishes this by adding parentheses around the atom A to create the new example ( (A) B C). This list meets all the constraints specified; as it is a solution to the problem, it is entered into the Examples-space as a new example constructionally derived from (A B C) . Thus * if the following problem is asked of the system, Give an example of a list of length 3, the depth of the first atom is 2, and the depth of the last is 3. the system can use the newly constructed its attempt to satisfy the new problem. example in Suppose one is working with the list (A B C>. If one first rectifies the length by adding two more elements, such as 1 and 2, to create the list (A B C 1 2 ), and then modifies this list to have the needed embedded list by “making-a-group” of length 2 around any of the first four elements, say to arrive atthelist(ABC(12) ),one has also modified the length as a side-effect of the grouping modification, i.e., one has messed up the length constraint. In other circumstances, it possible to set up totally contradictory constraints where satisfaction of one precludes satisfaction of the other. Thus a purely GPS approach is not sufficient to handle the complexity of constraint interaction. We are currently investigating the constraint interaction problem as well as issues concerning maintenance of agendas. For example, we are looking to employ planning-type control mechanisms for dealing with the constraint interaction problem. cc 131, C31) VI THE LARGER PICTURE --__I_ V THE HANDLING OF CONSTRAINTS - An application of our CEG system is in an The hand 1 ing of constraints is especially intelligent computer-assisted instruction tutoring important in the MODIFICATION phase where an environment . We are currently building a tutor to 257 teach students about programming languages such as LISP and PASCAL (C143, C151). In this context, CEC will serve the tutor in two ways: (1) it will generate examples to the specifications set by the tutor; and (2) it will evaluate student generated examples for the tutor. The same JUDGE used by CEG can be used to evaluate a student's example and help track down his misconceptions and bugs through analyses of differences between what the tutor requested and what the student acutually generated. In the future, we also plan to incorporate adaptation into the system. For example, the system can keep track of the performance of the various example ordering functions and choose the one that has the best performance. Also, we plan to apply hill-climbing techniques to the modifying processes themselves. That is, since there are alternative ways to massage and modify an example, those routines which lead to the most succeses should eventually be selected first. Adaptation on the modification techniques will be particularly important if the system is to be able to improve its performance, and thus "learn" from its own experience. The current implementation is only a “first-pass” and does not capture the richness of the CEG model. Nonetheless, we feel that it has demonstrated the utility of this model and we feel that subsequent implementations incorporating the characteristics of additional task domains should provide us with a rich environment to continue our investigation into the important process of example generation. Lll [21 c31 c43 c51 REFERENCES Bledsoe, W. (1977) A Maximal Method for Set Variables in Automatic Theorem Proving, Univ. of Texas at Austin, Math Dept. Memo ATP-33. Hanson, A. and Riseman E. (1978) VISIONS: A Computer System for Interpreting Scenes, in Computer Vision Systems, Hanson and Riseman, Eds. t Academic Press, New York. Hayes-Roth, B., ar,d F. Hayes-Roth (1978) Cognitive Process in Planning, Rand Report R-2366-ONR, The Rand Corporation, CA. Lakatos, I. ( 1963) Proofs and Refutations, - - British Journal for the Philosophy of Science, Vol. 19, May 1963. Also published by Cambridge University Press, London, 1976. Newell, A., Shaw, J., and Simon, H. (1959) Report on a General Problem-Solving Program. Proc. of the International Conference on Information Processing. UNESCO House, Paris. 163 Polya, G . (1973) now To Solve St, Second Edition, Princeton UXJ. Press, N.J, E7l Polya, G. (1968) Mathematics and Plausible Reasoning, Volunes I and II, S&?&d Edition, Princeton Univ. Press, N.J. C83 Rissland (Michener) , E. (1978a) Understanding Understanding Mathematics, Cognitive Science, Vol. 2, No. 4. [91 Rissland (Michener), E. (1978b) The Structure of Mathematical Knowledge, Technical Report No. 472, M.1.T Artificial Intelligence Lab, Cambridge. Cl01 Rissland, E. (1979) Protocols of Ex mple Generation 1 internal report, M.I.T., Cambr idge. [ill Rissland, E. and E. Soloway (1980) Generating Examples in LISP: Data and Programs, COINS Technical Report 80-07, Univ. of Mass, (submitted for publication). Cl21 Rissland, E., Soloway, E. , O’Connor, S., Waisbrot, S., Wall, R. , Wesley, L., and T. Weymouth (1980) Examples of Exaple Generation using the CEG Architecture. COINS Technical Report, in preparation. cl31 Sacerdoti, E. (1975) The Nonlinear Nature of Plans, Proc. 4th. Int. Joint Conf. Artificial Intelligence, Tbilisi, USSR. Cl41 Soloway. E. ( 1980) The Development and Ev aluafiion of Instructional Strategies for an Intelligent Computer-Assisted Instruction System, COINS Technical Report 80-04, Univ. of Mass., Amherst. El51 Soloway, E., and E. Rissland (1980) The Representation ar,d Organization of a Knowledge Base About LISP Programming for an ICAI System, COINS Technical Report 80-08, Univ of Mass., in preparation. [161 Soloway, E., and Woolf, B. ( 1980) Problems, Plans and Programs, Proc. of the ACM Eleventh SIGCSE Technical Symposium, Kansas City. 258	1980	27
20	STRUCTURE COMPARISON AND SEMANTIC INTERPRETATION OF DIFFERENCES* Wellington Yu Chiu USC Information Sciences Institute 4676 Admiralty Way Marina de1 Roy, California 90291 ABSTRACT Frequently situations are encountered where the ability to differentiate between objects is necessary. The typical situation is one in which one is in a current state and wishes to achieve a goal state. Abstractly, the problem we shall address is that of comparing two data structures between the two structures. work well in determining all of addressing the semantic syntactic issues. and determining all Current comparison differences techniques differences but fall short We address this gap by applying Al techniques that yield semantic interpretations of differences. I INTRODUCTION One frequently encounters situations where the ability to differentiate between objects is necessary. The typical situation is one in which one is in a current state and wishes to achieve a goal state. Such situations are encountered under several guises in our transformation based programming system research [I, 2, 31. A simple case is one in which we are given two program states and need to discover the changes from one to the other. Another case is one in which a transformation we wish to apply to effect a desired change does not match in the current state and one wishes to identify the differences. An extension of this second case is the situation where a sequence of transformations, called a development, is to be applied to (replayed on) a slightly different problem. Abstractly, the problem we shall address is that of comparing two data structures and determining all differences between the two structures. Current comparison techniques determining all syntactic differences but fall short the semantic issues. situations, comparisons address this gap by applying interpretations of differences. work well in of addressing In the replay and state differencing must be more semantically oriented. We This paper describes a semantic based differencer Al techniques that yield semantic part of my thesis: the design and its ongoing implementation. of a * This work woa rupportrd by Nitional Scknce the viowr cwpremed are thaw of thr author. Foundation Grant MCS 7683880. II AN EXAMPLE -- The following example is presented to show the types of information used to infer and deduce the semantic differences. Below are the before and after program states from a transformational development [ 1 J. BEFORE : uhile there exists character in text do 1 if character is linefeed 2 then replace character in text by space:3 uhile there exists character in text do i f Pfcharacter) : then remove character from text: 6 AFTER: while there exists character in text do begin ;: if character is Iinefeed then replace character in text by space: : if character in text then e i f P (character) f then remove character from text end: h9 The after state above was produced from the before state via application of a transformation but the following explanations of differences were generated without knowledge of the transformation. - The current syntactic differencing techniques [4, 5, 6, 7, 8, 9, lo] typically explain differences in the following terms: For BEFORE: Delete while in line 4 Delete there in line 4 For AFT&I:’ ’ Insert if in line f Insert then in line f . * . - A higher level syntactic explanation is achieved by generalizing and combining the techniques for syntactic differencing to explain differences in terms of embeds, extracts, swaps and associations, in addition to inserts and deletes, and by incorporating syntactic information about the structures being compared [3] The second loop is coerced into a conditional. 259 The conditional is embedded into the first loop, trees for 23, 5,6 and c,d,f,g are loop generators. - The proposed explanation of the semantic difference is: Loops merged. The following is the derivation of the syntactic explanation. It - Infer the similarity of 5,6 to e,f,g from the syntactic equivalence of 5,6 to f,g. 5,6 embedded in a test for generator consistency inferred from semantic knowledge of loop merging. - Conclusion: Loops merged. is presented to show the mechanisms upon which the semantic differencer will be based. - Syntactically, 2,3 (first loop body of the before state) is equivalent to c,d (part of the loop body of the after state) and 5,6 is equivalent to f,g. - Infer composite structure 1,2,3 similar to composite structure a,b,c,d,e,f,g,h based on 2,3 being equivalent to c,d. - tnfer an embed of composite structure 4,5,6 into the composite structure 1,2,3 to produce a,b,c,d,e,f,g. The support for this inference comes from 5,6 being equivalent to f,g, the adjacency of 1,2,3 to 4,5,6, and the adjacency of c,d to f,g. - Infer coercion of loop 4,5,6 to conditional e,f,g based on 5,6 being equivalent to f,g, and the similarity of the loop generator to the conditional predicate. - Conclude second loop embedded in the first loop. Our current syntactic differencer produces this type of difference explanation. It exte_nds the techniques currently The explanation generated by our syntactic clifferencer is not plausible because i? doesn’t make sense to transform a loop into a conditional only to embed this new conditional into a similar loop. The following is the desired explanation: The body of the decond loop is embedded in a conditional that tests for loop generator consistency. This is done without changing the functionality of the body. The two adjacent loops can now be merged subject to any side effects, caused by the first loop body, that will not be caught by the loop generator consistency check around the second loop body. Ill DESIGN OF THE SEMANTIC BASED DIFFERENCER -e--p We start by defining relations (profiles) on objects, where objects are the substructures of the structures we are comparing (see Appendix A). The information provided by this profile consists of: - Sequence of nonterminals from the left hand side of productions used in generating the context tree of the substructure. A context tree (i.e. super tree) is that part of the parse tree with the substructure part deleted. available by imposing structure on the text strings being advance, the explanations fall short of the desired semantics of compared, thereby making use of structural constraints and the the changes. In the SCENARIO section below, an explanation extra structural dimensions in its analysis [3]. Despite this - The sequence of productions used in the derivation of the substructure, given the context tree above as the starting point. Positional information. from the proposed semantic based differencer is presented. The domain we are dealing with is one where changes are made for the following reasons: 1. To optimize code Our syntactic differencer makes use of the information provided by the object profile to determine base matches. The techniques used in the differencer to determine base matches are combinations of the common unique, common residual and common super tree techniques described in [3, 4, 6, 7, lo]. 2. TO prepare code for optimization. Below is a brief description of the common unique and common residual techniques for linear structures. Within such a domain, we can use the constraints on the semantics of changes to derive the semantic explanation “loop merged” and at the same time rule out the syntactic explanation. - Longest Common Subsequence (LCS): The main coniern with LCS is that .of finding the minimal edit distance between two linear structures. The only edit ooerations allowed are: delete a substructure, Building on the mechanisms that generated the derivation above, or insert a substructure [6, 81. For the above the following is a proposed derivation of the comparison that example, the LCS is: 1,2,3,5,6 matching a,c,d,f,g. yield the semantic interpretation “loops merged”. - Common Unique KU): The key to this technique the - Syntactically, 2,3 (the first loop body) is equivalent to c,d and 5,6 (the second loop body) is equivalent to f,g. The context trees (i.e. super trees) containing 2,3 and 5,6 and the context tree for c,d,f,g are the same. Infer FACTORING of context trees with supports being the 2 to 1 mapping of substructures and the equivalence of context trees. use of substructures that are common to both structures and unique within each as anchors [4]. For the above example, the common unique substructures are: linefeed, replace, space, P, remove. We then build on these base matches by inferring matches of - Infer loop merge from the fact that the context substructures that contain substructures that are matches. An example of this is inferring that 1,2,3 is similar to a,b,c,d,e,f,g,h 260 from the assertion that 2,3 matches c,d. There are two types Of inferred matches: those without syntactic boundary conflicts and those with conflicts. Syntactic boundary conflicts result from embedding, extracting or associating substructures. The third type of profile is one that describes the relationship between substructures within a given structure. Considerations here are: adjacency, containment, and relative positioning. There are several semantic rules that describe a majority of structure changes. Some are: factoring, distributing, commuting, associating, extracting, embedding, folding, and unfolding. A oartial description of the factor semantics can be found in Appendix A. Factors currently considered by our semantic rules are: support for matches, the generating grammar, object profiles, and relations between substructures of the same structure. With each set of examples tried, we add to our set of semantic rules, and our intuitive guess, given our domain of changes due to optimization or preparation for optimization, is that this set will be fairly small when compared to the set of transformations needed in a transformation based programming system. IV A SCENARIO -- Our syntactic differencer makes use of structural information. For LISP programs it know about S-expressions. For programs written in our specification language, differencing is performed on the parse trees. The differencer first tries to isolate differences into the smallest composite substructure containing all changes. With this reduced scope, the differencer uses the common unique technique to assert relations on content base matches. In our example, substructure 2,3 is equivalent to c,d and substructure 5,6 is equivalent to f,g. Once all of these possible assertions have been made, we use them as anchors to assert relations based on positional constraint and content matches (residual structure matches). This residual technique works well as a relaxation of the uniqueness condition in the common unique requirement, and acts as a backup in case no substructures are common to both structures and unique within each. The super tree technique is used as a default when neither of the above techniques applies. The intuitive explanation for this third technique has to do with both objects occupying the same position with respect to a common super tree. With the super tree technique, content equivalence is relaxed. With the two asserted relations regarding substructures 2,3 being equivalent to c,d and 5,6 being equivalent to f,g, we now infer that 4,5,6 is similar to e,f,g because 5,6 is common unique with f,g and without conflicting evidence (i.e. boundary violations) the assertion is made. Once made, further analysis of this reduced scope shows relationships between the loop generator 4 and the conditional predicate e. Since we are given, via the super tree technique, that 1,2,3,4,5,6 matches a,b,c,d,e,f,g,h we assert the inference 2,3 to c,d even though conflicts due to boundary violations arise. The boundary violation in this instance is the mapping of two substructures into one (i.e. the segmental or n-m problem). Given that we want to produce all plausible explanations, we assert that 1,2,3 is similar to a,b,c,d,e,f,g,h because 2,3 and c,d are common unique matches. With this assertion, we could with our Embed Semantics say that the second loop is embedded in the first. But our knowledge about optimizations makes more plausible the Factor Semantics that is triggered by the segmental matching. When the Factor Semantics is triggered, the relationships within a given structure, such as adjacency and relative positioning, are asserted (see Appendix A). All the requirements except for body2 being equivalent to body4 are met. But once the cases are considered, we discover that the operator being factored is indeed a loop generator and that we can relax the requirement that body2 be equivalent to body4 to that of similarity of the two bodies. This follows from the support for Final analysis the relationship between body2 and body4. reveals that body2 is embedded in body4. V CONCLUSION Given the small example above, we see that the derivation of the semantic difference involves syntactic and semantic knowledge of the structure domain as well as techniques for managing and applying the knowledge. We present a design that addresses the issue of managing and applying both syntactic and semantic knowledge of the structures being compared so as to provide a semantic interpretation of changes. This allows us to bridge the gap that exists between the information provided by current differencers and the information needed by current differencing tasks. VI APPENDIX A: -- TEMPLATES FROM THE SEMANTIC DIFFERENCER ----I__ Every substructure of the structures being compared has associated with it an Object Profile that is an record with the following role names as fields: Content: Value of the substructure. Type Context: Sequence of nonterminals of productions, of the grammar, used in the derivation of the substructure. Posi t ional Position in parse tree. (i.e. a Context: sequence of directions in reaching the substructure from the root of the parse tree. Abetraction: If a grammar is used to generate the substructure, this refers to the sequence of productions used to generate the substructure itself. 261 A Relation Profile describes the relationships between two substructures, one from each of the structures being compared. The role names of this record are: Base Content (i.e. common unique) Matches: Context (i.e. positional determined from context trees) Positional constraint and Content (i.e. from the largest common residual substructure technique where uniqueness of context matches is relaxed). Inferred Conflict free (i.e. no syntactic Matches: boundary violations) With conflicts (inferences depends on heuristics regarding current substructure abstraction and weights associated with substructure matches). Happ i ngs: l-l substructure matches 2-1 substructure matches n-m substructure matches A second relation profile is one between substructures within a given structure. Some of the considerations here are: adjacency of two substructures, containment, and relative positioning. There are several semantic rules that describe a majority of structure changes. Some are factoring, distributing, commuting, associating, extracting, embedding, folding and unfolding. Below is a partial description of the Factor Semantics used in generating the semantic interpretation above: FACTOR SEllANTICS: FORM: LHS: opl body1 op2 body2 RI-IS: op3 body3 body4 KEY: Segmental matching REDUIREMENTS: requirements opl=opZ requirements opl =op3 requirements bodyl=body3 requirements bodyZ=body4 requirement9 relative positioning of body1 to body2 holds for body3 to body4 requirements adjacency of body1 to body2 holds for body3 to body4 CASES : opl is a loop generator relaxations body2=body4 relaxed to body2 similar to body4 RELAXATIONS: relaxations adjacency requirement relaxations relative positioning relaxations equivalence (=I to similar VII REFERENCES 1. Baiter, R., Goldman, N., and Wile, D., “On the Transformational Implementation Approach to Programming,” in Second International Conference on Software Engineering, pp. 337-344, IEEE, October 1976. 2. Balzer, R., Transformational Implementation: An Example, Information Sciences Institute, Research Report 79-79, September 1979. 3. Chiu, W., Structure Comparison, 1979. Presented at the Second International Transformational Implementation Workshop in Boston, September, 1979. 4. Heckel, P., “A Technique for Isolating Differences Between Files,” Communications of the ACM 21, (41, April 1978, 264-268. 5. Hirschberg, D., The Longest Common Subsequence Problem, Ph.D. thesis, Princeton University, August 1975. 6. Hirschberg, D., “Algorithms for the Longest Common Subsequence Problem,” journal of the ACM 24, (41, October 1977, 664-675. 7. Hunt, J., An Algorithm for Differential file Comparison, Bell Laboratories, Computer Science Technical Report 41, 1976. 8. Hunt, J., “A Fast Algorithm for Computing Longest Common Subsequences,” Communications of the ACM 20, (51, May 1977,350-353. 9. Tai, K., Syntactic Error Correction in Programming Languages, Ph.D. thesis, Cornell University, January 1977. 10. Tai, K., “The Tree-to-Tree Correction Problem,” Journal of the ACM 26, (3), July 1979, 422-433. 262	1980	28
21	Performing Inferences over Recursive Data Bases Shamim A. Naqvi and Lawrence J. Henschen Dept. of Electrical Engineering and Computer Science Northwestern University Evanston, Illinois 60201 Abstract The research reported in this paper presents a solution to an open problem which arises in sys- tems that use recursive production rules to re- present knowledge. The problem can be stated as follows: "Given a recursive definition, how can we derive an equivalent non-recursive program with well-defined termination conditions". Our solution uses connection graphs to first detect occurrences of recursive definitions and then synthesizes a non-recursive program from such a definition. I. Introduction In recent years, attention has focused on adding inferential capability to Codd's relational model of data (Codd 1970). This usually takes the form of defining new relations in terms of existing relations in the data base. The defined relations constituting the Intensional Data Base describe general rules about the data whereas explicit facts stored in the data base as base relations comprise the Extensional Data Base. This paper is concerned with the problem of find- ing a finite inference mechanism for a defined relation. Reiter (1977) suggests that for non-recursive data bases the essentially logical operations in- volved in unifying and resolving intensional lit- erals can be taken care of, i.e. "compiled", be- fore receiving queries, leaving only those opera- tions specifically related to information retrie- val from the extensional data base. We propose to extend this idea to the general case by analyzing what resolutions are possible that can lead to answers for a particular kind of query. In the case of recursive axioms this in- volves finding a pattern of data base retrievals instead of just a single data retrieval as in Reiter (1977). II. Problem Representation We shall view a data base as the ground unit clauses of a first order theory without function signs. The words literal and relation will be used interchangeably and all variables are as- sumed to be universally quantified. We propose to solve the problem of recursive def- initions by using connection graphs like those of Sickel (1976) in which nodes represent intensional axioms and edges connect unifiable literals of opposite signs. A loop is a Potential Recursive Loop (PRL) if the substitutions around the loop are such that the two literals at both ends of the loop are not the same literal and are not in the same instance of the clause partition. Figure 1 shows an example PRL in which E and F are base relations and letters at the end of the alphabet denote variables whereas letters at the start of the alphabet denote constants. In this case, starting from A(a,x,z,p) and resolv- ing around the loop (separating variables as we go) we eventually come back to clause 1 yielding an ultimate resolvent -El(x,y) A(a,y,z,b) B(y,y') lEl(y,y') in which the literal at the end of the cycle, A(a,y,z,b), is a different literal than the one we started the loop with. Two features of this loop traversal are noteworthy. First, the literal E causes data base accesses which provide possibly new values for y. Second, these values of y instantiate x for the next traversal around the loop and also cause data base accesses for F which provides answers to the query. 4 Flgure 1. Example PRL 263 III. Derivation of the Iterative Program Since non-atomic queries can be decomposed into equivalent atomic queries we shall only consider atomic queries in this paper. Before describing our method of deriving an iterative program for a recursive definition we notice that two kinds of edges exist in a PRL. A cycle edge contributes to the PRL whereas an exit edge provides an exit from the PRL. Extensional literals reached by traversing exit edges are called Exit Extensional Literals and those reached by traversing cycle edges are called Cycle Extensional Literals. For example in Figure 1 edges 2, 3 and 4 are cycle edges and edges 5 and 6 are exit edges; 1El(x,y) is a Cycle Extensional Literal whereas 1E3(u,v) and 1E4(q,r) are Exit Extensional Literals. We make the following observations about a PRL. Observation 1 A PRL must have an exit edge, which corresponds to the presence of a basis case for a recursive definition, in order for its clauses to contribute an answer to a query. In figure l the basis case is A(a,q,r,b) 1F(q,r). Notice that a literal having an exit edge has a non-exit edge which contributes to the cycle. Observation 2 In Horn data bases, if a PRL exists for a literal Q, then a literal- must exist which provides the closing edge for the PRL. We represent the defined relations as a connection graph and in a preprocessing step identify all PRLs. A set of answer expressions corresponding to a PRL is derived as follows: We note that the exit edges of Observation 1 above must be connect- ed to cycle literals. Starting from the intension- al axiom from which we expect to get a query, we first delete the literal which would resolve against the query. We then resolve around the cycle until we come to an exit edge. At this point the exit literal represents an expression which can be con- sidered as a call to a procedure. This procedure provides one way of obtaining some of the answers. Paving derived the expression for the first exit we proceed to successive exits in the same manner. These expressions are called answer expressions. In Figure 1 the answer expressions are -tEl(x,y) OR 1E3(y,z) and 1El(x,y) OR -rE4(y,z). A loop residue is obtained by resolving around the loop, starting from the intensional axiom from which we expect to get a query, and traversing only the cycle edges of the PRL. The ultimate resolvent is of the form E:= 1(El(argl,..) & . . . SC Ei(argl,..)) where the Ei (ihO) are base or defined relations. This expression is called the loop residue. In Figure 1 the loop residue is 1El(x,y). In order to derive a program from a PRL we use an algorithm given in Naqvi (1980). In this section we shall illustrate the working of this algorithm by considering two similar definitions of the an- cestor relation. Consider the first definition given below and the corresponding connection graph is shown in Figure 2. 7. 8. -'ANCESTOR(x,y) lFATHER(y,z) ANCESTOR(x,z) lMOTHER(u,v) ANCESTOR(u,v) It is straight-forward to show that with the +NCESTOR(w,a) we can generate the resolvent query 9. tiNCESTOR(w,y') lFATHER(y',y")....lFATHER(y,a) which corresponds to a left recursive definition of the ancestor relation. In this case the basis statement is used in the end to get the expression -+iOTHER(w,y') 1FATHER(y',y")...lFATHER(y,a). The data retrieval pattern is to find successive fathers of 'a' and then find a mother. In terms of the connection graph this corresponds to tra- versing the loop a certain number of times and then taking the exit edge. Examining the PRL we find that z, which is the variable that is expect- ed to be the driver, is replaced in the loop by y. Moreover, z determines y through the extensional evaluation of Father(y,z). This determination occurs within the loop without recourse to the basis statement. Our algorithm does the above kind of analysis and uses the answer expression derived from the loop and the substitutions from the closing edge of the loop to derive a program for the PRL. For this example it derives the following program fragment: 2: =a ENQUE(q,z) /* q is a queue / while (q 1= empty) do z:= DEQUE(q) x:= 1(MOTHER(x,y) & FATHER(y,z)) ENQUE (4, Y 1 od Now, consider the second definition of the ancestor relation given below and the corresponding connec- tion graph shown in Figure 3. 11. lFATHER(x,y) -'ANCESTOR(y,z) ANCESTOR(x,z) 12. -+lOTHER(u,v) ANCESTOR(u,v) Once again, the query -&NCESTOR(w,a) and 11 can be shown to generate the resolvent 13. 1FATHER(x,y) 1FATHER(y,y)...7ANCESTOR(y",a) In this case, our first answers come from resolv- ing (12) and (13) which corresponds in figure 3 to taking the exit edge to the basis case. Sub- sequent answers are derived by finding the succes- sive fathers which corresponds to going around the loop a certain number of times. Examining the PRL we find that the values of y derive the next set of answers, x. The expected driver vari- able z does not participate in this process. Our algorithm uses the resolvent of the basis case and the query to start the loop. The substitutions at the closing edge of the PRL identify the cor- rect variables which drive the loop and serve as place holders for answers. The loop residue de- rives all the subsequent answers. The program is as shown below. .- x.-w t'=lOTHER(w,a) ENQm (q ,w> while (q 1 = empty) do y:= DEQUE(q) 1 FATHER(x,y) ENQUE (4, x> od YlZ i ANCESTOR(u,v) LMOTHER(J Figure 2. First Definition of Ancestor Relation -rFATHER(x,y) -,ANCESTOR(y,z) ANCESTOR(x,z) (y’uz’y &?j z) (w/u,a/v) Figure 3. Second Definition of Ancestor Relation To review then, our algorithm analyzes the PRL oE a recursive definition to determine the loop resi- due, the answer expressions, the resolvent of the query and the basis case and whether the defini- tion is left or right recursive. It then derives a program whose structure corresponds to one of the two program structures outlined above. It now remains to discuss the termination condi- tions of the derived programs. Our termination conditions are designed for data bases without function signs. Briefly, we use a queue to store all unique values of the variables, indicated by the loop analysis, during each iteration. Each new iteration dequeues a previously enqueued value to generate some possibly new answers. Since the domain of discourse is finite in the absence of function signs the number of unique answers in the data base is finite. Thus the program queue will ultimately become empty. It should be noted that our technique for the detection of and generating programs for recursive definitions works in the presence of function signs. However, the termina- tion condition does not guarantee finite computa- tions in this case. v. Summary and Conclusions We have outlined an algorithm which derives itera- tive programs for recursively defined relations. The case where a defined relation is mutually recursive with some other definition (e.g. X & R -> R and R & Y -b X) leads to the derivation of mutually recursive programs. Transitive recursive axioms (e.g. ancestor of an ancestor is an ancestor) lead to the derivation of recursive programs. These situations require a fairly complicated con- trol mechanism for execution time invocation of the derived programs. This is discussed in detail in Naqvi (1980) and the algorithJz for deriving the programs is also given there. We can show the finiteness and completeness of our method (Naqvi 1980). Although we have considered a first order theory without function signs the method is ap- plicable to data bases containing function signs. The termination condition, however, may not be rigorous in this case. This is an obvious area for further research. References Chang, C. L., (1979) "On Evaluation of Queries Containing Derived Relations in a Relational Data Base", Workshop on Formal Bases for Data Bases, Toulouse, France. Codd, E.F., (1970) "A Relational Model of Data for Large Shared Data Banks", CACM 13, 6, 377-387. Naqvi, S. (1980) "Deductive Question-Answering in Recursive Rule-Based Systems", Ph.D. Diss., Northwestern University, (in preparation). Reiter, R., (1977) "An Approach to Deductive Question Answering", BBN Tech. Report no. 3649. Sickel, S. (1976) "A Search Technique for Clause Interconnectivity Graphs", IEEE Trans. on Comput- ers, Vol. C-25, No. 8. 265	1980	29
22	Automatic Goal-Directed Progrem ‘bvnsformetion Stephen Fickas USC/Information Sciences Institute* Marina del Rey, CA 90291 1. INTRODUCTION This paper focuses on a major problem faced by the user of a semi-automatic, transformation-based program-development system: management of low level details. I will argue that it is feasible to take some of this burden off of the user by automating portions of the development sequence. A prototype system is introduced which employs knowledge of the transformation domain in achieving a given program goal state. It is assumed that such a system will run in a real environment containing a large library of both generalized low level and specialized high level transformations. 2. THE TI APPROACH TO PROGRAM DEVELOPMENT The research discussed here is part of a larger context of program development through Transformational Implementation (or TI) [I, 21. Briefly, the TI approach to programming involves refining and optimizing a program specification written in a high level specification language (Currently, the GIST program specification language [8] is being used for this purpose) to a particular base language leg. LISP). Refinement and optimization are carried out by applying transformations to program fragments. This process is semi-automatic in that a programmer must both choose the transformation to apply and the context in which to apply it; the TI system ensures that the left hand side (LHS) of the transformation is applicable and actually applies the transformation. The TI system provides a facility for reverting to some previous point in the development sequence from which point the programmer can explore various alternative lines of reasoning. 3. CONCEPTUAL TRANSFORMATIONS AND JITTERING TRANSFORMATIONS In using the TI system, programmers generally employ only a small number of high level “conceptual” transformations, ones that produce a large refinement or optimization. Examples are changing a control structure from iterative to recursive, merging a number of loops into one, maintaining a set incrementally, or making non-determinism explicit. Typically these transformations have complex effects on the program; they may even have to interact with the user. Although only a relatively small number of conceptual transformations are employed in a typical TI development, the final sequence is generally quite lengthy. Because the applicability conditions of a conceptual transformation may be 9; This research was supported by Defense Advaked Research Projects Agency contract DAHCIS 72 C 0308 Views and conclusions contained in this document are fhooe of the authors and should not be interpreted as repreeenting the official opinion or policy of DARPA, the U.S. Government, or any other person or agency connected wilh them. quite specialized, usually with a number of properties to prove, much of the development sequence is made up of lower level transformations which massage the program into states where the set of conceptual transformations can be applied. I call these preparatory steps “jittering” steps. Examples of quite low level jittering steps include commuting two adjacent statements or unfolding nested blocks. More difficult jittering steps include moving a statement within a block from the kth position to the first position, merging two statements (eg. conditionals,loops) into one, or making two loop generators equivalent. 4. AN AUTOMATIC JITTERING SYSTEM Requiring the programmer to carry out the Jittering process detracls from his performance in several ways: it consumes a large portion of his time and effort; it disrupts his high level planning by forcing him to attend to a myriad of details. There is then strong motivation for automating some or all of the Jittering process. The following sections will discuss the types of mechanisms used to actually implement such a system thenceforth known as the Jitterer). 4.1. BLACK BOX DESCRIPTION The Jitterer is initially invoked whenever the TI system is unable to match the LHS of a transformation Tk selected by the user. The Jitterer’s inputs are 1) the current program state C, 2) a goal state G corresponding to the mismatched LHS of Tk, and 3) a library of transformations L to use in the jittering process. The Jltterer’s return value is either a failure message or a program state S matching G and a sequence of instantiated transformations from L which when started in C will result in S. If the Jltterer is successful, then Tk will be removed from its suspended state and applied as specified in state S. If G is a conjunction of sub-goals then currently a simple STRIPS like approach is employed in solving each in some determined order. This approach can be both inefficient and unbale to find solwtlons for certain jittering problems. Improving this process, possibly using some of the techniques proposed by problem solvers in other domains (see Sacerdoti [9] for a survey), remains a high priority item for future research. 4.2. THE GOAL LANGUAGE Contained in the TI system is a subsystem called the Differencer [53. The Differencer takes as input a current state pattern and a goal pattern, and returns a list of difference-descriptions between the two (an empty list for an exact match). Each element of the list is a description taken at a particular level of detail, and is written in a goal language I call GLl. For example, 68 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. suppose that the Differencer was passed “if P then FOO” as a current pattern, and “if Q then SA” as a goal pattern (The notation SX stands for a variable pattern matching a single statement). The output of the Differencer would be the following three descriptions in GLl: CHANGE-PREDICATE(P 0); CHANGE-ACTION(“if P then FOO” “if Q then SA”), ie. change from one conditional to another without going out of the current context; PRODUCE-ACTION-AT-POSITION(“if Q then $A” J), ie. use any means necessary, including looking at the surrounding context, to produce a conditional matching “if Q then $A” at position J, J being bound by the Differencer. The above three descriptions form a disjunction of goals, each of which gives a slightly wider perspective. Currently the Jitterer attempts to solve the most narrow goal first. If it is successful, then control returns to TI. If not, it attempts to solve the next higher goal and so on until the list is exhausted. De fault-bindings: none t various tactics 1 We now must define the individual jittering plans or Tactics which will achieve the CHANGE-PREDICATE goal. Each Toceic construct is composed of a set of constraints which further limit its applicability and an Actron for achieving the matching goal. Our three plans become formally STRATEGY change-predicate-from-a-to-b - _ - 1 . . Tactic (1) enlbed Applicabhty-condition: none Action: POST (EMBED (Predl Pred21) Taceic(2) extract Applicability-condition: NOT (VARIABLE (Predl I 1 Action: POST (EXTRACT tPred2 Predl) 1 Other goals of GLl not included in the above example include EMBED, EXTRACT, DELETE, ADD, COMMUTE, DISTRIBUTE, PROVE-PROPERTY, FOLD and UNFOLD. Because GLl is the language used to both describe pattern differences by the Dlfferencer and describe desired goal states by the Jitterer, it acts as a common interface between the two. 5. JiTTERIf’& PLANS A small number of jittering plans are capable of solving many of the goals of GLl. Suppose we take for example the goal of the previous section CHANGE-PREDICATE(P Q). This is a specific instance of the more general goal CHANGE-PREDICATE(pattern1 pattern2). There are three basic plans for solving this more general goal: 1) embed pattern1 in patternil, 2) extract pattern1 from pattern2, or 3) first embed pattern1 in something containing pattern2, and then extract pattern2. Since in our case each pattern is a simple variable, we can rule out the second plan. Similar plans exist for the other two goals of the example. In general, jittering plans can be defined for all but the most detailed goals of GLl. 5.1. STRATEGY AND TACTICS The plans available for solving a particular goal have been organized into a construct I call a STRATEGY. Each of the plans is known as a Tactic. Each STRATEGY construct takes the following form (see [4], [7] for similar planning constructs in other domains.): STRATEGY name Relevant-goal: goal the strategy matches Applicability-condition: used to narrow strategy’s scope Default-bindings: bind any unbound goal parameter 6 Tactic(l) . . . Tactic(n) . . . To illustrate, let us package the three plans described informally for solving the CHANGE-PREDICATE goal into a STRATEGY Tactic (31 embed-and-then-extract Applicabihy-condition: none Action: SEQ (POST (EMBED (Predl (#ANY-PRED fr Pred211) POST (EXTRACT tPred2 (#ANY-PRED Predl Pred21 Here, POST(G) says mark the goal G as a sub-goal to be achieved. *ANY-PRED will match to either AND or OR. SEQUENCE(A1 A2 . . . An) says execute the list of actions scqucntlally; if Ai is of the form POST(G), then do not move on to Ai+l until G has been achieved. There exist similar functions for executing a sequence of actions in parallel and in disjunctive form. In general, an Action can specify an arbitrarily ordered list of actions to take to achieve the STRATEGY’S goal. 5.2. BACKWARD CHAINING The transformations available for jittering, along with the Tactics introduced through the STRATEGY construct, define the methods available for solving a particular goal. Transformations are applied in backward chaining fashion: once a transformation’s RHS matches a goal G, variables are bound and the LHS is instantiated. The Differencer is then called in to see if a match exists between the current state and the instantiated LHS. If so, then G is marked as achieved by the application of the transformation. if there is a mismatch, then the disjunction of sub-goals produced by the Differencer will be marked as awaiting achievement. 6. THE JITTERlNG SCHEDULER Whenever a new goal is posted (marked to be achieved), the Jitterer attaches to it a list of methods (transformations and Tactics) which might solve it. It is at this point that the Scheduler is called in. The Scheduler first must decide among all active goals (ones with at least one untried method) which to construcl. Suppose that we wanted to rule out working on goals work on next. Once a goal IS chosen, it must decide which of the that attempt to change True to False or False to True, and that untried methods to employ. A set of domain dependent metrics we expected both of CHANGE-PREDICATE’s parameters to be which help the Scheduler make an intelligent decision in both bound. We would get the following STRATEGY header: cases has been identified. STRATEGY change-predicate-from-a-to-b Relevant-god: CHANGE-PREDICATE (Predl Pred2) .&n!:cability-conditions: NOT (Predl=True h Pred2=Fa I se) A NOT (Predl-Fa I se A Pred2=True) 6.1. Choosing among competing goals Lcngfh of path: the types of problems presented to the .litterer by Ti do not generally involve a large (over 10) number of transformation steps. Clcnce, as a path (sequence of 69 transformations) grows continuing it decreases. over a fixed threshold, the desirability of Conflict with high level motivation: if the Scheduler is able to determine the user’s current high level goal, it may be able to rule out certain goal states as un-productive. For example, if it is known that the user is trying to optimize his control structure by merging a number of loops into one, it may be unwise to try to achieve a sub-goal which produces still another loop. Such a sub-goal would be given low priority. Ease of achieving: A rough estimate is made of the cost to continue the goal by taking the minimum cost of the untried methods attached to it. This is a rough estimate because there is no easy way to compute exactly the final cost associated with any method, or in fact that any method will lead to a solution. 6.2. Choosing among competing methods Ease of application: a rough static estimate of how difficult the method may be to apply. In the case of a transformation, how complex is the LHS (eg, how many properties must be proven)? In the case of a Tactic, how many sub-goals must be achieved? User assistance: some methods call for the user to supply needed information. The preference is to avoid bothering the user as much as possible. Side effects: What undesirable actions will a method take besides the desired one (a qualitative judgment)? For instance, a method which unfolds a large procedure body to produce a certain type of goal pattern is seen as having a large side effect, ie. it tends to “flatten” the functional structure of the program. On the other hand, a method which simply changes a current statement into something matching a goal pattern has very little side-effect; it does nothing to disturb the surrounding context. Prefer small side effects over large ones. Tactic ordering rules: a STRATEGY writer may provide rules for ordering Tactics. These rules take the form “if Condition then Ordering”, where Condition can refer to any piece of knowledge known to the STRATEGY at match time and Ordering takes the form “Try Tactic(J) before Tactic(K)” or “Try TactdN) last”. The Scheduler makes use of this information when choosing among competing tactics for a particular goal. 7. CONCLUSION The purpose of this paper has been to present a prototype Jitterer with enough domain knowledge to deal competently with the types of jittering problems typically encountered in a TI environment. The prototype Jitterer described is currently being implemented in the Hearsay-ill knowledge representation language [3] and represents a preliminary system. Future systems will deal much more with performance issues (see for example the next section). 8. FURTHER RESEARCH There are generally many ways of achieving a given jittering goal. The metrics of section 5 give some help in ordering them. Even so, in some cases the Jitterer will produce a solution not acceptable to the user. A simple minded approach (and the one currently employed) is to keep presenting solutions until the user ik satisfied. A better approach is to allow the user to specify what he didn’t like about a particular solution and allow lhts information to guide the search for subsequent solutions. in fact, the user may not want to wait until an incorrect solution has been presented, but give jittering guidance when the Jitterer is initially invoked (Feather [6] proposes such a guidance mechanism for a fold/unfold type transformation system). Still another approach may be to “delay” jittering until more high level contextual information can be obtained. Both user guidance and delayed reasoning are being actively studied for inclusion in future jittering systems. Acknowledgments I would like to thank Bob Balzer, Martin Feather, Phil London and Dave Wile for their contributions to this work. Lee Erman and Neil Goldman have provided willing and able assistance to the Hearsay Ill implementation effort. 9, REFERENCES 1. Balzer, R., Goldman, N., and Wile, D. On the Transformational Implementation Approach to programming. Second International Conference on Software Engineering, October, 1976. 2. Balzer, R. TI: An example. Research Report RR-79-79, Information Sciences Institute, 1979. 3. Aalzcr, R., Erman, L., London, P., Williams, C. Hearsay-Ill: A Domain-Independent Framework for Expert Systems. First National Conference on Artificial Intelligence, 1980. 4. Bulrss-Rozas, J. GOAL: A Coal Oriented Command language for Int ractue Proff Constructlorr. Ph.D. Th., Computer Science Dept., Standford University, 1979. 5. Chiu, W. Structure Comparison and Semantic Interpretation of Differences. First National Conference on Artificial Intelligence, , 1980. 6. Feather, M. A System For Developing Programs by Transformatron. Ph.D. Th., Dept. of Artificial Intelligence, University of Edinburgh, 1979. 7. Friedland, P. Knowledge-based Hierarchical Planning in Molecular Genetics. Ph.D. Th., Computer Science Dept., Stanford University, 1979. 8. Goldman, N., and Wile, D. A Data Base specification. International Conference on the Kntity-Relational Approach to Systems Analysis and Design, UCLA, 1979. 9. Sacerdoti, E. Problem Solving Tactics. Technical Note 189, SRI, July 1979. 70	1980	3
23	PIAGET AND ARTIFICIAL INTELLIGENCE Jarrett K. Rosenberg Department of Psychology University of California, Bcrkelcy 94720 ASSTRACT Piaget’s Genetic Epistemology and Artl$cial Intelligence can be of great use to each other, since the strengths of each approach complement the weaknesses of the other. Two ways of bringing the two approaches together are suggested: elaborating the parallels between them (such as the concept of schemata), and building AI models based directly on Piaget’s theory. How this mighl benejit AI is illustrated by examining how Piagerian research on problem solving can suggest new ways of building programs that learn. I. INTRODUCTION. My thesis in this paper is that two such superficially disparate areas as Piaget’s Genetic Epistemology and Artificial Intelligence can inform each other immensely. This fact has already been noticed, as long ago as Seymour Papert’s work at Geneva [l] and as recently as Margaret Boden’s writings on Piaget [2, 31. Nevertheless, I think it still needs to be pointed out how the two fields both parallel and complement each other, and how one might go about trying to connect them. Piaget’s work is immense, stretching over half a century and dozens of volumes, and so only a glimpse of it can be given here, but hopefully it will be enough to stimulate some AI people (and maybe psychologists too). II. WHY SHOULD PIAGET AND RELEVANT? To address the problem of artificial intelligence is to presuppose what natural intelligence is, but AI work in general doesn’t really treat the latter in any comprehensive way; it’s assumed that playing chess is intelligent, or planning, or ordering a hamburger. But can we really hope to understand intelligent behavior, natural or artificial, without deciding what criteria we use for intelligence? If the answer is No. then it seems natural to look to psychology for a theory of intelligence. The only theory I find satisfying, and I suspect most AI people will as well, is Piaget’s. Piaget’s theory is relevant to AI because it is a fully developed, motivated theory of intelligence which is general enough to subsume not only human and animal intelligence, but purely artificial intelligence as well. By motivated I mean that the form of Piaget’s theory of knowledge is based on both the logical requirements of epistemology, and on the biological requirements of behavioral evolution. Hence there are good reasons for accepting his approach to intelligence, just as there are few, if any, compelling reasons to believe the (often implicit) definitions given in AI (or psychology, for that matter). Part of my argument, then, will be that AI rcscarchcrs might be able to use this foundation as a basis For their own work. Piaget, on the other hand, can learn 3 lot from AI, since it has developed ways of looking at the detailed structures and processes of complex behavior, thus suggesting ways of elaborating Piaget’s theory into one with a more “psychological” flavor. Piaget’s theory has always been more epistemologically than psychologically oriented (in the sense of information-processing psychology), and it has only been recently that Genevan researchers have addressed the question of the procedures involved in the acquisition and use cf knowledge- structures. Table 1 gives a general (and biased!) comparison of the strengths and weaknesses of Piaget’s theory and AI work taken as a whole. The important thing to note is that the strengths and weaknesses of the two approaches complement each other nicely. Let me discuss them briefly. The advantages of Piaget’s theory are: b It is both epistemologically and biologically motivated, that is. it starts from the most central issues of a theory of knowledge (e.g., how space, time, and causality are constructed), and places this analysis within the biological context of human evolution. This contrasts with most AI work, which considers intelligence outside of both its philosophical and biological context. b It is highly theoretical and comprehensive. As a biologist with philosophical interests, Piaget resolved to study the “natural history of knowledge”, and so developed a general theoretical framework which he then applied to a number of areas, particularly those of physical and logico-mathematical knowledge. This contrasts with the general lack of interest in AI for large-scale theories or applications. b It has a large amount of empirical validation: virtually everything he has proposed has been extensively replicated and consistently extended. b It proposes a “fundamental unit” for cognition, the scheme, from which all knowledge-structures are constructed. This supplies a uniform analytic framework for all research in cognition. b It has a formalism for the structures of knowledge. This formalism is an algebraic, rather than a computational one, and is only in its beginnings, but is nevertheless valuable. b It is a developnmtal theory of knowled&, promising a way to -- understand how knowledge-structures can radically change and yet prcscrve their integrity. It can thus provide an additional, developmental, set of constraints on AI theories: both structural ones (certain logical properties of cognition, with a logical and devclopmcntal ordering on them), and l%nctional/proccdural ones on the way those structures are constructed and used (e.g., forms of coordination). This should be one of its greatest attractions for AI. The disadvantages of Piaget’s theory are: b It is not very detailed about how knowledge is represented, or how it is processed; to someone looking for a theory at that level, Piagct’s seems mostly descriptive. b It says little about the proc&zrres used in cognition; only recently has this issue been addressed at Geneva. b Its emphasis is on physical and logico-mathematical knowledge, rather than other kinds such as social or linguistic. Here the field of AI has to offer prccirely what Piagctian research lacks, and a glance at the table shows th&t I consider the converse to be true. III. HOW CAN WE PUT THEM TOGETHER? Given that we want to bring these two areas into contact, two possible ways of doing it come immediately to mind. First, there are a number of parallels between Piagetian concepts and ones introduced in AI, the most obvious being that between Piaget’s notion of a scheme and the schema/frame/script ideas in vogue the past few years in AI. The most important such parallels are: b Schemata. For Piaget, the fundamental unit of knowledge is the scheme, a generalizable action-pattern. b Schematically-based processing As a consequence of his scheme- based theory, Piaget views cognitive processes as primarily top-down. b Knowlcdgc as action. In Piagct’s epistemology, no knowledge is --- gained without acting and its conscqucnt transfclmaticn of experience. These parallels can be used not only to suggest new ways of developing the AI notions, but as ways of formulating the Piagetian concepts in more detail. Second, we can attempt to create AI models of Piaget’s theory (not tasks, as has sometimes been done). This will introduce another paradigm into AI research, no worse than the others and hopefully better. Simultaneously, such work can serve to rigorously develop the processing aspects of Piaget’s theory, using all the insights of AI research. Let me give a brief example. One of the ways that Piagctian theory is relevant to AI is that it can give both theoretical and empirical suggestions for how to design programs that learn, i.e., develop new cognitive structures in order to solve problems. Current AI learning programs work by assuming some initial body of knowledge and learning techniques, and then showing how additional knowledge and techniques could be learned. However, this creates a regress: how the starting knowledge and techniques are acquired is left for someone else to explain. In addition, the initial knowledge used is often much more advanced than what is to be acquired; for example, Sussman’s HACKER uses fairly sophisticated debugging techniques to learn something children acquire by age two. Piaget’s theory, on the other hand, tries to account for all of cognitive development starting with only the simplest of structures. Piagetian research reveals a number of important characteristics of problem solving that are not tilly achieved until adolescence, being gradually constructed starting around age five (ignoring the sensori- motor precursors). They are: b The ability to construct and overgcncralize theories, and as a corollary, to interpret phenomena in terms of confirmation or disconfirmation of those theories. b The ability to construct refifations as tests, as well as confirmations. b The presence of articulated part-whole relations, e.g., as reflected in an appropriate decomposition of the problem. Table 1. A Comparison of Piagetian Theory and Artificial Intelligence. Piaget Al Strengths 1. Motivated (biologically & epistemologically) 2. Comprehensive theory 3. Lots of empirical validation 4. Fundamental unit of knowledge S..Formal model of knowledge structures 6. Developmental theory Weaknesses 1. Not detailed 2. Not procedural Strengths Weaknesses -- 1. Precise/formal 1. Unmotivated 2. Process-oriented 2. Not theoretical 3. Detailed 3. Little empirical 4. Body of validation techniques 4. Little agreement 5. Narrow 6. No model of development 267 ) The ability to reflect on procedures in order to switch or modify them. ) The ability to coordinate goals and subgoals. Piagetian studies of how ,these abilities are constructed can provide useful ideas for AI. Consider, for example, how children learn to balance blocks (see [4]). For very young children, actions of objects are more or less random unless assimilated to their own actions and intentions. Their attempts to balance blocks ignore completely the relevant factors in favor of pushing down on the block to maintain a balance. Somewhat older children, after much trial and error, recognize a regularity in the outcomes of their efforts, namely, that most things balance at their geometrical centers. Once this idea (a primitive theory) is grasped, it is rapidly gcncralized, to the point where exceptions to it are dismissed as chance. The eventual overgeneralization of this notion of regularity, and consequent rcpeatcd failures to confirm it, give rise to the idea that the exceptions to the theory may have regularities themselves. At this point, children will retain their original theory where it works, but construct an additional one for the counterexamples (balance based on weight distribution). As their powers of conceptual coordination grow, they can combine the two theories into a single, more general, one. Then finally the idea of a retitation test usually appears (the preference for confirmations stemming from the same source as that of overgeneralizing in the first place). What does this suggest for AI work on learning and problem solving? If Piaget is correct, then in order to be generally successful, AI programs should have the abilitieamentioned, and they should be able to develop ,them in the manner given above. In particular, we can imagine a class of AI programs that approach problem solving in the following way. Faced with a new problem they would: ) engage extcnsivcly in trial-and-error behavior, ) strive to find regularities in the outcomes (including regularities in the procedures used), ) use the regularities to construct and then overgeneralize a theory, ) construct confirmation tests of the theory and note the exceptions, ) construct a theory for the exceptions, if they are numerous enough, ) try to coordinate the two theories into a larger one, ) test this resultant theory by confirmation and refutation tests. Besides theories of problem situations, the programs would also construct theories of the problem-solving procedures themselves, i.e., they would study the relationships of procedures among themselves and with respect to their outcomes, rather than just theories of the relationships among states of the world. Some of the above is partially captured in various AI programs (Sussman’s HACKER, Winston’s arch builder, erc.), but none of them are configured in exactly this way, and all of them assume fairly sophisticated techniques and knowledge as a basis for learning. The proposal above is obviously not the last word in learning and problem solving, but at least provides a useful place to start. In addition, the vast amount of empirical data that Piagetians have collestcd can be invaluable as an empirical testing ground for AI models. Up to now, AI rcscarchers have only attempted to model performance on Pidgetian tasks, vithout rcalbing that it is both the empirical and the theoretical constraint5 that would make Piagetian work of use to Al: were a model to simulate performance on a Piagetian task. and do so in a way fully consistent with the theory, it would have met a very strong test of its validity. Moreover, the tasks themselves have a high degree of ecological validity and cpistcmological significance (for example, conservation of matter and number). Thus, by making use of the existing parallels between the two approaches, and by recasting Piagctian concepts in AI terms, the two might produce a hybrid that would have all the advantages of a Piagetian foundation as well as all the benefits of Al’s hard-won insights into knowledge-representation and information-processing. For this to happen, however, it will be necessary for both Piagctians and Al researchers to learn more about each other’s work; as mentioned above, previous AI work has ncglccted Piagct’s theory in favor of his tasks, while Piagetians have only the vaguest notion of what a computational model is all about. And of course, Piaget’s theory is not some sort of panacea-psychology’s gift to AI. On the contrary, it needs to be much more developed, along the lines that AI work is pursuing. It is precisely because each could profit from the other that I’ve presented these points here. ACKNOWLEDGEMENTS I would like to thank Jonas Langer, John Sccly Brown, and Tom Moran for their comments. No endorsements implied. REFERENCES [l] Papert, S. “Etude comparee de l’intelligence chez l’enfant et chez le robot.” in La ZYiafion des Structures. Etudes d’Epistemologie G&nCtique. vol. 15. Paris: P.U.F. 1963. [2] Boden, M. “Artificial Intelligence and Piagetian Theory.” Synlhese. 38: 389-414. 1978. 131 -- Piagel. New York: Viking. 1979. [4] Karmiloff-Smith, A. and B. Inhcldcr. “If you want to get ahead, get a theory.” Cognirion. 3: 195-212. 1975. 268	1980	30
24	RI: an Expert in the Computer Systems Domain’ John McDermott Department of Computer Science Carnegie-Mellon University Pittsburgh, Pennsylvania 15213 INTRODUCTION. Rl* is a rule-based system that has much in common with other domain-specific systems that have been developed over the past several years [l, 81. It differs from these systeins primarily in its use of Match rather than Generate-and-Test as its central problem solving method [2]; rather than exploring several hypotheses until an acceptable one is found, it exploits its knowledge of its task domain to generate a single acceptable solution. Rl’s domain of expertise is configuring Digital Equipment Corporation’s VAX-l l/780 systems. Its input is a customer’s order and its output is a set of diagrams displaying the spatial relationships among the components on the order; these diagrams are used by the technician who physically assembles the system. Since an order frequently lacks one or more components required for system functionality, a major part of Rl’s task is to notice what components are missing and add them to the order. Rl is currently being used on a regular basis by DEC’s manufacturing organization.3 THE DOMAIN. The VAX-11/780 is the first implementation of DEC’s VAX-1 1 architecture. The VAX-1 l/780 uses a high speed synchronous bus, the sbi, as its primary interconnect; the central processor, one or two memory control units, up to four massbus interfaces, and up to four unibus interfaces can be connected to the sbi. The massbuses and particularly the unibuses can support a wide variety of peripheral devices. A typical system contains about 90 components; these include cabinets, periperal devices, drivers for the devices, and cables. There are a large number of rules that constrain the ways in which these components may be associated. RI’S DEFINING CHARACTEKISTICS. Rl is implemented in OPS4, a production system language developed at Carnegie-Mellon University [3, 71. An OPS4 production system ‘This paper describes Rl as it exists in June of 1980; it is a highly condensed version of [5]. 2 Four years ago I couldn’t even say *‘knowledge engineer”, now I . . . 3The development of Ri was supported by Digital Equipment Corporation. The research that led to the development of OPS4, the language in which Rl is written, was sponsored by the Defense Advanced Research Projects Agency. (DOD), ARPA Order No. 3597, and monitored by the Air Force Avionics Laboratory under Contract F33615-78-C-1151. The views and conclusions contained in this document are those of the author and should not be interpreted as representing the official policies, either expressed or implied, of Digital Equipment Corporation, the Defense Advanced Research Projects Agency, or the U.S. Government. VAX, PDP-11, UNIBUS, and MASSBUS are trademarks of Digital Equipment Corporation. consists of a set of productions held in production memory and a set of data elements (eg, state descriptions) held in working memory. A production is a rule composed of conditions and actions: Pi (Cl C2 . . . Cn --> A, A2 . . . Am) Conditions are forms that are instantiated by memory elements. Actions add elements to working memory or modify existing elements. The recognize-act cycle repeatedly finds all production instantiations and executes one of them.4 Rl exploits this recognition match. Its rules have conditions that recognize situations in which a particular type of extension to a particular type of partial configuration is permissable or required; the actions then effect that extension. OPS4’s two memories have been augmented,. for this application, with a third. This memory, the data base, contains descriptions of each of the 420 components currently supported for the VAX. Each data base entry consists of the name of a component and a set of eight or so attribute/value pairs that indicate the properties of the component that are relevant for the configuration task. As Rl begins to configure an order, it retrieves the relevant component descriptions. As the configuration is generated, working memory grows to contain descriptions of partial configurations, results of various computations, and context symbols that identify the current subtask.’ Production memory contains all of Rl’s permanent knowledge about how to configure VAX systems. Rl currently has 772 rules that enable it to perform the task.5 These rules can be viewed as state transition operators. The conditional part of each rule describes features that a state must possess in order for the rule to be applied. The action part of the rule indicates what features of that state have to be modified or what features have to be added in order for a new state that is on a solution path to be generated. Each rule is a more or less autonomous piece of knowledge that watches for a state that it recognizes to be generated. Whenever 40PS4’s cycle time, though it is essentially independent of the size of both production memory and working memory [4], depends on particular features of the production system (eg, the number and complexity of the conditions and actions in each production); the average cycle time for OPS4 interpreting Rl is about 150 milliseconds. CPS4 is implemented in MACLISP; Rl is run on a PDP- 10 (model KL) and loads in 4 12 pages Of core. 5 Only 480 of these ru!es are “configuration rules”; the remainder contain more general (non domain-specific) knowledge that Rl needs in order to use the configuration rules. 269 that happens, it can effect a state transition. If all goes well, this new state will, in turn, be recognized by one or more rules; one of these rules will effect another state transition, and so on until the system is configured. English translations of two sample rules are shown in Figure 1. ASSlGN-UB-MODULES-EXCEPT-THOSE-CONNECTlNG-TO-PANELS-4 IF: THE CURRENT CONTEXT IS ASSIGNING DEVICES TO UNIBUS MODULES AND THERE IS AN UNASSIGNED DUAL PORT DISK DRIVE AND THE TY PE OF CONTROLLER IT REQUIRES IS KNOWN AND THERE ARE TWO SUCH CONTROLLERS NEITHER OF WHICH HAS ANY DEVICES ASSIGNED TO IT AND THE NUMBER OF DEVICES THAT THESE CONTROLLERS CAN SUPPORT IS KNOWN THEN: ASSIGN THE DISK DRIVE TO EACH OF THE CONTROLLERS AND NOTE THAT THE TWO CONTROLLERS HAVE BEEN ASSOCIATED AND THAT EACH SUPPORTS ONE DEVICE degree of conditionality in the configuration task. The fan-in of a rule is the number of distinct rules that could fire immediately before that rule; the fan-out is the number of distinct rules that could fire immediately after the rule. The average fan-in and fan-out of Rl’s rules is 3. The graph of possible rule firing sequences, then, has 666 nodes, one for each of the rules (excluding the 106 output generation rules); each of these nodes has, on the average, three edges coming into it and three going out. It should be clear that unless the selection of which edge to follow can be highly constrained, the cost (in nodes visited) of finding an adequate configuration (an appropriate path through the rules) will be enormous. It is in this context that the power of the Match method used by Rl becomes apparent. When Rl can configure a system without backtracking, it finds a single path that consists, on the average, of about 800 nodes. When Rl must backtrack, it visits an additional N nodes, where N is the product of the number of unsuccessful unibus module sequences it tries PUT-UB-MODULE-6 IF: THE CURRENT CONTEXT IS PUTTING UNIBUS MODULES IN BACKPLANES IN SOME 80X AND IT HAS BEEN DETERMINED WHICH MODULE TO TRY TO PUT IN A BACKPLANE AND THAT MODULE IS A MULTIPLEXER TERMINAL INTERFACE AND IT HAS NOT BEEN ASSOCIATED WITH ANY PANEL SPACE AND THE TY PE AND NUMBER OF BACKPLANE SLOTS IT REQUIRES IS KNOWN AND THERE ARE AT LEASTTHAT MANY SLOTS AVAILABLE IN A BACKPLANE OFTHE APPROPRIATETYPE AND THE CURRENT UNIBUS LOAD’ON THAT BACKPLANE IS KNOWN AND THE POSITION OFTHE BACKPLANE IN THE BOX IS KNOWN THEN: ENTER THE CONTEXT OF VERIFYING PANEL SPACE FOR A MULTIPLEXER Figure 1: Two Sample Rules (which is rarely more than 2) and the number of nodes that must be expanded to generate a candidate unibus module configuration (which is rarely more than 300). Rl ‘S EVOLUTION. In a period of less than a year, Rl went from an idea, to a demonstration system that had most of the basic knowledge required in the domain but lacked the ability to deal with complex orders, to a system that possesses true expertise. Its development parallels, in many respects, the development of the several domain-specific systems engineered by Stanford University’s Heuristic Programming Project [2]. Rl’s implementation history divides quite naturally into two stages. During the first stage, which began in December of 1978 and lasted for about four months, I spent five or six days being tutored in the basics of VAX system configuration, read and reread the two manuals that describe many of the VAX configuration constraints, It is usual to distinguish the matching of forms and data from search: for example, in discussing the amount of search occurring in a resolution theorem prover, the unification of clauses is considered to be part of the elementary search step. But Match is also a method for doing search in a state space [6]; it is analogous to methods such as Hill Climbing or Means-ends Analysis, though much more powerful. The characteristic that distinguishes Match from other Heuristic Search methods is that in the case of Match the conditions (tests) associated with each state are sufficient to guarantee that if a state transition is permissible, then the new state will be on a solution path (if there is a solution path). Thus with Match, false paths are never generated, and so backtracking is never required. Match is well suited for the configuration task because, with a single exception, the knowledge that is available at each step is sufficient to distinguish between acceptable and unacceptable paths. The subtask that cannot always be done with Match alone is placing modules on the unibus in an acceptable sequence; to perform this subtask, Rl must occassionally generate several candidate sequences. and implemented an initial version of Rl (consisting of fewer than 200 domain rules) that could configure the simplest of orders correctly, but made numerous mistakes when it tried to tackle more complex orders.” The second stage, which lasted for another four months, was spent in asking people who were expert in the VAX configuration task to examine Rl’s output, point out Rl’s mistakes, and indicate what knowledge Rl was lacking. RI was sufficiently ignorant that finding mistakes was no problem. Given a criticism of some aspect of the configuration by an expert, all that was necessary in order to refine Rl’s knowledge was to find the offending rule, ask the expert to point out the problem with the condition elements in the rule, and then either modify the rule or split it into two rules that would discriminate between two previously undifferentiated states. During this stage, Rl’s domain knowledge almost tripled. VALIDATION. During October and November of 1979, Rl was involved in a formal validation procedure. Over the two month period, Rl was given 50 orders to configure. A team of six experts The fan-in and fan-out of Rl’s rules provide a measure of the ‘During this first stage, Rl’s name was XCON. 270 examined RI’s output, spending from one to two hours on each REFERENCES order. In the course of examining the configurations, 12 pieces of errorful knowledge were uncovered. The rules responsible for the errors were modified and the orders were resubmitted to RI and were all configured correctly. Each of these 50 orders contained, ‘. Amarel, S. et al. Reports of panel on applications of artificial intelligence. Proceedings of the Fifth International Joint Conference on Artificial Intelligence, MIT, 1977, pp. 994-1006. on the average, 90 components; RI fired an average of 1056 rules 2. and used an average of 2.5 minutes of cpu time in configuring Feigenbaum, E. A. The art of artificial intelligence. Proceedings of the Fifth International Joint Conference on each order. Since January of 1980, RI has configured over 500 orders. It is now integrated into DEC’s manufacturing organization. It has also begun to be used by DEC’s sales organization to configure orders on the day they are booked. Artificial Intelligence, MIT, 1977, pp. 1014-1029. to enlarge its domain so that it can become a more helpful system. CONCLUDING REMARKS. RI has proven itself to be a highly competent configurer of VAX-l l/780 systems. The configurations that it produces are consistently adequate, and the information Work has already begun on augmenting RI’s knowledge to enable that it makes available to the technicians who physically assemble systems is far more detailed than that produced by the humans who do the task. There are, however, some obvious ways in which 3. Forgy, C. L. and J. McDermott. OPS, A domain-independent production system language. Proceedings of the Fifth International Joint Conference on Artificial Intelligence, MIT, 1977, pp. 933-939. 4. Forgy, C. L. RETE: A fast algorithm for the many pattern/many object pattern match problem. Carnegie-Mellon University, Department of Computer Science, I 980. 5. systems. Carnegie-Mellon University, Department of Computer McDermott, J. RI : a rule-based configurer of computer Science, 1980. it to configure other computer systems manufactured by DEC. In those capabilities. Ultimately we hope to develop a salesperson’s addition, we plan to augment its knowledge so that it will be able to assistant, an Rl that can held a customer identify the system that help with the scheduling of system delivery dates. We al,so plan to augment RI’s knowledge so that it will be able to provide best suits his needs. interactive assistance to a customer or salesperson that will allow him, if he wishes, to specify some of the capabilities of the system he wants and let RI select the set of components that will provide 6. Newell, A. Heuristic programming: ill-structured problems. ln Progress in Operations Research, Aronofsky, J. S.,’ Ed.,John Wiley and Sons, 1969, pp. 361-414. 7. Newell, A. Knowledge representation aspects of production systems. Proceedings of the Fifth International Joint Confeience on Artificial Intelligence, MIT, 1977, pp. 987-988. 8. Waterman, D. A. and F. Hayes-Roth. Pattern-Directed Inference Systems. Academic Press, 1978. ACKNOWLEDGEMENTS. Many people have provided help in various forms. Jon Bentley, Scott Fahlman, Charles Forgy, Betsy Herk, Jill Larkin, Allen Newell, Paul Rosenbloom, and Mike Rychener gave me much encouragement and many valuable ideas. Dave Barstow, Bruce Buchanan, Bob Englemore, Penny Nii, Ted Shortliffe, and Mark Stefik contributed their knowledge engineering expertise. Finally, Jim Baratz, Alan Belancik, Dick Caruso, Sam Fuller, Linda Marshall, Kent McNaughton, Vaidis Mongirdas, Dennis O’Connor, and Mike Powell, all of whom are at DEG, assisted in bringing RI up to industry standards. 271	1980	31
25	RULE-BASED MODELS OF LEGAL EXPERTISE D. A. Waterman and Mark Peterson The Rand Corporation 1700 Main Street Santa Monica, California ABSTRACT This paper describes a rule-based legal de- cisonmaking system (LDS) that embodies the skills and knowledge of an expert in product liability law. The system is being used to study the effect of changes in legal doctrine on settlement stra- tegies and practices. LDS is implemented in ROSIE, a rule-oriented language designed to facilitate the development of large expert systems. The ROSIE language is briefly described and our approach to modeling legal expertise using a prototype version of LDS is presented. L. INTRODUCTION We are currently engaged in designing and building rule-based models of legal expertise. A rule-based model of expertise is a computer program organized as a collection of antecedent-consequent rules [l] that embodies the skills and knowledge of an expert in some domain. The primary goal of our work is to develop rule-based models of the de- cisionmaking processes of attorneys and claims ad- justers involved in product liability litigation. We will use these models to study the effect of changes in legal doctrine on settlement strategies and practices. Some progress has already been made in developing computer systems to perform legal analysis. The LEGOL Project [2] has been working for a number of years on the construction of a language for expressing legislation. In addition, systems have been developed for analyzing cases on the basis of legal doctrine 13,41, investigating the tax consequences of corporate transactions [S], automating the assembly of form legal documents if313 and performing knowledge-based legal informa- tion retrieval [7]. Our legal decisionmaking system (LDS) is being implemented in ROSIE, a rule-oriented language designed to facilitate the development of large ex- pert systems. In section II the ROSIE language is briefly described. Section III discusses our ap- proach to modeling legal expertise and describes the operation of our prototype version of IDS. The conclusions are presented in section IV. II. METHODOLOGY - A rule-oriented system for implementing exper- tise (ROSIE) is currently under development-to pro- vide a tool for building expert systems in complex domains 181. ROSIE is a direct descendant of RITA [9] and more distantly MYCIN [lo] in that the models created are rule-based with data-directed control [ll], and are expressed in an English-like syntax. In addition, the models use special language primitives and pattern matching routines that facilitate interaction with external computer systems. not found The ROSIE design also includes features in these successor systems, such as a hierarchical data structure capable of supporting abstraction and inheritance in a general way, par- titioned rulesets that can be called as subroutines or functions, a clearer differentiation between rule antecedent matching and iterative control by permitting actions that involve looping through the data base, and a user support environment with ex- tended facilities for editing and explanation. In the latest version of ROSIE, directed modules data structure are divided knowledge or rules and the declarative knowledge or facts. Both rules and facts are represented as the pattern- used to examine and modify the antecedent-consequent into the imperative where the consequent is either an action to be executed (for rules) or a statement to be deduced (for facts). Rules operate via forward chaining and are of two basic types: existence-driven (IF-THEN) as in RITA, and event- driven (WHEN-THEN) as in ARS [12]. Facts, on the other hand, operate via backward chaining and are represented only as IF-THEN pairs. The facts in ROSIE are similar to RITA goals, but are more gen- eral since they are implicitly referenced by the rules and automatically executed whenever the rules need information the facts can supply. In effect, the information that can be inferred from the facts is a %irtual data base" or extension to the stan- dard ROSIE data base. The current ROSIE syntax is more English-like than that of RITA or earlier versions of ROSIE. It is intended to facilitate model creation, modifica- tion and explanation. This syntax is illustrated in Figure 1, which shows our definition of strict liability in the product liability domain. 272 IF: THEN: assert the defendant is liable theory of strict-liability. under the FIGURE 1. Definition of Strict Liability in ROSIE III. LEGAL MODEL - -- The model of legal decisionmaking we are building will contain five basic types of rules: those based on formal doctrine, informal princi- pies, strategies, subjective considerations and secondary effects (see Figure 2). The formal doc- trine evolves from court decisions and statutes, while the informal principles, strategies, etc. are shaped by example and experience. Sources for these rules include legal literature, case his- tories and interviews with experts. By separating the rules as described we can study both the relevant inference mechanisms and the influence of each type of knowledge on the decisionmaking pro- cess. We are using our model of legal decisionmaking to systematically describe how legal practitioners reach settlement decisions and to test the effect of changes in the legal system on these decisions. Individual cases are analyzed by comparing the chains of reasoning (the chains of rules) that lead to the outcomes to similar chains in prototypical cases. This helps clarify the relationships exist- FORMAL DOCTRINE: rules used as the basis for legal judgements such as legislation and common law. INFORMAL PRINCIPLES: rules that don't carry the weight of formal law but are generally agreed upon by legal practitioners. This includes ambiguous concepts (e.g., reasonable and proper) and generally accepted practices (e.g., pain and suffering = 3 J; medical expenses). STRATEGIES: methods used by legal practitioners to accomplish a goal, e.g., proving a product defective. SUBJECTIVE CONSIDERATIONS: rules that anticipate the subjective responses of people involved in legal interactions, e.g., the effect of plaintiff attractiveness on the amount of money awarded, or the effects of extreme injuries on liability decisions. SECONDARY EFFECTS: rules that describe the interactions between rules, e.g., a change in the law from contributory negligence to comparative negligence may change other rules such as strategies for settlement or anticipated behavior of juries. FIGURE 2. Components of Legal Decisionmaking ing between the formal doctrine, informal practices and strategies used in the decisionmaking. We are examining the effects of changes in legal doctrine, procedures and strategies on the processing of cases by modifying appropriate rules in the model and noting the effect on the operation of the model when applied to a body of selected test cases. This can provide insights that will suggest useful changes in legal doctrine and practices. Our current implementation of LDS is a small prototype model of legal decisionmaking containing rules representing negligence and liability laws. This prototype contains rules describing formal doctrine and informal principles in product liabil- ity. Future versions of the system will incor- porate the other rule types shown in Figure 2. The model consists of approximately 90 rules, half of which represent legal doctrine and principles. Given a description of a product liability case the model attempts to determine what theory of liabili- ty applies, whether or not the defendant is liable, how much the case is worth, and what an equitable value for settlement would be. Once a decision is reached the user may ask for an explanation in terms of the rules used to reach the decision. We will now describe the use of LDS to test the effect of a legislative change on a case out- come. The case is briefly summarized in Figure 3, while the operation of the model on this case is illustrated in Figure 4. The system was first ap- plied using the definition of strict liability given in Figure 1. It was determined that the de- fendant was partially liable for damages under the theory of comparative negligence, with the amount of liability lying somewhere between $21,000 and 273 The plaintiff was cleaning a bathtub drain with a liquid cleaner when the cleaner exploded out of the drain causing severe burns and permanent scarring to his left arm. Medical expenses for the plaintiff were $6000, and he was unable to work for 200 working days, during which time his rate of pay was $47 per day. The cleaner was manufactured and sold by the defendant, Stanway Chemical Company. The contents of the product were judged not to be defective by experts retained by the defendant. The product's label warned of potentially explosive chemical reactions from improper use of the product, but did not give a satisfactory description of means to avoid chemical reactions. The plaintiff was familiar with the product but did not flush out the drain before using the cleaner. The amount of the claim was $40,000. FIGURE 3. Description of Drain cleaner Case (Note: the model actually used a much more detailed description of of the case than is shown here-.) use was reasonable QI IU PI U)JW $29,000. The case was valued between $35,000 and $41,000. After the definition of strict liability was modified to state that the product must be un- reasonably dangerous for strict liability to apply, the defendant was found to be not liable. In this prototype implementation of LDS a somewhat more restrictive ROSIE rule syntax was used than is shown in Figure 1. v. CONCLUSIONS Our preliminary work with LDS has demonstrated the feasibility of applying rule-based modeling techniques to the product liability area. In spite of the inherent complexity of product liability law, the number of basic concepts manipulated by the rules is easily handled (in the hundreds), while the number of rules required to adequately represent legal doctrine and strategies is manage- able (in the thousands). The rules that represent legal doctrine in this area are basically declarative in nature. use was * rl lforeseeable no strict Droduct was defective defendant manufactured product - product not unreasonably dangerous victim’s responsibility = .4 victim was not a minor victim knew of hazard \- liability y - - - - defenclant -la/, Y “9 . . _ ,iabi.ity total amount of loss is between $35,000 and $41,000 location not dangerous z r6 medical expenses were $6136 lost 228 working days __IIc base pay of $47 per day / victim iparative :ligence pa I defendant’s liabilitv = .6 FIGURE 4. Inference Process for Drain Cleaner Case (Crosshatched area shows inference before law change) 274 Most of them are easily represented as definitions with complex antecedents and simple consequents that name the concept being defined. Rules of this sort can be organized as relatively unordered sets that are processed with a simple control scheme. Most of the action takes place in calls to other rule sets representing definitions of terms used by the initial set. This simple control structure fa- cilitates rule modification and explanation. In this application area improved methods are needed for dealing with vague or ambiguous concepts used in the rules. It is sometimes difficult to decide whether or not these concepts are applicable in a particular case, e.g., whether the use of the product was actually "reasonable and proper." Pos- sibilities include gradual refinement: a query scheme involving presenting the user with increas- ingly specific sets of questions, each of which may have ambiguous terms that will be further refined by even more specific query lists, and analogy: displaying case histories involving similar proto- typical concepts and having the user select the one closest to the term in question. ACKNOWLEDGMENTS This work has been supported by a grant from the Civil Justice Institute at the Rand Corpora- tion, Santa Monica, California. [71 [al [91 (101 [Ill [=I Process: A Computer That Uses Regulations and Statutes to Draft Legal Documents." American Bar Foundation Research Journal, No. 1, (1979) 3-81. -- Hafner, C.D., "Representation of Knowledge in a Legal Information Retrieval System" In Research and Development in Information Re- trieval. Proceedings of the Third Annual SI- GIR Conference, Cambridge, England, 1980. Waterman, D.A., Anderson, R.H., Hayes-Roth, F s:;. , Klahr, P., Martins, G., and Rosenschein, Design of 2 Rule-Oriented System for Implementing Expertise. - N-~~s~-ARPA, The Rand Corporation, Santa Monica, California, 1979. Anderson, R.H., and Gillogly, J.J., Rand In- - - telligent Terminal Agent (RITA): Design Phi- losophy. R-1809-ARPA, The Rand Corporation, Santa Monica, California, 1976. Shortliffe, E.H., Computer-Based Medical Con- --- sultations: MYCIN. New York: American El- sevier, 1976. Waterman, D.A., and Hayes-Roth, F. Pattern- Directed Inference Systems. New York: Academic Press, 1978. Sussman, G.J. "Electrical Circuit Design: A Problem for Artificial 1977 Intelligence Research." In Proceedings of the 5th Inter- --- - national Joint Conference on Artificial In- telligence, Cambridge, Massachusetts, 1977, 894-900. REFERENCES [II [21 [31 [41 [51 [61 Waterman, D. A., "User-Oriented Systems for Capturing Expertise: A Rule-Based Approach." In D. Michie (ed.) Expert Systems in the Mi- -- - cro Electronic &. Press, 1979. Edinburgh University Jones, S., Mason P.J., & Stamper, R.K. "LEGOL-2.0: A Relational Specification Language for Complex Rules." Information Sys- terns, 4:4, (1979). Meldman, J. A. (( A Structural Model for Computer-aided Legal Analysis." Journal of Computers and Law, 6 (1977) 27-71. - - -- Popp, W.G., and Schlink, B., "JUDITH: A Com- puter Program to Advise Lawyers in Reasoning a Case." Jurimetrics Journal, 15:4 (1975) 303-314. McCarty, L.T., and Sridharan, N.S., "The Representation of an Evolving System of Legal Concepts, Part One: Logical Templates" In Proceedings of the Third National Conference - _-- of the Canadian -- Society for Computational Studies of Intelligence', Columbia,T980. Victoria, British Sprowl, J.A., "Automating the Legal Reasoning 275	1980	32
26	EXPLOITING A DOMAIN MODEL IN AN EXPERT SPECTRAL ANALYSIS PROGRAM David R. Barstow Schlumberger-Doll Research, Ridgefield, Connecticut 06877 ABSTRACT Gamma ray activation spectra are used by nuclear physicists to identify the elemental composition of unknown substances. Neutron bombardment causes some of the atoms of a sample to change into unstable isotopes, which then decay, emitting gamma radiation at characteristic energies and intensities, By identifying these isotopes, the composition of the original substance can be determined. GAMMA is an expert system for performing this interpretation task. It has a detailed model of its domain and can exploit this model for a variety of purposes, including ratings for individual isotopes and elements, ratings based on multiple spectra, complete interpretations, and even calibration. GAMMA's performance is generally quite good when compared with human performance. 1. INTRODUCTION Gamma ray spectra are commonly used by nuclear physicists to identify the elemental composition of a substance. One kind of gamma ray spectrum (an "activation spectrum") is produced by bombarding a sample of the substance with neutrons. This causes certain changes in some of the atoms in the sample, many of which result in unstable isotopes that then begin to decay. As they decay, the unstable isotopes emit gamma rays at characteristic energies and intensities. By measuring these, the unstable isotopes (and from these, the elements of the original sample) can be identified. For example, Figure 1 shows such a spectrum, with peaks identified and labeled by a physicist. In this case, the peaks were produced by emissions from the isotopes Na-24, Cl-37, and S-37; the original substance was a sample of salt. An expert system, called GAMMA, has been developed to perform this task, and GAMMA's performance compares well with human interpreters. The basic strategy employed in developing GAMMA was to develop a detailed model of the domain and then to exploit this model for a variety of tasks and situations. Early work on GAMMA was discussed in another paper[ll; in this paper, recent progress will be discussed. 2. The Domain Model GAMMA's domain model was described in detail in the earlier paper and will only be summarized here. Basically, the process that produces gamma ray activation spectra can be seen at six different levels as follows: Figure 1: Gamma Ray Activation Spectrum 276 (1) elements in the original sample 3.1. Isotopic and Elemental Ratings (2) isotopes in the original sample (3) isotopes after bombardment by neutrons (4) decays of unstable isotopes (5) gamma ray emissions during decay (6) g-a ray detections during decay Level 6 represents the actual spectrum and level 1 represents the ultimate goal of the interpretation. Level 3 is a convenient intermediate level used by most practicing nuclear physicists. GAMMA’s use of this domain model involves considering hypotheses at each of the levels. A hypothesis consists of a set of triples, each triple consisting of an object appropriate to the level (e.g., naturally occurring isotope for level 2; gamma ray emissions at a particular energy for level 5), an estimated concentration for the object (e.g., the number of decays of a given unstable isotope for level 4), and a label which encodes the path from level 1 to the triple (e.g., “NA-23/NG/NA-24,’ denoting that the unstable isotope Na-24 was produced when the naturally occurring isotope Na-23 underwent an N-y transition). Relationships between levels can be expressed in terms of several formulae that have been derived from both theoretical considerations and empirical observations. These formulae involve such parameters as the likelihood of particular isotopic transitions during neutron bombardment, the half-lives of unstable isotopes, and the characteristic gamma ray energies and intensities for different isotopes. Nuclear physicists consult published reports when they need such information; the data from one such source[21 has been converted into a LISP data base for GAMMA’s use. Further details of the formulae and data base are available elsewherec 11. In GAMMA’s case, the formulae are all used predictively; that is, given an hypothesis at one level, the appropriate formula can be used to predict hypotheses at the next level down. By chaining such predictions together, GAMMA can go from hypothetical interpretations at levels 1 or 3 down to predicted gamma ray detections that can be compared against spectral data. 3. Applications of the Doglain Model The accuracy with which predictions can be made and the high resolution of this particular detector enable GAMMA to exploit the domain model in a variety of tasks and situations. Some of these were discussed earlierElI, and will be mentioned here only briefly. GAMMA was first used to “rate,, the likelihood that any particular unstable isotope was present after neutron bombardment. This was done by hypothesizing decays of that isotope at level 4, predicting detections at level 6, and using an evaluation function to compare the predictions with the spectral data. The evaluation function was designed to take positive and negative evidence into account, allowing both for background radiation and for noise and errors in the prediction and detection processes. The peaks were individually rated for both energy and intensity, and the final rating was the average of the individual ratings, weighted by intensity (i.e., stronger peaks were more important). When a predicted peak had a corresponding detected peak, a positive rating was given; when no peak was detected, a negative rating was assessed, unless the predicted intensity was low enough that the peak could have been obscured by background radiation. Noise and errors were taken into account by using what we call the trapezoidal rule. For example, the trapezoidal rule for peak energies is shown in Figure 2. If a peak was predicted at energy E, then a detected peak within the range (E-6 E+b > was considered a perfect match, peaks out&e thJ range (E- ’ all, and peaks in 8; E?a!$e~e~~-~t ?Za!?:edazdt (E+& E+h2)were scaled to provide 2a’ con&nuous fun&on. Such trapezoidal rules were used throughout GAMMA’s evaluation function, and has proved quite adequate. GAMMA’s performance at the isotopic rating task was moderately good compared with that of human experts: although it gave high ratings to isotopes identified by experts, it also occasionally gave such ratings to isotopes considered implausible by the experts. GAMMA’s second task was to do a similar rating for elements in the original sample (i.e., hypotheses at level 1). The same predict-and-match technique was used, and GAMMA, s performance was again moderately good, although not quite as good as in the isotopic case: fewer implausible elements were rated but some elements identified by the human experts received low ratings. This was due largely to certain simplifying assumptions in the formulae relating levels 2 through 4. Further details of GAMMA’s rating scheme are given elsewhere[ll. Recently, GAMMA’s repertoire has been expanded to include several other tasks, and its performance seems to have improved with age. rating 1.0 -1 .o -I -Gy+ A-A-+- energy 62 blE% b2 Figure 2: Trapezoidal Rule for Peak Energies 277 3.2. Ratings for Multiple Spectra GAMMA's next major task was to do similar ratings for individual isotopes and elements, but to do so on the basis of multiple spectra: in a typical experimental situation, not one but several spectra are recorded, each for a different time interval. Generally, the first few spectra are for comparatively short time periods (10 to 30 seconds), and the later spectra may be for periods as long as several hours. The primary advantage of multiple spectra is that they permit greater use of half-life information: unstable isotopes with short half-lives will appear on the earlier spectra but not on the later ones; isotopes with longer half-lives emit ganrma rays at roughly constant rates, so they appear most distinctly on the later spectra (for which the detection time is longer). The technique used by GAMMA is to find the isotopic (or elemental) ratings for the individual spectra, use these to hypothesize an initial concentration for the isotope, redo the predictions based on this concentration, and finally combine the ratings for these predictions into a single overall rating. The hypothetical concentration for an isotope is determined by considering all spectra for which the isotope's rating is sufficiently high, taking the concentrations (byproducts of the original prediction-and-match rating) that agree sufficiently well (within one order of magnitude), and then computing the average. This technique is designed to ignore those results which, for any of several reasons, deviate from the norm and in practice seems to work quite well. Given this hypothesized concentration, the prediction-and-match rating is again computed for each individual spectrum. These ratings are then averaged to determine the overall ,,multiple spectra,, rating for the isotope. In our first attempt to average these ratings, we weighted them by the total predicted intensity for a spectrum, as was done for ratings within individual spectra. But this seemed to attach too much weight intensities, to spectra with high predicted so on our second attempt, we took the simple average of the ratings for all spectra for which the evidence was significant (either positive or negative), and the results were much better. (Interestingly, the first symptom of this problem was due to an INTERLISP error: under certain circumstances, INTERLISP ignores floating point overflow and underflow, thereby producing a very large number when multiplying two very small ones. With simple averaging, such isolated erroneous computations no longer have much overall effect. In fact, we now take this as a maxim: no numeric rating scheme should depend too heavily on any single data point.) GAMMA's performance on multiple spectra is generally much better than on individual spectra, primarily because of the value of half-life information. GAMMA's ratings generally compare well with those of human experts, and implausible isotopes (Or elements) are only rarely given high ratings. 3.3. Producing a Complete Interpretation The major problem with the tasks described so far is that the ratings are given to isotopes and elements as if they were totally independent of each other. The fact that the same peak may be caused by emissions from two different isotopes does not detract from the rating of either one. The ultimate interpretation of spectral data should not be ratings for individual elements, but rather a set of elements (and concentrations) which, taken together, explain the data well. A first pass at coming up with such a complete interpretation might be to take all and only those elements with sufficiently high ratings, but that does not take into account the interaction between the elements, and is simply inadequate. GAMMA's solution to this problem is essentially a hill-climbing algorithm designed to maximize an "interpretation measure,,. For this algorithm, a complete interpretation is defined to be a set of <element, concentration> pairs, and a mapping of detected peaks to sets of labels. (The labels describe the path from one of the <element, concentration> pairs to the detected peak. Under this definition, the same detected peak may have several different labels, a situation which actually occurs in the spectra under consideration.) The interpretation measure that GAMMA currently uses is based on two different considerations. First, the individual spectra are rated in terms of (1) how many peaks have no labels (i.e., are there peaks which are not explained by the interpretation?), (2) how many labels have no peaks (i.e., are there predictions which do not appear in the detected spectra?), and (3) how well the peaks and associated labels match (i.e., do the energies and intensities of the detected peaks match well with the energies and intensities predicted for the associated labels?). The second consideration is that the relative concentrations of the elements be plausible. This is used only as negative evidence: if the concentration of an element is high (relative to the concentrations of the other elements), but the rating for that element is low, then the interpretation is suspect, since the detector and model can be expected to be quite accurate for relatively pure substances; if the concentration is below a certain threshold, then the interpretation is also suspect, since the detector simply cannot be expected to find elements in such smali concentrations: The task is thus to find the set of <element, concentration> pairs that maximizes this measure. GAMMA uses the following hill-climbing algorithm: INTERPRETATION := 0; CANDIDATES := {<element, concentration> I rating is above a threshold); consider all interpretations formed by moving one element from CANDIDATES to INTERPRETATION or from INTERPRETATION to CANDIDATES; if no such interpretation increases the measure then quit else select that which maximizes the measure; update INTERPRETATION and CANDIDATES; repeat. 278 While we have no theoretical basis for claiming that this algorithm does, indeed, find the subset of candidates with maximal measure, our experience indicates that it performs very well, and the interpretations that GAMMA produces are quite good. 5. References Cl1 Barstow, D.R. Knowledge engineering in nuclear physics. In Sixth International Joint Conference on Artificial Intelligence, pages 34-36.- Stanford Computer Science Department, 1979 3.4. Calibration GAMMA’s final task is to calibrate the spectra. Up to this point, it has been assumed that the input spectra have already been calibrated (spectral channels have been associated with gamma ray energies), and this is a task which has hitherto been performed by physicists before the data are given to GAMMA. We have not yet completed the attempts at solving this problem, but our first results are encouraging : we have developed a technique for recalibrating a spectrum more precisely, given an initial approximate calibration. The basic technique is to use a set of good calibration priori by physicists) isotopes (as identified a and look for any which satisfy two criteria: (1) there is exactly one match within a given range of the initial calibration; (2) that match gets a fairly high rating. A linear least squares fit of these points gives the recalibration. Experience with this approach is quite good, and it is currently used to correct for calibration differences among the individual spectra within a set of spectra. c21 Erdimann , G . Neutron Activation Tables. WChemie , New York, 1976 . c31 Shortliffe, E.H. MYCIN: Computer-Based Medical Consultations. American Elsevie-w-1975 . Our plan for finding the initial calibration is to take a sum spectrun for the entire set (thereby increasing the data/noise ratio) and apply a similar strategy: find any ,,calibration,, isotope which has exactly one match anywhere on the sum spectrum and for which the match rating is quite high. We hope to know soon whether this approach will succeed. 4. Shallow and Deep Domain Models GAMMA’s success is due largely to its use of a relatively detailed model of its domain. This may be compared with systems such as MYCINC31 whose success is due largely to shallow models that encode, in a sense, an expert’s compiled version of a more detailed model that the expert may or may not know explicitly. In comparing these two approaches, several observations can be made. First, while a deep model can be put to great use (as it was in GAMMA), there are several circumstances in which a shallow domain model is necessary: (1) a deep model doesn’t exist (e.g., there is no computational theory of the entire hunan body) ; and (2) a deep model is not computationally feasible (e.g., one cannot hope to do weather prediction based on the fundamental properties of gases). Second, although shallow models will often suffice, it seems likely that future expert systems based on shallow models will require access to deep models for difficult cases. Third, “deep,, and ,,shallow,, are obviously relative terms : GAMMA’s ,deep model is shallow when viewed as a model of subatomic physics. The relationship between deep and shallow models seems to be an important topic for future work on expert systems. 279	1980	33
27	Project EPISTLE: A system for the Automatic Analysis of Business Correspondence Lance A. Miller IBM Thomas J. Watson Research Center Yorktown Heights, New York 1 0598 ABSTRACT: The developing system described here is plan- ned to provide the business executive with useful applications for the computer processing of correspondence in the office environment. Applications will include the synopsis and abstraction of incoming mail and a variety of critiques of newly-generated letters, all based upon the capability of understanding the natural language text at least to a level corresponding to customary business communication. Succes- sive sections of the paper describe the Background and Prior Work, the planned System Output, and Implementation. I. BACKGROUND AND PRIOR WORK We conclude from these behavioral findings that there are indeed extensive regularities in the characteristics of business letters, determined primarily by the purpose objec- tives. It is these constraints that most strongly indicate to us the feasibility of developing automatic means for recognizing content-themes and purposes from the letter text (as well as the converse, generating letter text from information about purposes). Other analyses have been undertaken to estimate the linguistic complexity and regularities of the texts. The aver- age letter appears to contain 8 sentences, with an average of 18 words each; in the 400 letter-bodies there are roughly 57900 words and 4500 unique words totaL An ongoing hand analysis of the syntactic structure of sentences in a 50-letter sample reveals a relatively high frequency of subject-verb inversions (about 1 per letter) and complex lengthy complementizers (l-4 per letter). These features, along with very frequent noun phrase and sentence coordina- tion, accompanied by a wide variety of grammatical but un- systematic structure deletions, indicate an exceptionally high level of grammatical complexity of our texts. With respect to overall text syntax we have analyzed 10 letters for text cohe- sion, using a modification of Halliday and Hasan’s coding scheme E41; 82 percent of the instances of cohesion detect- ed were accounted for by 4 categories: lexical repetitions (29%), pronouns (28%), nominal substitutions (9%, e.g., “one”, “same”), and lexical collocations (words related via their semantics, 16%). In an extension of this discourse structure analysis we are analyzing 50 letters, coding all occurrences of functional nouns in terms of (1) the grammat- ical case function served and (2) the cohesive relation to prior nouns. Preliminary results indicate consistent patterns of case-shift and type of cohesion as a function of the prag- matic and content themes. The results of these linguistic analyses will help determine the strategy ultimately adopted for selecting surface parses and meaning interpretations. II. SYSTEM OUTPUT The planned system will provide the following for each letter in our database: (1) surface syntactic parses for each sen- tence; (2) meaning interpretations for each sentence, adjust- ed to the context of prior sentences; (3) a condensed synop- sis of the overall meaning content of the letter; (4) a critique of each letter’s spelling, punctuation, and grammaticality; (5) a mapping of the meaning content onto common business communication themes; and (6) some characterization of the author’s style and tone. In addition to the above, we plan to develop a limited facility to generate short letters of a certain type (e.g., information requests) conforming to a particular author’s normal “style” and “tone”. III. IMPL.EMENTATION COMPONENTS A. Semantic Representations: Many of the lexical items in our texts appear to have two or more literal word-sense usages (as well as occasional non-literal ones); it also appears that the discriminating semantic features among highly- related lexical items cannot be ignored if the intended letter nuances are to be preserved. We therefore do not expect much reduction in cardinality when mapping from the lexical to the concept space; we also anticipate that our representa- tions will have to be unusually rich, in terms of both a large number of features distinguishing the concepts underlying lexical items and the capability to relate different concepts together. Among the most important anticipated semantic features are those describing the prcxmdhns and cmse- quences of ACTIONS and those characterizing the internal states of ACTORS (e.g., their intentions, expectations, and reactions). B. Parsing: We will employ a system called NLP as the basic “operating system” for our application development. This system has been used for other understanding projects and provides all of the general components required, including a word-stem dictionary, a parser, knowledge representation, and natural language generation I: 25 1. In particular, the parser proceeds left to right, character by character, in proc- essing a sentence, generating all possible descriptions of text segments in a bottom-up fashion by application of rules from an augmented phrase structure grammar -- essentially a set of context-free rules augmented with arbitrary conditions and structure-building actions. In writing the grammar we are attempting to keep to an absolute minimum the use of se- mantic information, to increase the applicability of the parser over a variety of semantic domains. Our general strategy to minimize the number of surface parses (and increase parsing efficiency) is to attach lower syntactic constituents (e.g., post-nominal prepositional phrases) to the highest-possible unit (e.g., to the verb-phrase rather than the noun-phrase), with the final decision as to the most appropriate attachment to be resolved by the interpretation components. C. Meaning-Assignment: Our planned strategy for choosing the meaning to assign to a sentence is basically to find that action-concept whose case-relations are most completely “satisfied” by all of the concepts implied by the sentence. In the expected frequent case of only partial “fits” to several action-concepts, prefer- ence among these will be based on several factors, including: (1) the number of case relations of a concept “filled” or “unfilled” by elements of the present (or prior) text, and the relative importance in the intensional definition of each of these; (2) the “directness” of the mappings of text segments to underlying concepts; and (3) the syntactic structure of the sentence (e.g., syntactically “higher” components usually will be preferred to “lower” ones). D. Text-Interpretation: We propose to keep separate lists of each action-concept and entity-concept encountered in the text. Following the meaning-assignment to a sentence, the sentence will be re- examined to determine if it supplies qualification information for any prior-mentioned action or entity; if so, these separate representations will be so augmented, a process called “updating”. Statistics of each such updating of information will be kept for each sentence for subsequent characteriza- tion of. style. Next, these separate entity/action representa- tions will be examined directly to determine whether they can be combined as elements of some broader concept. By this process we will therefore be able to update and condense our representations as we go along, facilitating eventual synopsis and abstraction of content-themes. In addition to the above semantic interpretations for the complete text, we will also build up a composite representa- tion of the syntactic structure of the text. We, first, are hopeful of being able to discover a relatively small number of schema for characterizing syntactic structures within sen- tences:wetllenheJieYethatthe~Qfleuer~canbe accounted for these schemas. in terms of E. Adaptation: frequently occurring patterns of to As a unique feature, we plan to implement the capability dynamically modify or adapt our system so as to change the manner in which word-senses are selected or meanings assigned as a function of the system’s experience with various kinds of texts. This would be accomplished by assigning “weight” attributes to each lexical item and to each underly- ing concept (and its attribute-values); the weight-values along the path finally selected for mapping text to concepts would then be incremented by some amount. Given that preference ordering of text to concept paths is determined by such overall path-weights, the system could thus achieve self-adaptation to word-usages of particular application do- mains. This facility could also be employed to characterize individual authors’ styles. F. Abstraction and Critique of - Letters: Concerning Content-Themes and Purposes, we plan to map the system’s meaning interpretations onto a set of com- mon business content-themes and communication purposes, and we are presently conducting behavioral and analytical studies to determine these. With respect to Grammaticality, we anticipate being able to detect incomplete sentences, subject-verb disagreements, and inappropriate shifts in verb tenses; in addition, we will be able to identify ambiguities and some instances of clearly “awkward” syntax. Spelling errors of the “non-word” type are easily caught, and certain spelling errors in which the misspelled word is in the diction- ary may also be caught if they contain sufficient syntactic information. In addition, some fraction of “spelling” errors involving semantically inappropriate words should be detecta- ble. Finally, we may be able to discover a number of Punctuation errors. The last aspect of critiques is that of style and tone. We are aware of the several “indices” for measuring various aspects of these but consider them to be at best very crude indicators C 61. As a starting point we have identified five dimensions for each concept, and we will implement the capability to assess texts on these dimensions until we are better informed. For S~J&, defined as “the organizational strategy for conveying content”, the dimensions are: sentence precision, sentence readability, reference clarity, information- value, and cohesion. Tone, defined as “the connotations of interpersonal attitudes”, is to be rated on the dimensions of: personal-ness, positive-ness, informal-ness, concrete-ness, and strength. We plan to output and highlight those text segments which fall below a certain level of acceptability on these measures. REFERENCES Cl1 Miller, L. A. “Behavioral studies of the programming process. ” IBM Research Report RC 7367, 1978. c 21 Heidorn, G. E. “Augmented phrase structure gram- mars”. In Nash-Webber, B.L. and Schank, R. C. (Eds.), Theoretical Issues in Natural Language --- - - -- Processing. Association for Computational Linguis- t.ic& .hle+ 1975. 281 c31 Miller, L. A. and Daiute, C. “A taxonomic analysis of c51 Heidorn, G. E. “Automatic programming through natu- business letters”. IBM Research Report, in prepara- ral language dialogue: A survey”. IBM Journal of --- tion, 1980. Research and Development, 1976, 20 302-3 13. -- - 141 Halliday, M. A. K. and Hasan, R. Cohesion in English. -- London: Longman Group Ltd., 1976. C61 Dyer, F. C. Executive’s Guide to Effective Speaking --- and Writing. Englewood Cliffs, N. J.: Prentice-Hall, Inc., 1962. 282	1980	34
28	A KNOWLEDGE BASED DESIGN SYSTEM FOR DIGITAL ELECTRONICS* Milton R. Grinberg Department of Computer Science University of Maryland College Park, Maryland 20742 1. Overview and Goals of the SADD System v----p the goal of a human ex e;ia;; such as "build a digits P display interface into a This is a problem solving activity in which the roblem expert's general purpose R solving abilities interact with his rich nowledge of the digital world. initial ideas By translatin through successively more !! the re ined "sketches", the human expert gradually arrives at a chip-level digital schematic that realizes the initial goals, and which may have bugs that will only be discovered after the circuit has been simulated or built and tested. The Semi-Automatic Digital Designer (SADD) is an experimental interactive knowledge-based design s stem, II whose domain of expertise is digital e ectronics. First, I want The SADD project has two goals. to provide an intermediate digital design problem solver for which a human expert can interactively provide the high-level functional descriptions-of-a circuit and which can take . high-level circuit descri;tion and refine it in?: ciycuit Second, schematic that performs the required task. corn uter guman iii I am ;EaEyptmg to discover and express in that_.knowledge esig;t; an expert in.digital that makesT;F; desi n. * includes f eneric information K or each hi~h-le~~l d;g,f;zdsfun;;;c!;d(e.g., counter, clock) transform these functional descriptions into r%izable circuits. As a first case study in relative1 desi n, sophisticated TV video % I adopted a called 4: isplay circuit t e Screensplitter. This is a real circuit of moderate overall complexity. 2. Digital Design -- How does a digital desi designing a circuit? The % ner go about the task of esigner generally starts with a well-defined goal for a circuit with only an ill-defined solution for accomplis$;ng Using his design experience, that goal. designer can pinpoint some of the required and hence pivotal components needed in the circuit. The circuit is slowly sketched around these The pivotal sketching process allows components. specify where the inputs come from the designer to and where the outputs go. concerning The designer of:;: makes notes characteristics of I$f;;;g this sketch phase, components. the expert discovers that components are needed and adds them to the design. each Eventually the designer starts to refine component by selecting and interconnecting chips to implement the component. What are the "primitives" that a designer uses? At the im leme$afion(i.$., final circui;idlevel R $h;;; a;zdt ree prim+tives : (2)(l) a chip its wire and si nal. (3) a f The"%!;tisPin?&rction those &tputs are de ined by its current state ;~~,,a;;,,current A wire i?!iE:siefining anizleFtrica1 equivalence. A ST:: physical f is the electrical information present on a wire. At the sketch ad level "primitives": (17 a there are two additional f unction and (2) a connection. This research is sup orted b the Office of Naval Research under Grant 800014-765-0477. A function performs some desi nated digital task. I have concentrated on nine 5ifferent functions (i.e., counter, shifter, memory, combinational d divider, selector, and clockj in the Screensplitter. These are components that a designer uses in the A connection is an information path between- two functions. It identifies where information flows and when it is allowed to flow. 3. SADD Organization There are three distinct design phases in the In the specification acquistion phase, the user describes the functionatsstructure offEk;e;ircuit in English. During . phase, are introduced relevant information about the frames is filled in, and the interrelationship among the frames established. From builds a model of this description, SADD the semantic net. circuit, expressed as a component is selected using the conceptual function &zracteristics (which may-have to be deduced from characteristics provided by the designer and the functions relationship to the other functions in the model). the circuit schematic is implemented using '%: selected and conceptual function description. strategy In the circuit simulation phase, the correctness of the circuit is determined by simulating the circuit on a conceptual simulator. circuit proves not to If the designed fulfill the goal of the desi ner during the simulation phase, it can be modi led and redesigned. 3. 3.1. Screensplitter Circuit In order to develop SADD, the Screensplitter was chosen as a benchmark. Fig. 1 illustrates one of DCCBR READ DM ADDRESS DCCBR LOAD STCOL Fig. 1. Schematic for the DCC Counter the 12 logical components (the Display Character 283 Counter) from the original Screensplitter circuit schematic. the circuit schematic design scenerYii%s develoned a verbal that descsibes the funcrional components, characteristics of these functional components and the interconnections. Fig. 2 shows the relevant portion of the input 1. 2. 3. 4. 5. THE DISPLAY CHARACTER COUNTER (DCC) COUNTS FROM . _ 0 TO 3519. WHEN THE COUNT OF THE SLC EQUALS 4, THE COUNT OF THE DCC CAPTURED-IN THE LOAD OF THE DCC-BASE REGISTER (DCCBR) WHEN THE HORIZONTAL BLANK (HBLANK) BEGINS. _ THE DCC CAPTURED-FROM THE DCC-BASE REGISTER (DCCBR) WHEN HORIZONTAL BLANK (HBLANK) BEGINS. EACH-TIME THE COUNT OF THE PIXEL COUNTER (PC) EQUALS 5 ENDS, THE DCC INCREMENTS. THE COUNT OF THE DCC CAPTURED-IN THE ADDRESS OF THE DISPLAY MEMORY (DM). Fig. 2. Scenerio for the DCC Counter description for the Display Character Counter. The complete scenerio is 41 sentences. 3.2. Parser A phrase-keyword parser using procedurally encod ed case-framewo rks for the verbs was develoned to interpret the input. Phrase-keyword means that the parser is $way;r;;z;ng to build up one of seven digitalt$gE?gn wor P that are common in the d. identify It uz;; the keyworcl; tot:::: phrases. the beginnings endings After a sentence has been Darsed into Dhrases. the procedure associated w!ith the verb' in the sentence is verifies applied. This processing first that the sentence is acceptable (that the phrase semantically for the verb types Then those are legitimate currently in circuit objects not t h g circuit model are introduced into the model, and the verb's manipulations of the model are processed. The parser uses 5 directives to manipulate the model. These directives either add new information to the model or interrelate existing parts of the model. The directives are: (a> (b) cc> (d) (e> 3.3. _ To Specify a function - a function (e.g., clock, counter) is introduced into the model. Assi n a value ~0 a function's.aspect -. one i?i o,fvt,l;efunction s characteristics 1s assigned Define i conceptual si assumed to be a port nal - a global name is o B either a known or unknown function. Define the source of a conceptual signal - the source function of a global signal is identified. Make a connection - identify an information path between two functions and a condition gating the information flow. Example - -- Specification Acquisition illustrate the specification acquistion g* hase, the five sentences shown above for the isplay Character Counter are reduced to their effects on the model. The effects for each sentence are preceded by a letter which references the corresponding directive type from the above list of directives. 1. THE DISPLAY CHARACTER COUNTER (DCC) COUNTS FROM . . ?aT"Iz?k%uce a COUNTER function named DISPLAY CHARACTER COUNTER (DCC). (b) Assi n a value to the Aspect of (b 3519). COUNT-SEQUENCE 2. WHEN THE COUNT OF THE SLC EOUALS 4. THE COUNT OF THE DCC CAPTURED-IN THE LOAD OF zTHE DCC-BASE REGISTER t (DCCBR) WHEN THE HORIZONTAL BLANK HBLANK) BEGINS. c) Define a conceptual signal named HORIZONTAL ?IANK (HBLANK). 3. 4. 5. (e) Make a connection between (DCCBR LOAD) and (DCC COUNT) under the condition (AND (HIGH SLC DS~) (RISING UNKNOWN HBLANK)). THE DCC CAPTURED-FROM THE DCC-BASE REGISTER DCCBR WHEN HORIZONTAL BLANK (HBLANK) BEGINS. [e) EPIC; cpon;;ti;;de:etween (DCC LOAD) and the condition (RISING UNKNOWN HBLANK). EACH-TIME THE COUNT OF THE PIXEL COUNTER (PC) . _ EQUALS 5 ENDS, THE DCC INCREMENTS. (a) Introduce COUNTER function named PIXEL COUNTER (PC). (e) Make a connection between (DCC COUNT-UP) and T under the condition (FALLING PC DSS). THE COUNT OF THE DCC CAPTURED-IN THE ADDRESS OF THE DISPLAY MEMORY (DM). (a) Introduce MEMORY function named DISPLAY . ~ MEMORY (DM). (e) Make a connection between (DM ADDRESS) and (DCC COUNT). The model of the circuit after just these five sentences are processed is shown in Fig. 3. After (I I - HBLANK Fig. 3. State of the Model - (1, 2, 3, 4, and 5) this description, the specified and its DCC has been completely the next section. implementation is discussed in 3.4. Function Each function type has an associated frame structure that provides the prototypical knowledge about that function. There are two corn onents the function frame: the ASPECT and the 8 to ONPT. The ASPECT identifies the im ortant characteristics of the function which mig R t in the input scenerio. be mentioned by the user The CONPT identifies the ports (the input and output lines) that are associated with the function. The width of the CONPT is also identified aseeither 1 or the value of one of the ASPECTS of function. When a particoulfar function is introduced into the model, a co Y x the protot pe Eg. instantiated wit K is copied intFhkhe database the name of function. 4 is the instantiation of the DCC counter as represented in the model after sentences 1, 2, 3, 4, and 5 have been processed. The asterisked items are those processing. that were modified during the sentence The ex ert component E has a conce tual K view einer described. T e function of the frame with the- associated filled-;;easpect val;:; re,pL;ese:;," that conceptual view component. - exnert I 4. Designing a Circuit for a Function ---_ There are two phases involved ' the implementation of a function. First a method for implementing the function must be selected. Then the selected method must be processed to produce the chip-level design. 284 (COMPONENT DCC SADD COUNTER) * ‘TYPE DCC DEVICE (DISPLAY CHARACTER COUNTER)) * DESC-VAL WIDTH DCC 10-2) DESC-VAL COUNT-SE UENCE DCC (0 3519)) ASPECT WIDTH DCC 8 IL) ASPECT OUTPUT-CODING DCC NIL) ASPECT COUNT-SEQUENCE DCC NIL) ASPECT COUNT-DIRECTION DCC NIL) ASPECT DISTINGUISHED-STATES DCC NIL) ASPECT LOADABLE DCC NIL) ASPECT RESETABLE DCC NIL CONPT LOAD DCC WIDTH Gl 0)) 4 CONPT READ DCC WIDTH G265 G171)) CONPT COUNT-UP DCC 1 1 CONPT COUNT-DOWN DCC 1 NIL CONPT LOAD-LINE DCC 1 NIL) CONPT CARRY DCC 1 NIL) where IO-2 G150 G265 = 12 - (CONNECT 1 ~~~~~~ t - (CONNECT (DC t % (DM c Mom) (DCCBR C COUNT-UP)(:P CBR LOAD) (DCC ADDRESS) (Dd! READ) UNKNOWN HBLANK) 'D,g;TpC DSS)) IGH COMB1 OUTPUT COUNT) T) Fig. 4. DCC Counter Function - (1, 2, 3, 4, and 5) 4.1. Implementation Method The first ste F in the selection E recess is to deduce values or with the many of the ASP CTs associated function and to verify that there are no inconsistencies ' deductions made flZr t'k ASPECT values DCC are that it The is loadable, and that the count resetable, has a binary out ut, direction * K Te set of implementation methods aiiocii!ld with the type of function is then considered. An implementation method has 4 components: Prere uisites, Fig. 2 Eliminators, Traits, and Procedure. illustrates two im lementation methods for a counter, one based on a 7 e 161 loadable counter and the other on a 7493 counter. The Prerequisites are (STRATEGY BIN7493 COUNTER (PREREQUISITES I COUNT-DIRECTION UP OUTPUT-CODING BIN) (ELIMINATORS ) I (LOADABLE YES) mms (RESETABLE YES)) > PROCEDURE $BIN7493) Fig. 5. Example Counter Implementation Methods a list of ASPECTS and their values that must be valid for the function in order to select the method. In the BIN74161 function being im lementation.method, the implemente 8 must output coding that only counts up. re ;i;e a bi~~;~ 9 1s true for the BIN7493 implementation method. The Eliminators are a list of ASPECTS and their values which if matched by the function under im lementation, P eliminate that method from being se ected to implement the function. a list of ports that can be explicitly available if the method is selected. The Procedure is the recipe used to build the circuit. It is re resented procedurally in LISP code and is !i re erenced by name in Fig. 5. only Each method that can be used to implement a function is processed in the selection ph;s;liz checking the Prerequisites and Eliminators. of all acceptable methods is selection. compiled during the If there is only one acceptable method it is used to implement the function. If more than one method is acceDtable then the Traits of all candidates are used method. to find the most appropriate If this does not narrow the list to a single method then one user chooses the method. randomly chosen or the this examnle for the DCC assuming that there are I 5 "ZZilk%e two implementation methods from Fig. im lement 51 a counter, the BIN74161 method is tlt$ on y acceptable method and is selected. The procedure associated with the selected method is then run. This procedure consults the frame associated with the function and the connections information involving the function and from this constructs a real diiital interconnecting- circuit, introducin and collective y implement the function. 7 chips that 4.2. Implemented Circuit After the first levels of two phases, there are three the DCC design. is the DCC function and the A:e:;e,e;op level there connections. At the bottom level are the chins and wires used to im lement in erfaces f %,e Dee* The -intermediate level other two levels bv Drovidine. interconnections between chip pins or wire's and th: connection points on the function. Hence the hierarchy of the components is maintained and. if necessary, any circuit can be altered without much effect on the fragment rest of the circuit. The im lementation BIN7416P for the DCC as designed b the implementation method is illustrate in i Fig. 6. DCCBR READ DM ADDRESS DCCBR LOAD vcc Fig. 6. DCC Implementation 5. Conclusion the SADD is a general purpose design system based on ideas of structured, an interactive modular circuit design via user interface. The current 41 sentences in the innut scenerio have been run successfully thro 3 h the parser creating a database of approximately 6 0 entries. counters in the The design of the 3 Screensplitter are czmpleted and circuits functional1 used in the P equivalent to those Screensp ac tuaktz itter have been designed. design of a symbolic simulator is currently in progress. This simulator will allow the desi ner to test and debug the circuits and will camp ete f the design S&i?%?1 envisioned for SADD. When completed, extensible E rovide a generaioFurpose and design. knowledge- ased system digital The Traits are a list of ASPECTS and their values that are true if the method is selected and 285	1980	35
29	THEORY DIRECTED READING DIAGNOSIS RESEARCH USING COMPUTER SIMULATION Christian C. Wagner John F. Vinsonhaler The Institute for Research on Teaching Michigan State University East Lansing, Michigan 48823 Abstract A five year project studying the diagnosis and remediation of children with reading problems is described. The discussion includes observational studies of reading diagnosticians at work, observations of diagnostician training programs and computer simulation of theories about decision making in reading diagnosis. The results of the observational studies are mentioned and the theories and systems for computer simulated diagnosis are described. I Introduction The Institute for Research on Teaching is a federally funded project whose purpose is to investigate teaching, where teaching is conceptualized as an information processing task. The Clinical Studies project of the Institute is just finishing its first five year plan studying teachers who diagnose and remediate children with reading problems. This paper should serve as a short introduction of this work to the AI community. The Clinical Studies project has been primarily concerned with understanding reading diagnosis and remediation - whether performed by classroom teacher, reading specialist, learning disability specialist or school psychologist. Theories or models have been developed to account for the significant behaviors that occur when one of the teaching professionals works with a child. These theories have been tested against 1)direct observational studies of the specialists working with cases, 2)training studies observing the instruction of new specialists, and 3)computer simulation studies observing the behaviors implied by the theory through simulation. Results of these studies have shown that an individual's decision making may be very unreliable - suggesting that individual behavior may not warrant simulation. Before turning to the computer simulation studies, consider briefly the theory and results of the other studies. II Theories and Models The content-independent theory that attempts to account for the problem solving behavior of the clinicians is termed the Inquiry Theory Cll. The reading clinician-child interaction is viewed by this theory as follows: 1)The case is considered to be an information processing system that must perform a certain set of tasks. Some information processing abilities are critical to the adequate performance of these tasks. An adequate diagnosis of a problematic case must contain a statement relevant to each such critical ability. There may exist prerequisites for these critical abilities. An adequate diagnosis will include a statement relevant to all prerequisites for each deficient critical ability. Finally, a good diagnosis will include a statement of the history of the case that has led to current deficiencies. 2)The clinician must diagnose a case as described above. This is accomplished by the application of elementary information processing tasks to an understanding of how a case performs its function. The elementary tasks include ones such as hypothesis generation, hypothesis testing, cue collection, cue interpretation, etcetera [21. 286 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. processing task, the MORAL describes other significant aspects of behavior required such as the allocation of information processing capacity and conditions of information overload. At the current time the MORAL describes the critical abilities for human reading. It further details what other factors might effect these abilities and cause deficiences and how correction of any deficiency might be attempted. To date it appears to be quite effective in determining the reading problems in a case. The MORAL does not at the current time make any attempt to describe how reading comprehension takes place. Instead the various types of reading or listening comprehension tasks which people must perform to be good readers are listed. Of course the MORAL is at best incomplete and possibly incorrect. But it does serve to direct research and diagnose cases of reading difficulty. One final note with respect to theory - it is the intersection of these two relatively independent models (a content independent model of clinical problem solving and a content dependent model of the reading process) that is our area of concern. The questions that arise include: how do the specialists diagnose and treat children, how should they do it, how can they be trained to do what they should, what impact does the model of reading have on the decision making process, etc. III Observational and ApDlication Studies The observational studies of teaching professionals and clinicians in reading have basically indicated one thing - the specialists are not reliable. Careful observation in seven interconnected studies of these professionals diagnosing a simulated case of reading difficulty has shown correlations generally not significantly different from zero, whether from the same clinician diagnosing the same case twice or different clinicians each diagnosing the same case once. This finding holds for professionals selected by their peers as the best; for classroom teachers, reading specialists, and learning disability specialists. After eliminating most counter explanations of low reliability through carefully designed replication studies, the reliabilities are still of borderline significance from zero (e.g., 0.10 - 0.15). c41 Training studies have indicated more optimistic possibilities - in thirty hours of instruction, the correlations may be raised from 0.1 to 0.4 - a value close to physician reliability. Close observation of the training process and its transfer to work settings will hopefully uncover means by which reliability may be raised further. IV Computer Simulation Studies With this backdrop, consider the contribution of computer simulation to this program of research. For this discussion we will ignore the use of case simulation which has been so vital for stimulus control in our experimental design and turn ins tead to clinician simulation theories described earlier. on the All of our simulations of reading specialists have been simulations based on the Inquiry Theory and the Model of Reading and Learning to read described earlier. In this way, the results of each simulation can be used to expand and refine a model that directs our reseach efforts in simulation, training and observation. All studies described here were run on an interpretive procedural language whose primitives were based on the Inquiry Theory. This system is entitled BMIS - the Basic Management Information System. Effectively, a system subroutine was created for each elementary information processing task described by the Inquiry Theory (e.g., hypothesis generation, cue collection, diagnostic determination, etcetera). Each subroutine could be called up by a command in an interpretive language. An initial hypothesis directed program was set up in which the hypotheses generated about a case direct the collection of information about a case and information interpretation, which generated more hypotheses, and so on. On the basis of any decision that was made (accept hypothesis X as part of the diagnosis, reject hypothese Y, etc.) sub procedures might be activated to handle the peculiarities of the particular decision. This system was designed for theory investigation and was not intended to be easy to use or flashy. Furthermore, there were many restrictions on its input to bypass the natural language communication problems. As time permitted, a new system was created to rectify these and other identified shortcomings of BMIS. The new production oriented system with similar types of primitives is entitled MOSES, the Multiple Object Simulated Encounter System. Both systems are available through TELENET on the Wayne State University Amdahl 470~6. The SIMCLIN Modeling Study: The first simulation study was basically a modeling study. Given the framework provided by the Inquiry Theory, memory structures were created by systematic interview with a senior reading clinician. Such things as hypotheses, cues, observations, diagnoses, strategies, etc. were defined. The goal was the creation of a simulation that would closely emulate this specialist's problem solving behavior. Comparisons were drawn between the human specialist and the computer analog as they diagnosed the simulated cases of reading difficulties mentioned earlier. The results indicated that the simulation was a very effective model in terms of all measures used - the number and order of cues collected, the diagnosis and suggested remedial plan, etc. The Pilot SIMCLIN Reliability Study: With the very low human clinician reliability, it became clear that modeling of individual people was a pointless procedure. Instead we directed our efforts to the simulation of behavior of 287 groups of clinicians; i.e., to the simulation of models of diagnosis agreed upon by clinicians. At this point, then, the emphasis turned to the creation of intelligence that would be reliable and valid with respect to group reading diagnosis and still be teachable to unaided human specialists. It was at this point that the development of the Model of Reading and Learning was begun - it would serve to define the content of clinician memory. This study examined the reliability of a computer diagnostic system that was based on the Inquiry Theory and the newly developed MORAL. The simulated clinician (SIMCLIN) was set up and asked to diagnose four simulated cases twice (no SIMCLIN memory of previous runs was allowed but different initial contact settings were used). These diagnoses were compared with respect to reliability with the diagnoses of human clinicians. The results were that the SIMCLIN had a reliability of 0.65 compared to human reliability of 0.10. Further, commonality scores - which indicate how an individual agrees with a criterion group diagnosis - indicated that the SIMCLIN included 80% of the categories agreed upon by the group of human clinicians while the mean for individual human clinicians was 50%. *The Pilot SIMCLIN Validity Study: Finally, a simulation study has been run to get a first measure of the validity of the SIMCLIN's diagnostic decisions when those decisions are directed by the Inquiry Theory and MORAL. Reading case records were taken from Michigan State University's reading clinic for SIMCLIN workup. Records were selected which indicated correct diagnosis and others that indicated poor diagnosis (as measured by the child's response to treatment). The areas of concern were the adequacy of the SIMCLIN as an embodiment of the theories, the reliability of the SIMCLIN diagnosis and the validity of the SIMCLIN diagnosis. It was hoped that the SIMCLIN would agree closely with the clinic's diagnosis for the correctly diagnosed case and not as closely for the poorly diagnosed one. The SIMCLIN did, in fact, behave as dictated by the MORAL - the simulation checked out the critical abilities of reading and the prerequisite factors and past history of those that were problematic. The reliability of the diagnostic decisions was essentially 1. Adherence to the MORAL almost guarentees this. With respect to the SIMCLIN diagnosis on the well and poorly diagnosed cases, the results were equivocal. The reason for this is that data required by the SIMCLIN was not present in the clinic files. Such things as classroom observation of engaged academic time, listening comprehension scores, and change scores over time were not available. In fact, indications are that these types of data are not routinely collected by reading clinicians, although the SIMCLIN considers them significant. The model and its simulation might well demonstrate inadequacies in the state of the art in reading diagnosis. V Conclusion In conclusion, the research paradigm described here has been quite effective. Models and theories direct and focus research designs. These designs - whether observational, training or simulation - reflect back to expand and refine the theories. Substantial data has shown that an individual's decisions may be very unreliable. Training in decision making models and content area theories can improve the reliability. But the key to effective problem solving seems to be the validity of the theories that are used to direct decision making. One effective means for examining the validity of such theories is through computer simulation. The next step will be the completion of a production oriented SIMCLIN that will be used as a preceptor during instruction of student clinicians and a decision aid by reading specialists in schools. The validity of the MORAL SIMCLIN will be checked by following its recommendations and watching the results for real children. The research will continue to be theory oriented. Further information on many aspects of this research program may be obtained by contacting the Institute for Research on Teaching at Michigan State University. References Cl1 c21 c31 c41 Wagner, C.C. and J.F. Vinsonhaler "The Inquiry Theory of Clinical Problem Solving: 1980", The Institute for Research on Teaching, Michigan State University, 1980. Elstein, A.S., L.S. Shulman, S. Sprafka, H. Jason, N. Kagan, L.K. Akkak, M.J. Gordon, M.J. Loupe, and R.D. Jordon. Medicd Problem Solving: An Analvsis of Clinical Reasoning Press, l&8. Cambridge: Harvard University Sherman, G.B., J.F. Vinsonhaler, and A.B. Weinshank "A Model of Reading and Learning", The Institute for Research on Teaching, Michigan State University, 1979. Vinsonhaler, J.F. "The Consistency of Reading Diagnosis", The Institute for Research University, 51:80. Teaching, Michigan State 288	1980	36
30	A WORD-FINDING ALGORITHM WITH A DYNAMIC LEXICAL- SEMANTIC MEMORY FOR PATIENTS WITH ANOMIA USING A SPEECH PROSTHESIS Kenneth Mark Colby, Daniel Christinaz, Santiago Graham, Roger C. Parkison The Neuropsychiatric Institute/UCLA 760 Westwood Plaza Los Angeles, California 90024 ABSTRACT Word-finding problems (anomia) are common in brain-damaged patients suffering from various types of aphasia. An algorithm is described which finds words for patients using a portable micro- processor-based speech prosthesis. The data- structures consist of a lexical-semantic memory which becomes reorganized over time depending on usage. The algorithm finds words based on partial information about them which is input by the user. I WORD RETRIEVAL PROBLEMS We are developing an "intelligent" speech prosthesis (ISP) for people with speech impairments Cll. An ISP consists of a small, portable computer programmed to serve a number of functions and interfaced with a speech synthesizer. ISPs can be operated by pressing keys on a keyboard or by eye- movements using a specially-designed pair of spectacles. How words are stored, organized, and re- trieved from human lexical memories constitutes a lively area of research in current psychology, computational linguistics, neurology, aphasiology and other cognitive sciences [2], [3], [4], [5], C61, 1171, 181. Words in a lexical memory can be associated to other words by means of several relations - synonymy, antonymy, part-whole, spatio-temporal contiguity, etc. [33. It can also be assumed that the process of word-finding for production begins with semantic concepts to which word-signs are connected. Once the word-represen- tation for the concept is found, it is subjected to phonological rules if it is to be spoken, and to graphological rules, if it is to be written. In the final stage of output, articulatory rules governing muscle movements for speech are utilized or rules for hand and finger movements for writing are applied. Impairment in word expression can be due to failures at any stage of this process from concept to utterance. Our interest here is in those instances in which the speaker has the word, or part of the word, or some information about the word, in consciousness but cannot produce the target word. Our efforts have been directed -_------ This research was supported by Grant #MCS78-09900 from the Intelligent Systems Program of the Mathematics and Computer Science Division of the National Science Foundation. towards writing and field-testing a computer program which can find words in a lexical-semantic memory, given some working information about the target word. It has long been known that some aphasic patients, who cannot produce a word, can indicate how many syllables the word contains by finger tapping or squeezing the examiner's hand [9]. Both Barton [lo] and Goodglass et. al. Lll] reported that aphasics know some generic proper- ties of the target word such as its first letter, last letter, etc. Our own experience with patients having anemic problems have confirmed and extended these observations. The word-expression disorders in which we are interested are commonly divided into two groups which are (weakly) correlated with the locations of brain lesions and accompanying signs and symptoms. The first group consists of patients with lesions in the anterior portion of dominant cerebral hemisphere. These patients have many concrete and picturable words in consciousness, and they perform well, although slowly, on naming tasks [12], [13]. The naming disruption is part of a more generalized disturbance which includes laborious articulation and often a disruption of productive syntax. The second group of disorders in this classification scheme involves lesions in the posterior region of the dominant hemisphere. Although these patients often fail to provide the correct name of an object, their substituted response can be related to the target word as described above in the studies of Rinnert and Whitaker [6]. Since the substitute word is so often systematically related to the target word, we thought it might be usable as a clue or pointer in a lexical search for the target word. II A WORD-FINDING ALGORITHM --- The first step involves constructing the data base, a lexicon of words stored on linked lists at various levels. The highest level is a "Topic Area" (TA), representing the idea being talked ahout. Currently we use 15 topic areas. The TA consists of list of words, each connected in a large lexicon of words, falling within that topic area. Each word within a particular topic area is also linked to a list of selected words, as in the 289 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. following example: Assuming that each slot in the pattern is (BODY (ACHE (HEAD HURT PAIN ( ANKLE (FOOT LEG) ) STOMACH) ) (WOUND (BLOOD CUT HURT ) ) ) The organization of the word lists changes dynamically over time according to their usage, as will be described below. In attempting to find a word, the program first asks the user to identify the topic-area from a list presented to him on a display. We start with an initial set of word-lists but change them in accordance with the individual user's environment. A topic-area specifies where the user is linguistically in the discourse, not where he is physically or socio-psychologically. He is then asked about certain properties of the target word, the questions appearing on the display as follows: (1) What is the topic area? (2) What is the first letter of the word? (3) What is the last letter of the word? (4) What letters are in the middle of the word? (5) What word does this word go with? Question (5) attempts to obtain a word associated to the target word however idiosyncratic it might be. It may even resemble the target word in sound. Our starting lexical memory of about 1800 words was based on known high-frequency discourse words, associations from Deese [81, word associa- tion norms cl41 and word-pairs from Rinnert and Whitaker [6] . Word-associations can be highly idiosyncratic to the individual. For example, if one asks 1,000 college students which word they associate to the word table, 691 says chair but the remainder of the responses are distributed over 32 other words including jewels like big, cards, and tennis [151. Hence, with each user, we add his particular associations and words that do not appear in our starting lexicon. This data is collected by a spouse, friend, research assistant or speech pathologist in conversations with the user. After the user has offered clues about the target word, we can represent the responses using a clue pattern of the form: CLUE = TA + L + L + STRING + GWW -t;rh-= TA = Topic area L = Letter STRING = One or more letters GWW = GOESWITH word correct and not null, the program first finds a short list of numbered candidate target words (zero to a maximum of 20 words) meeting the input criteria of the clue pattern. The search will be illustrated by cases. Case (1). Suppose the target word were steak. The clue pattern might be: CLUE = FOOD + S + K + A + meat. The word meat is looked up on the FOOD list and then a search is made on the list of words linked with the word meat which begin with S and end with K and have an A in between. If steak were on the list of words linked to meat, it would be display- ed visually and auditorily as: 1. STEAK When he sees or hears the target the user enters its number (in this case "1.") and the word is then inserted in the utterance of the ISP. (Some- times, on seeing or hearing the target, the user can utter the word himself.) In this first illustration of the program's operations, we have assumed full and correct entries have been made in the clue pattern and that steak was on the list of words which [GOESWITH] meat. But suppose steak is not on the meat list of FOOD. Case (2). If steak were not found on the meat list of FOOD, the "meat" part of the clue is ignored and all words under FOOD beginning with (S) , ending in (K), and with an (A) in between are retrieved. The word steak might appear in this group and if so, the program will add it to the meat list after the user signifies this is his target word. Thus the lexical memory becomes automatically reorganized over time. Case (3). If still no acceptable word is retrieved, the "FOOD" part of the clue is ignored and a search is made on all topic-area lists for (S + K + A) words. The word steak might appear in the ANIMAL topic-area associated with the word cow. After the user indicates this is the target word, steak is added to the meat list under FOOD. With repeated usage, steak becomes promoted to the top of the meat list. Case (4). If the target-word is still not retrieved bv an exhaustive search of all topic- areas, the word does not exist in the lexicon and the program ends. One might consider varying the constraints of (S + K + A) clues and searching further but in our experience this is rarely productive. Time is not a problem for the program since an exhaustive search requires only a few seconds. But a large number of candidate words are retrieved when most of the clues are ignored. And it is too time-consuming for the user to search through the list of candidates looking for the desired word. Case (5). It might be that the user cannot 290 answer completely the questions required by the clue pattern or the entries may be in error. For example, he may not know the first and last letters of steak, but he does know the topic-area is FOOD and the [GOESWITH] word is meat. If steak is on the meat list, the search will succeed. If not, there is no point to displaying all the words under FOOD because the candidate-set is too large. In our experience thus far we have found that at least 2-3 pieces of information are necessary to retrieve a target word. Further experience may indicate that some users can benefit from variations in the clue pattern. For example, with some patients, we have found it expedient to ask them first for the topic-area and the [GOESWITH] word. If this fails, the program then asks the letter questions. With other patients questions regard- ing word-size, number of syllables, and "what do you use it for?" may be helpful. We are currently field testing the program with a variety of anemic patients using keyboard or ocular control to study the value of utilizing different clue patterns.* In the meantime, others in the field of communica- tion disorders may wish to utilize and improve on the word-finding algorithm reported here. Such a program can also be used by speech pathologists as a therapeutic aid. The patient can be given linguistic tasks to practice on using the ISP at home. Repeated exercise and home- practice may facilitate the finding functions. patient's own III SUMMARY We have described a computer program for an intelligent speech prosthesis which can find words in a lexical-semantic memory given some informa- tion about the target words. The program is currently being tested with a variety of anemic patients who have word-finding difficulties. A future report will describe the results of this field-testing. REFERENCES [ll Colby, K.M., Christinaz, D., Graham, S. "A Computer-Driven Personal, Portable and Intelligent Speech Prosthesis." Computers and Biomedical Research 11 (1978) 337-343. [2] Caramazza, A., Berndt, R.S. "Semantic and Syntactic Processes in Aphasia: A Review of the Literature." Psychological Bulletin 85 (1978) 898-918. -------- *We are grateful to Carol Karp and patients of the Speech Pathology Division of Northridge Hospital Foundation, Northridge, California (Pam Schiffmacher, Director) for their collaboration. c31 c41 c51 [61 c71 c81 c91 Cl01 Cl11 Cl21 Cl31 Cl41 Cl51 Miller, G.A., Johnson-Laird, P.N. Language and Perception. Cambridge: Harvard University Press, 1976. Goodglass, Naming and Aphasia." 374. H ., Baker, E. "Semantic Field, Auditory Comprehension in Brain and Language 3 (1976) 359- Zurif, E., Caramazza, A., Myerson, R., Galvin, J. "Semantic Feature Representation for Normal and Aphasic Language. Brain and Language 1 (1974) 167-187. Rinnert, C , Whitaker, H.A. "Semantic Confusions by Aphasic Patients." Cortex, 9 (1973) 56-81. Geschwind, N. "The Varieties of Naming Errors." Cortex 3 (1967) 97-112. Deese, J. The Structure of Associations in Language and Thought. Baltimore: Johns Hopkins Press, 1965. Lichtheim, L. "On Aphasia." Brain 7 (1885) 433-484. Barton, M. "Recall of Generic Proper Words in Aphasic Patients." Cortex 7 73-82. ties of (1971) Goodglass, H., Kaplan, E., Weintraub, S., Ackerman, N. "The "Tip-of-the-Tongue" Phenomenon in Aphasia." Cortex 12 (1976) 145-153. Wepman, J.M., Bock, R.D., Jones, L.V., Van Pelt, D. "Psycholinguistic Study of the Concept of Anomia." Journal of Speech and Hearing Disorders 21 (1956) 468-477. Marshall, J.C., Newcome, F. "Syntactic and Semantic Errors in Paralexia." Neuro- psychologia 4 (1966) 169-176. Postman, L., Keppel, G. (Eds.). Norms of Word Association. New York: Academic Press, 1970. Palermo, D.S., Jenkins, J. Word Association Norms. Minneapolis: University of Minnesota Press, 1964. 291	1980	37
31	TROUBLE-SHOOTING BY PLAUSIBLE INFERENCE * Leonard Friedman Jet Propulsion Laboratory, California Institute of Technology Pasadena, CA 91103 ABSTRACT The PI system has been implemented with the ability to reason in both directions. This is combined with truth maintenance, dependency directed backtracking, and time-varying contexts to permit modelling dynamic situations. Credibility is propagated in a semantic network, and the belief transfer factors can be modified by the system, unlike previous systems for inexact reasoning. I TROUBLE-SHOOTING LOGIC The PI (for Plausible Inference) system enables a user to trouble-shoot physical systems in a very general way. The trouble-shooting process requires that our user first define a physical model which represents what takes place in normal operation with everything functioning correctly. If this physical model is translatable to the representation used by PI, it can be stored in computer memory and used to guide the search for the most likely failure. -In order to make the process clearer, we shall describe a few of the many methods of reasoning employed by human beings in their trouble-shooting. These methods are the ones we can at present imitate in the PI system. Suppose we have a desired goal state defined in our physical model, and this state depends on three conditions being true to attain the goal. If we execute the process and observe that the goal state was not attained, we conclude that at least one of the three conditions on which it depended must have been false, and all are possibly false. If we then perform some test designed to verify whether one of the conditions is actually true and the test shows that it is indeed true, we conclude that at least one of the remaining two untested conditions must be false. If all but one of the remaining conditions has been eliminated from consideration by further testing, we may conclude that the single condition remaining must be the guilty party. The process of elimination just described is the one normally employed by humans, and it is this process we have implemented on the computer. Of course, the three conditions may in turn have conditions on which they depend. In that case the method just described may be applied This paper presents the results of one phase of research performed at the Jet Propulsion Laboratory, California Institute of Technology, sponsored by the National Aeronautics and Space Administration under Contract NAS 7-100. recursively to narrow the fault down further, at least to the granularity of the conditions employed in the representation. This method fails if there are conditions on which the goal state depends for realization and which are not explicitly represented in the model. Nevertheless, the exercise may serve as a valuable guide to help a user to focus attention on specific, possibly false, areas as likely sources of failure. Another difficulty with the method is the fact that either a test does not exist to determine whether a specific sub-goal was reached, or the sub-goal state in question was changed by a later event occurring in the model. In this case it is difficult to verify whether the changed sub-goal state was ever achieved. Only if there were long-lasting side-effects will it be possible to verify. Such difficulties plague human trouble-shooters as well. The present implementation can not reason about such "vanished" events, in a hypothetical mode, from a past context. II IMPLEMENTATION AND THEORY The PI system is part of a larger system called ERIS, named for the Greek goddess of argument. The basic module of the system, described in [ll, performs deduction and modeling in the Propositional Calculus. A planning module has been built by M. Creeger by augmenting the basic deduction module with many special features, including a "causal" connective that supplements the standard logical connectives (AND, OR, etc.). Similarly, the PI module has been built by augmenting Propositional Calculus with extra rules of inference, and another belief besides truth-value associated with each assertion. This additional belief, which we call credibility, is a subjective numerical measure of the confidence in the truth-value, with values between -1 and 1. The basic ERIS module generates a network of nodes linked by connectives as it reads in its knowledge base of assertions; this feature is retained in the other modules. The techniques used in ERIS make it possible to perform deduction without rewriting. Instead, "specialists" for each connective propagate the correct values of the beliefs to the assertions which they link. A theoretical foundation for this approach, aw lying 292 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. both to Propositional Calculus and First Order Predicate Calculus, is given by the method of Analytic Tableaux" [Zl. Because rewriting is totally avoided, inference, planning, model-revision, and dependency-directed backtracking can be performed in a single integrated system. Al(Q( PI 2) I :T; +.5 :F; -.5 Arcs link antecedent/consequent in implication Fig. 1 Transfer Factors in a Semantic Net Plausible Inference introduces two rules of inference besides Modus Ponens and Modus Tollens, called the Method of Confirmation and the Method of Denial. These extra modes permit the propagation of truth values, true or false, even in directions forbidden bY Propositional Calculus. Simultaneously, the associated credibilities are propagated through the net, employing all four modes as appropriate. A calculus of Credibility transfer between arbitrary logical expressions has been worked out to specify exactly the process of Credibility propagation through the net. The calculus is described in [31, and is based on equations employed in MYCIN [41. The basic quantities controlling propagation between antecedents and consequents in implication are transfer factors or DELTAl's, and there are four for each antecedent/consequent pair, one for each mode. (See Fig. 1) Both MYCIN and PROSPECTOR are limited to a single reasoning mode and transfer factor for each implication and use a static transfer factor structure that is specified by human "experts". The PI system, in trouble shooting, recalculates the appropriate transfer factors on the basis of incoming evidence. In addition to a dynamic transfer factor structure, PI also incorporates the use of default values for the transfer factors in trouble-shooting when the users do not have better information. III APPLICATION EXAMPLE Our example of automated trouble-shooting uses a toy case selected from the application domain, the mission control of spacecraft. It is a simplified representation of signal transmission from a space craft to Earth. The desired goal state is "Signal Received" from the SIC. Fig 2 shows a plan to accomplish this which has been generated by the ERIS planner. The user supplies three basic "action CAUSES state" relations: (1) "Ground Antenna Receiving" CAUSES "Signal Received", (2) "Point SIC" CAUSES "Pointed Correctly", and (3) "Transmit Signal" CAUSES "Signal Transmitted". The pre-conditions for the first relation are the end states achieved in the GROUND ANTENNA RECEIVING S/C POINTED CORRECTLY ) -“.“yT) -0.001 TRANSMIT SIGNAL RSS POWER ON -0.001 SIGNAL RECEIVED (3 77 CORRECT DATA MODE SIGNAL TRANSMITTED Fig. 2 Plan for Command Sequence Generation 293 second and third. In effect we are saying, "In order for an action to achieve the goal state certain pre-conditions must be true. In order to make those pre-conditions true certain actions will cause them which may also have pre-conditions to succeed". Thus complex sequences may be built up. The ERIS planner links the basic strategies in the manner shown, using the pre-conditions as hooks. It then collects the actions in a list, and supplies that list as the desired command sequence. The plan generated in this way is a descriptive model of the signal transmission process and constitutes our trouble-shooting knowledge base. The propagation of beliefs that takes place with the CAUSES connective is identical to the belief propagations of an implication as defined in [3], although the timing of belief propagation of CAUSES are different. We define the belief propagation-equivalent implication form of the CAUSES relation as (IMPLIES (AND action precondition1 . . preconditionN) goal-state). At the start, the assumption is made that all states are true. Suppose that the sequence is executed and the ground station fails to receive the signal. Then "Signal Received" is false, and this can be entered into the data base. The effects of this change of belief are propagated through the data base by a modified Modus Tollens, making all the events on which "Signal Received" depends Possibly-False (or PF). If a test is then performed by the human controllers like causing the s/c to roll while transmitting, and a signal is received during the roll, we may conclude that the action "Point space craft" worked, "Signal Transmitted" is true, and "Ground Antenna Receiving" has been verified. Inputting these facts into the data base causes the PI system to do two things: (1) Those preconditions required by "Signal Transmitted" are changed from PF to T by Modus Ponens. (Possibly False to True) (2) "Pointed Correctly" is changed to False, F, rather than PF. In addition, the PI system raises the credibility of failure for those events on which "Pointed Correctly" depends. Their truth value remains Possibly False because there are multiple possibilities. If one of these latter events is shown to be true by testing, the remaining one may be the only possibility left. For example, if the Sun Sensor and Canopus Sensor can be shown to work, and their truth status is input, the system will conclude that the Starmap must be at fault, even though the a priori credibility of such a mistake was extremely low. How the credibilities change at various stages of operation can be described now. At the start there are two possibilities: either a priori credibilities may be entered or default credibilities generated. Figure 2 shows a priori credibilities entered on the branching lines. These are subjective measures of the likelihood of failure of the respective events to which they lead given that the state which depends on them is false. Thus, for "Signal Received", "Pointed Correctly" has an a priori credibility associated with false of -.7, "Signal Transmitted" a value of -.3 and "Point space craft" a value of -.OOl . When we start, the assumption is that every state is true (at the appropriate time) with a credibility of 1.0 (equivalent to certainty). At the next stage, when all we know is that "Signal Received" is false, the a priori credibilities are assigned to all the states. If we used the default mode, credibility would be assigned equally; i. e., for two events, each would get .5, for three .33, for four .25, etc. Whenever an event is eliminated as true, the remaining credibilities are raised by an empirical formula that reflects a reasonable sharing of suspicion, based either on the a priori splits or an equal partition. Thus, in our example, "Starmap Accurate" went from true (cred 1.0) to Possibly False (cred -.007) to Possibly False (cred -.Ol> to False (cred -1.0). This, in a simplified way, describes the operation of PI in trouble-shooting using reasoning by Plausible Inference. Of course, humans employ many other methods in trouble shooting, such as analogy. For example, a person may say "This problem resembles one I encountered in another area. Maybe it has the same cause I deduced then." By such techniques, humans can often vector in on a problem, bypassing step-by-step elimination. We hope to implement some of these techniques eventually. References 111 Thompson A. M., "Logical Support in a Time-Varying Model", in Lecture Notes in Computer __I_-- Science, Springer-Verlag, Berlin, 1980. Fifth Conf. on Automated Deduction, July 8-11, Les Arcs, Savoie, France. [21 Smullyan R. M, First-Order Logic, Springer-Verlag, Berlin, 1968 [3] Friedman L., "Reasoning by Plausible Inference" in Lecture Notes 580. Computer Science Springer-Verlag, Berlin, Proc. Fifth Conf. on Automated Deduction, July 8-11, Les Arcs, Savoie, France. [41 Shortliffe E. H. and Buchanan B. G., "A Model of Inexact Reasoning in Medicine" , Math. Biosci. 23, PP. 351-379, 1975. Also chapt. 4 of Computer-Based Medical Consultations: MYCIN, Elsevier, New York, 1976. 294	1980	38
32	AN APPLICATION OF THE PROSPECTOR SYSTEM TO DOE’S NATIONAL URANIUM RESOURCE EVALUATION* John Gaschnig Artificial Intelligence Center SRI International Menlo Park, CA 94025 Abstract A practical criterion for the success of a knowledge-based problem-solving system is its usefulness as a tool to those workilb in its specialized domain of expertise. Here we describe an application of the Prospector consultation system to the task of estimating the favorability of several test regions for occurrence of uranium deposits. This pilot study was conducted for the National Uranium Resource Estimate program of the U.S. Department of Energy. For credibility, the study was preceded by a performance evaluation of the relevant portion of Prospector’s knowledge base, which showed that Prospector’s conclusions agreed very closely with those of the model designer over a broad range of conditions and levels of detail. We comment on characteristics of the Prospector system that are relevant to the issue of inducing geologists to use the system. 1. Introduction This paper describes an evaluation and an application of a knowledgerbased system, the Prospector consultant for mineral exploration. Prospector is a rule-based judgmental reasoning system that evaluates the mineral potential of a site or region with respect to inference network models of specific classes of ore deposits. Here we describe one such model, for a class of “Western statesIr sandstone uranium deposits, and report the results of extensive quantitative tests measuring how faithfully it captures the reasoning of its designer across a set of specific sites (used as case studies in fine-tuning the model), and with respect to the detailed subconclusions of the model as well as its overall conclusions. Having so validated the performance of this model (called RWSSU), we then describe a pilot study performed in conjunction with the National Uranium Resource Evaluation (NURE) program of the U.S. Department of Energy. The pilot study applied the RwSSU model to evaluate and compare five target regions, using input data provided by DOE and USGS geologists (using the medium of a model-specific questionnaire generated by Prospector). The results of the experiment not only rank the test regions, but also measure the sensitivity of the conctusions to more certain or less certain variations in the input data. One interesting facet of this study is that several geologists provided input data independently about each This research was supported by the U. S. Geological Survey under USGS Contract No. 14-08-0001-l 7227. Any opinions, findings, and conclusion or recommendations expressed in this report are those of the author and do not necessarily reflect the views of the U.S. Geological Survey. test region. Since input data about each region varies among the responding geologists, so do the conclusions; we demonstrate how Prospector is used to identify and resolve the disagreements about input data that are most significantly responsible for differences in the resulting overall conclusions. This paper is a condensation of portions of a larger report [4]. 2. Validation of the Model The practical usefulness of an expert system is limited if those working in its domain of expertise do not or will not use it. Before they will accept and use the system as a working tool, such people (we shall call them the “domain users”) usually expect some evidence that the performance of the system is adequate for their needs (e.g., see [8]). Accordingly, considerable effort has been devoted to evaluating the performance of the Prospector system and of its various models [2, 31. In the present case, we first needed to validate the performance of the uranium model to be used r- the pilot study for the U.S. Department of Energy. The methodology used to evaluate Prospector’s performance is discussed in detail elsewhere [2, 31. For brevity, here we outline a few relevant factors. The Prospector knowledge base contains a distinct inference network model for each of a number of different classes of ore deposits, and a separate performance evaluation is performed for each model. Here we are concerned with one such model, called the regional-scale “Western states” sandstone uranium model (RWSSU), designed by Mr. Ruffin Rackley. Since there exist no objective quantitative measures of the performance of human geologists against which to compare that of Prospector, we instead use a relative comparison of the conclusions of a Prospector model against those of the expert geologist who designed it. To do so, first a number of test regions are chosen, some being exemplars of the model and others having a poor or less good match against the model. For each such case, a questionnaire is completed detailing the observable characteristics that the model requests as inputs for its deliberation. Prospector evaluates each such data set and derives its conclusion for that test case, which is expressed 081 a scale from -5 to 5. As a basis of comparison, we also independently elicit the model designer’s conclusion about each test case, based on the same input data, and expressed on the same -5 to 5 scale. Then we compare Prospector’s predictions against the target values provided by the model designer. Table 1 compares the top-level conclusions of Prospector (using the RWSSU model) against those of the model designer for eight test regions. 295 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Table 1. Comparison of RWSSU model with Designer for Eight Test Cases Test Region Designer’s Prospector Di fference Target Score -------------------------------------------------- Black Hi Ils 3.50 4.33 -0.83 Crooks Gap 4.70 4.26 0.44 Gas Hills 4.90 4.37 0.53 Shirley Basin 4.95 4.13 0.82 Ambros i a Lake 5.00 4.39 0.61 Powder River 4.40 4.40 0.00 Fox Hi I I s 1.50 2.17 -0.67 Oi I Mountain 1.70 3.32 -1.62 ------------------------------------------------- Average: 0.69 Table 1 indicates that the average difference between the Prospector score and the corresponding target value for these eight cases is 0.69, which is 6.9% of the -5 to 5 scale. Besides the overall conclusions reported above, quite detailed information about Prospector’s conclusions was collected for each test case. One feature of the Prospector system is the ability to explain its conclusions at any desired level of detail. In its normal interactive mode, the user can interrogate Prospector’s conclusions by indicating which conclusions or subconclusions he wishes to see more information about. The same sort of information is presented in Table 2 (using the Gas Hills region as an example), in the form of Prospector’s overall evaluation, the major conclusions on which the overall evaluation is based, and the subconclusions that support each major conclusion. For brevity, each section of the RWSSU model represented in Table 2 is identified by its symbolic name, which is indented to show its place in the hierarchy of the model. For comparison, we first elicited from the model designer his target values for each section of the model listed in Table 2; these values are included in Table 2. Table 2. Detailed Comparison of RWSSU Model with Designer for Gas Hills Designer’s Prospector Difference Target Score ------------------------------------------------ RWSSU 4.90 4.37 .53 FTRC 4.80 4.64 .16 TECTON 4.50 4.50 .oo AHR 5.00 4.95 .05 FAVHOST 4.80 5.00 -. 20 S EDTECT 4.80 4.88 -.08 FAVSED 4.90 4.68 .22 FLUVSED 4.90 4.68 .22 KAlti’lESE:) -3.54 -2.03 -1.43 AEOLSED -2.50 -2.10 -.40 FMA 4.95 4.41 .54 RBZONE 5.00 4.60 .40 AIZONE 4.00 4.77 -.77 M I NZONE 5.00 5.00 l oo Average difference = 0.36 (Average of absolute values) The data in Table 2 indicate that Prospector not only reaches essentially the same numerical conclusions as its designer, but does so for similar reasons. This detailed comparison was repeated for each of the eight cases, resulting in 112 distinct comparisons between Prospector’s prediction and designer’s target value (i.e., 8 test regions times 14 sections of the model). The average difference between Prospector’s score and designer’s target value over these 112 cases was 0.70, or 7.0% of our standard lo-point scale.* Gaschnig [4] also reports sensitivity analysis experiments showing the models to be rather stable in their conclusions: for the RWSSU model, a 10% perturbation in the input certainties caused only a 1.2% change in the output certainties. 3. Results of the NURE Pilot Study Having established the credibility of the RWSSU model by the test results just discussed, we then undertook an evaluation of five test regions selected by the Department of Energy. For this purpose USGS and DOE geologists completed questionnaires for this model. As a sensitivity test, several geologists independently completed questionnaires for each test region. For comparison, the model designer, R. Rackley, also completed questionnaires for the five test regions. The overall results are reported in Table 3, in which the abbreviations M.H., P.B., MO., N.G., and W.R. denote the names of the test regions, namely Monument Hill, Pumpkin Buttes, Moorcroft, Northwest Gillette, and White River, respectively. Table 3. Overall Conclusions for Five Test Regions Geologist A B C D USGS Rackley Range team data ------------------------------------------------ M.H. 4.17 3.32 3.97 4.40 1.08 P.B. 4.20 3.30 4.19 4.40 1.10 MO. 3.92 3.88 4.00 0.12 N.G. 3.64 0.10 3.42 3.54 W.R. 0.13 0.01 0.12 The results in Table 3 indicate that the Monument Hill, Pumpkin Buttes, and Moorcroft regions are very favorable, and about equally favorable, for occurrence of “Western States” sandstone uranium deposits. Northwest Gillette is scored as moderately favorable, whereas White River is neutral (balanced positive and negative indicators). Note that each respondent has had different exposure to the target regions, in terms of both first-hand, on-site experience and familiarity with field data reported in the literature. These differences in experience are reflected in their answers on the questionnaires. Since different inputs yield different conclusions, one would expect a spread in the certainties about each region, reflecting the differences in input data provided by the various geologists. Inspection of Table 3 reveals, however, that the scores derived from different geologists’ input data about the same region agree rather closely for each region except Northwest Gillette (see the column labeled IrRange”). These generally close agreements reflect the capability of Prospector models to synthesize many diverse factors, mechanically ascertaining general commonalities without being unduly distracted by occasional disparities. In cases such as Northwest Gillette in which a large difference in conclusions occurs, it is easy to trace the source of the disagreement by comparing the individual conclusions for different sections of the model (representing different geological subconclusions), as in Table 4. Table 4. Comparison of Detailed Conclusions About Northwest Gi I lette Geologist C D Rackley Avg. data ----------------------------------------- RWSSU .lO 3.66 3.42 3.56 FTRC 4.67 3.80 4.63 4.37 TECTCN 4.90 4.50 4.50 4.63 AHR 4.95 1.03 4.94 3.64 FAVHQST 5.00 5.00 5.00 5.00 S EDTECT 4.98 4.33 4.78 4.69 FAVSED .04 3.92 4.79 2.92 FLUVSED .04 3.92 4.79 2.92 MAR I NESED -4.60 3.34 .02 -.41 AEOLSED -4.99 -2.10 -3.23 -3.44 FMA .27 2.45 1.33 2.18 RBZQNE 4.10 4.83 4.73 4.55 AIZQNE -3.29 2.40 0.00 -0.30 M I NZQNE .41 2.82 2.59 1.94 Inspection of Table 4 reveals that the conclusions agree fairly closely for the FTRC section of the model, and less closely for the FAVSED and FMA sections. Tracing the differences deeper, one sees that of the three factors on which FMA depends, there is fairly good agreement about RBZONE, but larger differences in the cases of the AIZONE and MINZONE sections. In some cases, such a detailed analysis can isolate the source of overall disagreement to a few key questions about rhich the respondents disagreed. These can then be resolved by the respondents without the need to be concerned with other disagreements in their questionnaire inputs that did not significantly affect the overall conclusions. Prospector has also been applied to several other practical tasks. One evaluated several regions on the Alaskan Peninsula for uranium potential [l 1, as one of the bases for deciding their ultimate disposition (e.g., wilderness status versus commercial exploitation). Another app!iccztian das concc’rnad with measuring quantitatively the economic value of a geological map, resulting in statistically significant results [ 71. 4. Discussion We have measured Prospector’s expertise explicitly and presented a practical application to a national project, demonstrating in particular how the Prospector approach deals effectively with the variabilities and uncertainties inherent in the task of resource assessment. This work illustrates that expert systems intenderr for actual practical use must accommodate the special characteristics of the domain of expertise. In the case of economic geology, it is not rare for field geologists to disagree to some extent about their observations at a given site. Accordingly, the use of various sorts of sensitivity analysis is stressed in Prospector to bound the impact of such disagreements and to isolate their sources. In so doing, we provide geologists with new quantitative techniques by which to address an important issue, thus adding to the attractiveness of Prospector as a working tool. Other domains of expertise will have their own peculiarities, which nust be accommodated by designers of expert systems for those domains. A more mundane, but nevertheless important, example concerns the use of a questionnaire as a medium for obtaining input data to Prospector from geologists. Most geologists have little or no experience with computers; furthermore, access to a central computer from a remote site may be problematic in practice. On the other hand, geologists seem to be quite comfortable with questionnaires. Our point is simply that issues ancillary to Al usually have to be addressed to ensure the practical success of knowledge-based Al systems. References 1. 2. 3. 4. 5. 6. 7. 8. Cox, D. P., D. E. Detra, and R. L. Detterman, “Mineral Resources of the Chignik and Sutwick Island Quadrangles, Alaska,” U.S. Geological Survey Map MF-1053K, 1980 in press. Duda, R.O., P.E. Hart, P. Barrett, J. Gaschnig, K. t\onolige, R. Reboh, and J. Slocum, “Development of the Prospector Consultation System for Mineral Exploration,” Final Report, SRI Projects 5821 and 6415, Artificial Intelligence Center, SRI International, Menlo Park, California, October 1978. Gaschnig, J. G., “Preliminary Performance Analysis of the Prospector Consultant System for Mineral Exploration,” Proc. Sixth International Joint Conference on Artificial Intelligence, Tokyo, August 1979. Gaschnig, J. G., “Development of Uranium Exploration Models for the Prospector Consultant System,” SRI Project 7856, Artificial Intelligence Center, SRI International, Menlo Park, California, March 1980. National Uranium Resource Evaluation, Interim Report, U.S. Department of Energy, Report GJO-111(79), Grand Junction, Colorado, June 1979. Roach, C. H., “Overview of NURE Progress Fiscal Year 1979,” Preprint of Proceedings of the Uranium Industry Seminar, U. S. Department of Energy, Grand Junction, Colorado, October 16-17, 1979. Shapiro, C., and W. Watson, “An Interim Report on the Value of Geologic Maps,” Preliminary Draft Report, Director’s Office, U.S. Geological Survey, Reston, Virginia, 1979. Yu, V.L., et al., “Evaluating the Performance of a Computer-Based Consultant,” Heuristic Programming Project Memo H PP-78-17, Dept. of Computer Science, Stanford University, September 1978. 297	1980	39
33	INCREMENTAL, INFORMAL PROGRAM ACQUISITION’ Brian P. McCune Advanced Information & Decision Systems 201 San Antonio Circle, Suite 286 Mountain View, California 94040 AbJrract. Program acquisition is the transformation of a program specification into an executable, but not necessarily efficient, program that meets the given specification. This paper presents a solution to one aspect of the program acquisition problem, the incremental construction of program models from informal descriptions [II, in the form of a framework that includes (1) a formal language for expressing program fragments that contain informalities, (2) a control structure for the incremental recognition and assimilation of such fragments, and (3) a knowledge base of rules for acquiring programs specified with informalities. 1. Introduction The paper describes a LISP based computer system called the Program Model Builder (abbreviated “PMB”), which receives informal program fragments incrementally and assembles them into a very high level program model that is complete, semantically consistent, unambiguous, and executable. The program specification comes in the form of partial program fragments that arrive in any order and may exhibit such informalities as inconsistencies and ambiguous references. The program fragment language used for specifications is a superset of the language in which program models are built. This program modelling language is a very high level programming language for symbolic processing that deals with such information structures as sets and mappings. 2. The Problem The two key problems faced by PMB come from processing fragments that specify programs incrementally and informally. The notion of incremental program specification means that the fragments specifying a program may be received in an arbitrary order and may contain an arbitrarily small amount of new information. The user thus has the most flexibility to provide new knowledge about any part of the program at any time. For example, a single fragment conveying a small number of pieces of information is the statement “A is a collection.” This identifies an information structure called A and defines it as a collection of objects. However, the fragment says nothing about the type of objects, their number, etc. These details are provided in program fragments occurring before or after this one. Informality means that fragments may be incomplete, semantically inconsistent, or ambiguous; may use generic operators; and may provide more than one equivalent way of expressing a program part. An incomplete program model part may be completed either by use of a default value, by inference by PMB, or frog later fragments from the user. ’ This paper describes research done at the Stanford Artificial Intelligence Laboratory and Systems Control, Inc., and supported by DARPA under Contracts MDA903-76-C-0206 and NOOOl4-79 C-0127, monitored by ONR. Fellowship support was provided by NSF, IBM, and the De Karman Trust. The views and conclusions contained in this paper are those of the author. Program model consistency is monitored at all times. PMB tries to resolve inconsistencies first; otherwise, it reports them to the user. For example, the membership test fragment x E A requires that either A have elements of the same type as x (whenever the types of R and x finally become known) or their types inferred to be the same. Because a fragment may possess ambiguities, its interpretation depends upon the model context. So PMB specializes a generic operator into the appropriate primitive operation, based upon the information structure used. For example, part-oflx,A) (a Boolean operation that checks if information structure x is somehow contained within A) becomes x E R, if A is a collection with elements of the same type as X, and an k.componenr if A is a plex (record structure). PMB is capable of canonization, the transformation of equivalent information and procedural structures into concise, high level, canonical forms. This allows subsequent automatic coding the greatest freedom in choosing implementations. Interesting patterns are detected by specific rules set up to watch for them. For example, expressions that are quantified over elements of a set are canonized to the corresponding expression in set notation. 3. Control Structure The model building problem is to acquire knowledge in the form of a program model. The control structure of PMB is based upon the “recognition” paradigm 121, in which a system watches for new information, recognizes the information based upon knowledge of the domain and the current situation, and then integrates the new knowledge into its knowledge base. PMB has one key feature: subgoals may be dealt with in an order chosen by the user, rather than dictated by the system. Subgoals are satisfied either externally or internally to PMB. The two cases are handled by the two kinds of data driven antecedent rules, response rules and demons, which are triggered respectively by the input of new fragments or changes in the partial model. When new information arrives in fragments, appropriate response rules are triggered to process the information, update the model being built, and perhaps create more subgoals and associated response rules. Each time a subgoal is generated, an associated “question” asking for new fragments containing a solution to the subgoal is sent out. This process continues until no further information is required to complete the model. To process subgoals that are completely internal to PMB, demon rules are created that delay execution until their prerequisite information in the model has been filled in by response rules or perhaps other demons. 4. Knowledge Base PMB has a knowledge base of rules for handling constructs of the program modelling language. processing informalities in fragments, monitoring consistency of the model, and doing limited forms of program canonization. Rules about the modelling language include facts about five different information structures, six control structures, and approximately twenty primitive operations. The 71 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. control structures are ones that are common to mose high level languages. The modelling language’s real power comes from its very high level operators for information structures such as sets, lists, mappings, records, and alternatives of these. Below are English paraphrases of three rules that exemplify the major types of rules used in PMB. Rule I is a response rule for processing a new loop. Rule 2 is a demon that checks that the arguments of an i~-&~et operation are consistent. Rule 3 is a canonization demon that transforms a case into a test when appropriate. [II A loop consists of an optional initialization, required body, and required pairs of exit tests and exit blocks. Each exir test must be a Boolean expression occurring within the body. [21 Require that the two arguments of an is-J&Jet operation both return collections of the same prototypic element. [31 If the statement is a case, the case has two condition/action pairs, and the first condition is the negation of the second condition, then change the case into a test. Both response rules and simple demons are procedural. Compound demons (i.e., those whose antecedents test more than one object in the model) use declarative antecedent patterns that are expanded automatically into procedural form by a rule “compiler”. 5. Example of PMB in Operation The model building excerpt below displays (1) growth of the program model tree in a fashion that is generally top down, but data driven, and (2) completion and monitoring of parts of the model by demons. Note that this excerpt does justice neither to the concept of arbitrary order 6f fragments nor the types of programming knowledge in PMB. The trace discusses three program fragments generated from an English dialog. Each fragment is followed by a description of how it was processed by PMB, a snapshot of the partial model at that point, and a list of the outstanding demons. A detailed trace for the first fragment shows PMB focusing on individual slots of a fragment, creating model templates, and creating subgoals. The other fragments emphasize the creation and triggering of demons. Names preceding colons are unique template names that allow fragments to refer to different parts of the model. Missing parts of the partial model are denoted by “???“. Newly added or changed lines are denoted by the character ‘I” at the right margin. [The excerpt starts after the first fragment has already caused the partial program model shown below to be created. It only contains the names of the model, CLASSIFY, and the main algorithm, “algorithm-body”. No demons exist.] Current program model: program classify; algorithm-body: ??? Current demons act ive: None [The second fragment describes the top level algoriehm as a control structure having type composite and two “Input-concept” steps called and “classify-loop”. This fragment might have arisen from a sentence from the user such as “The algorithm first inputs the concept and then classifies it.“] Inputting fragment: algorithm-body: begin input-concept classify-loop end CA composite is a compound statement with an optional partial ordering on the execution of its subparts. The response rule that processes the composite creates the following two subgoals, along with response rules to handle them (not shown).] Processing ALGORITHM-BOOY.TYPE = COMPOSITE Creating subgoal: ALGORITHM-BOOY.SUBPARTS - 711 Creating subgoal: ALGORITHM-BOOY.OROERINGS = 711 Done processing ALGORITHM-BOOY,TYPE = COMPOSITE [Within the same fragment the two subparts are defined as operational units with unique names, but of unknown types. An operational unit can be any control structure, primitive operation, or procedure call. Two new templates are created and their types are requested.1 Process i ng ALGORI THM-BODY. SUBPARTS = (INPUT-CONCEPT CLASSIFY-LOOP) Creating template INPUT-CONCEPT with value I NPUT-CONCEPT. CLASS - OPERAT I ONAL-UNI T Creating subgoal: INPUT-CONCEPT.TYPE - 131 Creating template CLASSIFY-LOOP with value CLASSIFY-LOOP.CLASS = OPERATIONAL-UNIT Creating subgoal: CLASSIFY-LOOP. TYPE = ??? Done processing ALGORITHM-BOOY.SUBPARTS = (INPUT-CONCEPT CLASSIFY-LOOP) [At this point, the of the composite.] model IS missing the definitions of the two parts Current program model: program classify; begin Input-concept: ???; classify-loop: ??? end Current demons active: None [The third fragment, which defines “input-concept” to be an input primitive operation, is omitted. Information structures from this fragment are not shown in the models below. The fourth fragment defines the second step of the composite. This fragment might have come from “The classificaeion step is a loop with a single exit condition.“] Inputting fragment: classify-loop: until exit (exit-condition) repeat loop-body finally exit: endloop [This fragment defines a loop that repeats “loop-body” (as yet undefined) until a Boolean expression called “exit-condition” is true. At such time, the loop is exited to the empty exit block, called “exit”, which is associated with “exit-condition”. Since PMB doesn’t know precisely where the test of aexit-condition” will be located, it is shown separately from the main algorithm below. The response rule that processes the loop needs to guarantee that “exit-condition” is contained within the body of the loop. Since 72 this can’t be determined until the location of “exit-condition” is defined in a fragment, the response rule attaches Demon 1 to the template for “exit-condition” to await this event. Similarly, Demon 2 is created to await the location of “exit-condition” and then put it inside a test with an asserl-exit-connlilon as its true branch. This will cause the loop to be exited when the exit condition becomes true.1 Current program model8 program classlf3r; begirl concept c input(concept#otot~pe, user, concept-prompt); until exit repeat I loop-body: ??? finally I exit: endloop I end exit-condition: ??? I Current demons active: Demon 1: awaiting control structure containing “ex i t-cond i t i on” I Demon 2: awaiting control structure containing “ex i t-cond i t i on” I [The fifth fragment defines the body of the loop, thus triggering the two demons set up previously. A possible source of this fragment is ‘The loop first inputs a scene, tests whether the datum that was input is really the signal to exit the loop, classifies the scene, and then outpu&this classification to the user.“] Inputting fragment: loop-body: begin loop-input; exit-condition; classification; output-classification end [“Loop-body” is a composite with four named steps. PMB now knows where “exit-condition” occurs and that it must return a Boolean value. Demon I is awakened to find that “exit-condition” is located inside the composite “loop-body”. Since this isn’t a loop, Demon 1 continues up the tree of nested control constructs. It immediately finds that the parent of “loop-body” is the desired loop, and succeeds. Demon 2 is also awakened. Since it now knows “exit-condition” and its parent, Demon 2 can create a new template between them. The demon creates a test with “exit-condition” as its predicate and an assert-exit-condition that will leave the loop as its true action.] Current program mode I : program classify; begin concept t input(concept.$ototype, user, concept-prompt); until exit repeat begin loop-input: ???; I if exit-condition: ??? then assert-exit-conditio(exit); I classification: ??!; output-classification: ??? end finally exit: endloop end Current demons act i ve: None I [At the end of the excerpt, five of 32 fragments have been processed.] 6. Role of PMB in a Program Synthesis System PMB was designed to operate as part of a more complete program synthesis system with two distinct phases: acquisition and automatic coding. In such a system the program model would serve as the interface between the two phases. Automatic coding is the process of transforming a model into an efficient program without human intervention. The model is acquired during the acquisition phase; the model is coded only when it is complete and consistent. PMB may work within a robust acquisition environment. In such an environment, program fragments may come from many other knowledge sources, such as those expert in traces and examples, natural language, and specific programming domains. However, the operation of PMB is not predicated on the existence of other modules: all fragments to PMB could be produced by a straightforward deterministic parser for a surface language such as the one used to express fragments. 7. Conclusion PMB has been used both as a module of the PSI program synthesis system E31 and independently. Models built as part of PSI have been acquired via natural language dialogs and execution traces and have been automatically coded into LISP by other PSI modules. PMB has successfully built a number of moderately complex programs for symbolic computation. The most important topics for future work in this area include (1) extending and revising the knowledge base, (2) providing an efficient mechanism for testing alternate hypotheses and allowing program modification, and (3) providing a general mechanism for specifying where in the program model a program fragment is to go. The last problem has resulted in a proposed program reference language I1 3. 8. References [I J Brian P. McCune, Building Program Models Incrementally from informal Descriptions, Ph.D. thesis, AIM-333, STAN-CS-79 772, AI Lab., CS Dept., Stanford Univ., Stanford, CA, Oct. 1979. [2J Daniel G. Bobrow and Terry Winograd, “An Overview of KRL, a Knowledge Representation Language”, Cognifive SCienCe, Vol. 1, No. 1, Jan. 1977, pp. 3-46. [33 Cordell Green, Richard P. Gabriel, Elaine Kant, Beverly I. Kedzierski, Brian P. McCune, Jorge V. Phillips, Steve T. Tappel, and Stephen J. Westfold, “Results in Knowledge Based Program Synthesis”, 1 JCAI-79: Proceedings of the Sixth International Joint Conference on Artijciai Intelligence, Vol. 1, CS Dept., Stanford Univ., Stanford, CA, Aug. 1979, pp. 342-344. 73	1980	4
34	SOME REQUIREMENTS FOR A COMPUTER-BASED LEGAL CONSULTANT L. Thorne McCarty Faculty of Law, SUNY at Buffalo Laboratory for Computer Science Research, Rutgers Although the literature on computer-based consultation systems has often suggested the possibility of building an expert system in the field of law (see, e.g., [2]) it is only recently that several AI researchers have begun to explore this possibility seriously. Recent projects include: the development of a computational theory of legal reasoning, using corporate tax law as an experimental problem domain C61 C73 C81; the development of a language for expressing legal rules within a data-base management environment [Y]; the design of an information retrieval system based on a computer model of legal knowledge [ 31; and the design of an artificial intelligence system to analyze simple tort cases [lo]. This paper attempts to identify the principal obstacles to the development of a legal consultation system, given the current state of artificial intelligence research, and argues that there are only certain areas of the law which are amenable to such treatment at the present time. The paper then suggests several criteria for selecting the most promising areas of application, and indicates the kinds of results that might be expected, using our current work on the TAXMAN project [q] as an example. I. Poter&i& &nlications. One can imagine numerous applications of artificial intelligence techniques, in several diverse areas of law, but most of these would fall into one of the following categories: (I.) m-al Svstema. There are a number of systems in operation today which maintain data bases of statutes and decided cases, in full text, and which are capable of searching these texts for combinations of key words, using standard information retrieval techniques. (For a comparative survey of the principal domestic systems, LEXIS and WESTLAW, see cf21.1 These retrieval systems have turned out to be useful for certain kinds of legal research tasks, but only when used in conjunction with the traditional manual digests and indices, all of which are organized according to a rigid conceptual * This research has been funded by the National Science Foundation through Grant SOC-78- 1 1408 and Grant MCS-79-21471 (1979-81). classification of the law (usually: the West "key number systemV1). With the use of artificial intelligence techniques, however, the retrieval systems could be augmented to provide a form of automated conceptual searching as well, and without the rigidities of the manual indices. For a discussion of these possibilities, see [6] and c31. (2.) I&g,& Analvsis and Planni. Svstems. A step more sophisticated than a retrieval system, a legal analysis and planning system would actually analyze a set of facts, or propose a sequence of transactions, in accordance with the applicable body of legal rules. This is the kind of system that most often comes to mind when one speculates about computer-based legal consultation, for it is the system most similar to the successful systems in chemical and medical domains: a lawyer, engaged in a dialogue with a computer, would describe the facts of his case, and the computer would suggest an analysis or a possible course of action. In fact, there are systems of this sort under development today, using techniques much less powerful than those available to the artificial intelligence community, and they seem close to commercial application: see, e.g., [I31 . The advantages of artificial intelligence techniques for these applications have been discussed by [ 61. (3.) Integrated u Information Svstems. Instead of looking only at the tasks of the private attorney, we could focus our attention more broadly on the legal system as a whole. One of the tasks of the legal system is to manage information and to make decisions about the rights and obligations of various individual actors, and there seems to be no reason, in principle, why some of this information and some of these decisions could not be represented entirely within a computer system. For a current example, using conventional programming technology, consider the computerized title registration systems which are now being used to manage real estate transactions (see, e.g., c511. With the availability of artificial intelligence techniques, a large number of additional applications come to mind: financial transactions, securities registration, corporate taxation, etc. At present it appears that these possibilities are being taken more seriously by European lawyers than by American lawyers (see, e.g., [II and [Ill). From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. If we consider the potential role of artificial intelligence techniques in all of these applications, a basic paradigm emerges. A computer-based legal consul tat ion system must represent the 11facts11 of a case at a comfortable level of abstraction, and it must represent the I1 1 awl1 in the chosen area of application. The “lawl’ would consist of a system of “concepts” and ltrulesl’ with the following characteristics : (a. 1 they would be relatively abstract, that is, they would subsume large classes of lower-level factual descriptions; and (b. ) they would have normative implications, that is, they would specify which actions were permitted and which actions were obligatory in a given situation. Legal anal ysi 3, in its simplest form, would then be a process of applying the lllawll to the “facts.” . Put this way, the paradigm seems to be an ideal candidate for an artificial intelligence approach: the “facts11 would be represented in a lower-level semantic network, perhaps ; the ltlawll would be represented in a higher-level semantic description ; and the process of legal analysis would be represented by a pattern-matching routine. The difficult problems with this paradigm, however, are the representation problems. In the existing knowledge-based systems, in other domains, the representation of the llfactsll and the I1 1 awl1 has been relatively straight forward. In DENDRAL, for example, the ground-level description of all possible chemical strut tures could be represented in a simple graphical notation, and the rules for splitting these structures in a mass spectrograph coul d be -represented as simple operations on the links of the graphs. In MYCIN, the basic facts of a case could be represented as a set of features listing the presence or absence of certain symptoms and the results of certain laboratory tests, and the diagnostic rules could then be represented as a probabilistic judgment that a given symptom or test result implied a certain disease. By contrast, the facts of a legal case typically involve al 1 the complexities of daily life : human actions, beliefs, intentions, motivations, etc. , in a world of ordinary objects like houses and automobil es, and complex institutions like businesses and courts. Even if the facts of a particular case could be represented in a computer system, the rules themselves would often be problematical. Some rules, usually those embodied in statutes, have a precise logical strut ture , and this makes them amenable to the existing artificial intelligence techniques. But it is a commonplace among lawyers that the most important legal rules do not have this form at all: instead they are said to have an ltopen texture”; their boundaries are not fixed, but are llconstruc ted” and ltmodifiedll as they are applied to particular fat tual situations. A sophisticated legal consul tation system would not be able to ignore these complexities, but would have to address them directly.- II. Possible mroaches. Since the representation problems for a legal consul tation system are so difficult, it is tempting to start with the “simplestl’ possible 1 egal issue 3, such as the subject matter of the first-year law school courses. We might therefore be tempted to investigate assault and battery cases from the first-year torts course [ lOI, or offer and acceptance cases from the first-year contracts tour se . But these cases are ltsimplelV for law students primarily because they draw upon ordinary human experience, and this is precisely what makes them so difficult for an artificial intelligence system. To understand tort cases, we must understand all the ways in which human beings can be injured, intentionally and unintentionally, mental 1 y and physically , with and without justification. To understand contract cases, we must understand the expectations of real people in concrete business situations, and the ambiguities of human language in expressing particular contractual intentions. If we abstract away these details, we will miss entirely the central features of legal reasoning, and our consultation systems will tend to produce only the more trivial results. Paradoxically, the cases that are most. tractable for an artificial intelligence system are those cases, usually involving commercial and corporate matters, which a lawyer finds most complex . There is a simple reason why this is so. A mature 1 egal system in an industrialized democracy i a composed of many levels of legal abstractions: the law initially defines llrightsll in terms 0 f concrete objects and ordinary human actions , but these rights are then treated as llobjects I1 themselves, and made subject to further human llactionsll ; by repeating this process of reification many times, a complex body of commercial law can be developed. Because of their technical complexity, the legal rules at the top levels of this conceptual hierarchy are difficult for most lawyers to comprehend, but this would be no obstacle for an artificial intelligence system. The commercial abstractions, in fact, are artificial and formal systems themselves, drained of much of the content of the ordinary world; and because of the commercial pressures for precision and uniformity, they are, by legal standards, well structured. A reasonable strategy for developing a computer-based legal consul tation system, then, would be to start here. This is the strategy we have followed in the TAXMAN project C61 [7] . The TAXMAN system operates in the field of corporate tax law, which is very near the apex of the hierarchy of commercial abstractions. The basic 11facts11 of a corporate tax case can be captured in a relatively straightforward representation: corporations issue securities, transfer property, distribute dividends, etc. Below this level there is an expanded representation of the meaning of a security interest in terms of its component rights and obligations: the owners of the shares of a common stock, for example, have certain rights to 299 the "earnings", the "assets", and the llcontrol" of the corporation. Above this level there is the lllawll : the statutory rules which classify transactions as taxable or nontaxable, ordinary income or capital gains, dividend distributions or stock redemptions, etc. Although these rules are certainly complex, the underlying representations are manageable, and we have concluded from our earlier work that the construction of an expert consultation system in this area of the law is a feasible proposition C61 . In our current work [7] we are taking this model one step further, in an attempt to account for the "open texture" of legal concepts and the process by which a legal concept is "constructed" and "modified" during the course of an argument over a contested case. In many areas of the law this would be an impossible task: the complexity of the representation would be overwhelming, and the structural and dynamic properties of the concepts would be obscured. But in the world of corporate abstractions the task appears to be feasible. Looking at a series of corporate tax cases before the United States Supreme Court in the 1920's and 19301s, we have been able to construct a model of the concept of "taxable income" as it appeared to the lawyers and judges at the time (see [9]) . Although the concept is sometimes represented as a "logicalfl pattern which can be "matched" to a lower-level semantic network (we call this a -ical temblate structure), the more important representation for the process of legal analysis consists of a prototvpe structure and a sequence of deformations of the prototype. We are currently involved in an implementation of these ideas, and we will describe them in detail in a future paper (for an initial description of the implementation, see [81). We believe that the "prototype-plus-deformation" structure is an essential component of a system of legal rules, and that it ought to play an important role in a sophisticated legal consultation system. III. Prosbects. This paper has emphasized the difficulties in constructing a computer-based legal consultation system, but it has also suggested some feasible approaches. The main difficulty is the representation problem: the factual situations in a legal problem domain involve complex human actions, and the most important legal rules tend to contain the most amorphous and malleable legal concepts. By selecting legal problems from the commercial and corporate areas, however, we can construct a representation of the legally relevant facts without having to model the entire human world, and we can begin to develop the necessary structures for the representation of the higher- level legal rules. We have had some success with this strategy in the TAXMAN project, and we believe it can be applied elsewhere as well. REFERENCES T: 11 Bing, J., and Harvold, T., &Q&- Decisions m Information Svstems (Universitetsforlaget, Oslo, 1977). [2] Buchanan, B.G., and Headrick, T.E., "Some Speculation About Artificial Intelligence and Legal Reasoning," 23 Stanford ti Review 40- ( 1970). [3] Hafner, C., "An Information Retrieval System Based on a Computer Model of Legal Knowledge," Ph.D. Dissertation, University of Michigan (1978). [4] Jones, S., Mason, P., and Stamper, R., "LEGOL 2.0: A Relational Specification Language for Complex Rules, )1 4 Information Svstems 293-305 ( 1979). [5] Maggs, P.B., "Automating the Land Title System," 22 American Universitv Law Review 369-91 (1973). C63 McCarty, L.T., flReflections on TAXMAN: An Experiment in Artificial Intelligence and Legal Reasoning,11 90 Harvard & Review 837-93 (1977). [7] McCarty, L.T., "The TAXMAN Project: Towards a Cognitive Theory of Legal Argument," in B. Niblett, ed., Computer &g& & Legal &gl&@Jg (Cambridge University Press, forthcoming 1980). 183 McCarty, L.T., and Sridharan, N.S., "The Representation of an Evolving System of Legal Concepts: I. Logical Templates," in Proceedings. Third National Conference of the Canadian Societv for Computational Studies of A- Intelligence, Victoria, British Columbia, May 14-16, 1980. [9] McCarty, L-T., Sridharan, N-S., and Sangster, B.C., "The Implementation of TAXMAN II: An Experiment in Artificial Intelligence and Legal Reasoning," Report LRP-TR-2, Laboratory for Computer Science Research, Rutgers University (1979). [ IO1 Meldman, J.A., "A Preliminary Study in Computer-Aided Legal Analysis," Ph.D. Dissertation, Massachusetts Institute of Technology, Technical Report No. MAC-TR-157 (November, 1975). [ 111 Seipel, P., Computing & (LiberForlag, Stockholm, 1977). [I21 Sprowl, J.A., A Manual for Commuter-Assisted Legal Research (American Bar Foundation, 1976). El31 Sprowl, J.A. "Automating the Legal Reasoning Process: A Computer that Uses Regulations and Statutes to Draft Legal Documents," 1979 American & Foundation Research Journal l-81. 300	1980	40
35	A FRAME-BASEDPRODUCTION SYSTEM ARCHITECTURE David E. Smith and Jan E. Clayton Heuristic Programming Project* Department of Computer Science Stanford University ABSTRACT We propose a flexible frame-structured representation and agenda-based control mechanism for the construction of production-type systems. Advantages of this architecture include uniformity, control freedom, and extensibility. We also describe an experimental system, named WHJXZE, that uses this formalism. The success of MYCIN-like production systems 141 [7] [9] has demonstrated that a variety of types of expertise can be successfully captured in rules. In some cases, however, rules alone are inadequate necessitating the USC of auxiliary representations (e.g. property lists for paramzters in MYCIN). Other limitations result from the use of goal-directed control. In this paper we outline a flexible schemata for constructing high performance production-like systems. The architecture consists of two components: 1. An extensible representation (utilizing a frame-structured language) which captures production rule knowledge. 2. An agenda-based control mechanism allowing considerable freedom in tailoring control flow. We have used this architecture in the development of a system named WHEEZE, which performs medical pulmonary function diagnosis based on clinical test results. This syst%m is based on two earlier efforts, PUFF [7], an EMYCIN-based production rule system Ill], and CENTAUR [l] 121, a system constructed of both rules and prototypes. AN ~LTERNATXVE RI~PRESENTATION FOH PRODUCTIONS Figure 1 shows how a typical PUFF rule would be transformed into our representation. Each assertion is represented as a frame in the knowledge. base, with anteccdcnt sub-assertions appearing in its h4auzjdarion slot. The number associated with each manifestation is its corresponding importance. Similarly, the certainty factor and findings from the rule are given separate slots in the assertion. Assertions appearing in the SuggesliveOf and ComplemenlaryTo . . ..-........... *This research was supported in part by ONR contract N00014- 79C-0302 and by .NIH Biotechnology Resource Grant RR-00785. 54 slots are those worth investigating cornfirmed or denied respectively assertions are suggesfivilies). Implicit in the production rule if the original assertion is (numbers following these representation is a fimction which indicates how to compute the “belief’ of the consequent assertions given belief in the antecedent assertion. Unfortunately, evaluation of the antecedent assertion involves modal logic (since greater shading is required than simple binary values for belief and disbelief). Therefore, a “HowToDetermineBelief’ slot is associated with each assertion indicating how its belief is to be computed. If: 1) The severity of Obstructive Airways Disease of the patient is less than or greater to mild, al’;1 2) The number of pack-years smoked is greater than 0, and 3) The number of years ago that the patient quit smoking is 0 Then: It is definite (1000) that the following is one of the conclusion statements about this interpretation: Discontinuation of smoking should help re1iev.e the symptoms. OADwithSmoking: Manifestation SuggestiveOf ComplementaryTo Certainty Findings ((OAD-Present 10) (PatientHasSmoked 10) (PatientSillSmoking 10)) ((SmokingExaccrbatedOAD 5) (SmokingInduccdOAD 5)) ((OADwithSmoking-None 5)) 1000 “Discontinuation of smoking should help relieve the symptoms. ” HowToDetermincBelicf finelion for compuring the minimum of the beliefs of the manifestations Figure 1. English translation of PUFF rule (top) and Corresponding WHEEZE Frame for OADwithSmoking (bottom). Numbers appearing in the Manifestation, SuggcstiveOf and Complcmcntaryl’o slots arc importance and suggcstivity wcightings. From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. The declarative nature of this representation facilitates modification and extension. For example, the addition of related knowledge, such as justifications, explanations, and instructional material, can be accomplished by the addition of slots to already existing assertions. The single uniform structure alleviates the need for any auxilliary means of representation. Considerable efficiency has been gained by the use of rule compilation on production systems [lo] [ll]. We feel that a technique similar to this could also be used effectively on our representation but have not yet fully investigated this possibility. AN AGENDA BASED CONTROL MECHANISM Depth-first, goal-directed search is often used in production systems because questions asked by the system are focused on specific topics. Thus, the system appears to follow a coherent line of reasoning, more closely mimicking that of human diagnosticians. There are, however, many widely recognized limitations. No mechanism is provided for dynamically selecting or ordering the initial set of goals. Consequently, the system may explore many “red herrings” and ask irrelevant questions before encountering a good hypothesis. In addition, a startling piece of evidence (strongly suggesting a different hypothesis) cannot cause suspension of the current investigation and pursuit of the alternative. Expert diagnosticians use more than simple goal-directed reasoning. They seem to work by alternately constructing and verifying hypotheses, corresponding to a mix of data- and goal- directed search. Furthermore, they expect these systems to reason in an analogous manner. It is desirable, therefore, that the system builder have control over the dynamic reasoning behavior of the system. To provide this control, we employ a simple relaxation of goal- and data-directed mechanisms. This is facilitated by the use of an agenda to keep track of the set of goals to be examined, and their relative priorities. The control strategy is: 1. Examine the top assertion on the agenda. 2. If its sub-assertions (manifestations) arc known, the relative belief of the assertion is determined. If confirmed, any assertions that it is suggestive of are placed on the agenda according to a specified measure of suggestivity. If denied, complementary assertions are placed on the agenda according to a measure of suggestivity. 3. If it cannot be immediately verified or rejected then its unknown sub-assertions are placed on the agenda according to a measure of importance, and according to the agenda level of the original assertion. By varying the importance factors, SuggesfiveOf values, and the initial items placed on the agenda, numerous strategies are possible. For example, if high-level goals are initially placed on the agenda and subgoals are always placed at the top of the agenda, depth-first goal-directed behavior will result. Alternatively, if low-level data are placed on the agenda initially, and assertions suggested by these data assertions arc always placed below them on the agenda, breadth-first data driven behavior will result. More commonly, what is desired is a mixture of the two, in which assertions suggest others as being likely, and goal directed verification is employed to investigate the likely assertions. The example below illustrates how this can be done. FEVl/FVC<80 QAD FEVl/FVC>SO Figure 2. A simplified portion of the WHEEZE knowledge base. The solid lines indicate Manifesatation links (e.g. OAD is a manifestation of Asthma), the dashed lines represent SuggestiveOf links. The numbers associated with the links are the corresponding “importances” and “suggestivities” of the connections. In -the knowledge base of figure 2, suppose that RDX-ALS is confirmed, suggesting RLD to the agenda at level 6 and ALS at level 4. ‘RLD is then examined, and since its manifestations are unknown, they are placed at the specified level on the agenda. The agenda now contains FEVl/FVC>80 at level 8, RV<80 and RLD at level 6, and ALS at level 4. FEVl/FVC>80 is therefore selected, and suppose. that it is found to be false. Its complementary assertion (FEVl/FVC<80) is placed at level 8 on the agenda. and is immediately investigated. It is, of course, true, causing OAD to be placed at level 8 on the agenda. The diagnosis proceeds by investigating the manifestations of OAD; and, if OAD is confirmed, Asthma and Bronchitis are investigated. While many subtleties have been glossed over in this example, it is important to note that: 1. The manipulation of SuggestiveOf and importance values can change the order in which assertions are examined, therefore changing the order in which questions are asked and results printed out. (In the example, FEVUFVC was asked for before RV.) 155 2. Surprise data (data contrary to the hypothesis currently being investigated) may suggest goals to the agenda high enough to cause suspension of the current investigation. (The surprise FEVl/FVC value caused suspension of the RLD investigation in favor of the OAD investigation. If the suggestivity of the link from FEVl/FVC<80 to OAD were not as high, this would not have occurred.) 3. Low-level data assertions cause the suggestion of high- level goals, thus selecting and ordering goals to avoid irrelevant questions. (In the ex‘ample, RLD and ALS were suggested and ordered by the low-level assertion RDX-ALS.) Thus, extreme control flexibility is provided by this mechanism. Besides the mechanism proposed above, there have been several other attempts to augment simple goal directed search. Meta-rules [S] can be used to encode strategic information, such as how to order or prune the hypothesis space. They could also be used, in principle, to suspend a current investigation when strong alternatives were discovered. In ‘practice, however, meta-rules for accomplishing this task could be quite clumsy. In the CENTAUR system [l] [2], procedural attachment mechanisms (in disease prototypes) are used to capture the control information explicitly, and “triggering” rules serve to order the initial hypothesis space. Our solution differs from these earlier attempts by proposing a single uniform control mechanism. lt is sufficiently straightforward that tailoring of the control flow could potentially be turned over to the domain expert. RIZSULTS Not suprisingly, WHEEZE exhibits the same diagnostic behavior as its predecessors, PUFF and CENTAUR, on a standard set of 10 patient test cases. In refining the knowledge base, suggestivities and importance factors were used to great advantage to change the order in which questions were asked and conclusions printed out. This ‘eliminated the need to carefully order sets of antecedent assertions. The reprcscntation described has proven adcquatc for capturing the domain knowledge. In some cases, several rules were collapsed into a single assertion. In addition, the combination of representation and control structure eliminated the need for many awkward interdependent rules (e.g. rules with screening clauses). Representation of both the rule and non-rule knowledge of the PUFF and CENTAUR systems has been facilitated by the flexibility of the architecture described. This flexibility is the direct result of the uniform representation and control mechanism. Further exploitations of this architecture appear possible, providing directions for future research. ACKNOWLEDGEMENTS We would like to thank Jan Aikins, Avron Barr, James Bennett, Bruce Buchanan, Mike Genesereth, Russ Greiner, and Doug Lenat for their help and comments. 111 PI t31 141 151 161 171 PI [91 [lOI REFERENCES Jan. S. Aikins. Prototypes and Production Rules: an Approach to Knowledge Representation for Hypothesis Formation. Proc 6th IJCAI, 1979. Pp. l-3. Jan S. Aikins. Profolypes and Producrion Rules: A Knowledge Represenrarion for Computer Consullations Doctoral dissertation, Dept. of Computer Science, Stanford University. Jan S. Aikins. Representation of Control Systems. Proc Zsr AAAI, 1980. Knowledge in Expert James Bennett, Lewis Creary, Robert Englemorc and Robert Melosh. SACON: A Knowledge Based Consultant for Sfructural Analysis. Computer Science Report CS-78-699, Dept. of Computer Science, Stanford University, September 1978. Randall Davis and Bruce G. Buchanan. Meta-Level Knowledge: Overview and Applications. Proc Slh IJCAI, 1977. Pp. 920-927. Russell Greiner, Douglas Lenat. A Language. Proc Zsf AAAI, 1980. Representation Language J. C. Kunz, R. J. Fallet, et. al.. A Physiological Rule Based System for Inlerpreting Pulmonary Function Test Results. HPP- 78-19 (Working Paper), Heuristic Programming Project, Dept. of Computer Science, Stanford University, December 1978. Stephen G. Pauker and Peter Szolovits. Analyzing and Simulating Taking the History of the Present Illness: Context Formation. In Schneider and Sagvall Hein (Eds.), Compulational Linguislics in Medicine. North-Holland, 1977. Pp. 109-118. E. hi. Shortliffe. Computer-based Medical Consultations: MYCIN. New York: American Elsevier, 1976. William van Melle. A System for Consultation 923-92s. Domain-independent Production-rule Programs. Proc 6th IJCAI, 1979. Pp. [ll] William van Melle. A Domain-independenr Syslem thai Aids in Construcling Knowledge-based Consultation Programs. Doctoral dissertation, Dept. of Computer Science, Stanford University. 156	1980	41
36	ABSTRACT Knowledge Embedding in the Description System Omega Carl Mewitt, 6iuseppe Attardi, and Maria Simi M.I.T. 545 Technology Square Cambridge, Mass 02139 Omega is a description system for knowledge embedding which combines mechanisms of the predicate calculus, type systems, and pattern matching systems. It can express arbitrary predicates (achieving the power of the o-order quantificational calculus), type declarations in programming systems (Algal, Simula, etc.), pattern matching languages (Planner, Merlin, KRL, etc.). Omega gains much of its power by unifying these mechanisms in a single formalism In this paper we present an axiomatization of basic constructs in Omega which serves as an important component of the interface between implementors and users+ Omega is based on a small number of primitive concepts. It is sufficiently powerful to be able to express its own rules of inference In this way Omega represents a self-describing system in which a great deal of knowledge about itself can be embedded. The techniques in Omega represent an important advance in the creation of self-describing systems without engendering the problems discovered by Russell Meta-descriptions (in the sense used in mathematical logic) are ordinary descriptions in Omega. Together with Jerry Barber we have constructed a preliminary implementation of Omega on the MLT. CADR System and used it in the development of an office workstation prototype. I -- Introduction First Order Logic is a powerful formalism for representing mathematical theories and formalizing hypotheses about the world. Logicians have developed a mathematical semantics in which a number of important results have been established such as completeness. These circumstances have motivated the development of deductive systems based on first order predicate calculus [FOL, PROLOG, Bledsoe’s Verifier, etc. ] However, First Order Logic is unsatisfactory as a language for embedding knowledge in computer systems. Therefore tnany recent reasoning system have tried to develop their own formalisms [PLANNER, FRL, KL-ONE, KRL, LMS, NETL, AMORD, XPRT, ETHERJ The semantics and deductive theory of these new systems however has not been satisfactorily developed. The only rigorous description of most of them has been their implementations which are rather large and convoluted programs. 2 -- Overview The syntax of Omega is a version of template English, For example we use the indefinite article in instance descriptions such as the one below: (a Son) Instance descriptions like the previous one in general describe a whole category of objects, like the category of sons, in this example Such description however can be tnade more specific, by ‘prescribing particular attributes for the instance ‘description So for example, (a Son (With father Paul) (With mother Mary)) describes a son with father Paul and with mother Mary. Otnega differs from systems based on records with attached procedures (SIMULA and its descendants), generalized property lists (FRL, XRL, etc.), frames ‘(Minsky), and units (KRL) in several important respects. One of the most important differences is that instance descriptions in Omega cannot be updated. This is a consequence of the monotonicity of knowledge accutnulation in Omega Change in Omega is modeled through the use of viewpoints [Barber: 19801 Another difference is that in Omega an instance description can have more than one attribution with the same relation, For example (a Human (with child Jack) (with child Jill)) is a description of a human with a child Jack and a child Jill Statements inheritance can be relation deduced is. For of because example transitivity of the 157 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. (John is (a Man)) can be deduced from the following statements (John is (a Son)) ((a Son) is (a Man)) In order to aid readability we will freely mix infix and prefix notations. For example the statement (John is (a Man)) is completely equivalent to (is John (a Man)) 3 -- Inheritance The inheritance relation in Omega differs somewhat from the usual ISA relation typically found in semantic networks. For example from (John iS (a Human)) ((a Human) is (a Mammal)) (Human is (a Species)) we can deduce (John is (a Mammal)) but cannot conclude that (John is (a Species)). However we can deduce that (John is (a (a Species))) which says that John is something which is a species We can often avoid the use of explicit universal quantifiers. For instance the following sentence in quantificational calculus Vx Man(x) * Mortal(x) can be expressed as ((a Man) is (a Mortal)) In this case we avoid the need to restrict the range of a universal quantifier by means of a predicate, as it is usually necessary in the quantificational calculus. When unrestricted quantification is required, or when we need to give a name to a description which occurs several times in the same expression, we denote a universally quantified variable by prefixing the symbol = to the name of the variable, wherever it appears in the statement, as in: (=x is (a Man)) * (=x is (a Mortal)) The scope of such a variable is the whole statement. Thus the above statement is an abbreviation for (for-a// =x ((=x is (a Man)) 3 (=x is (a Mortal)))) Occasionally it is necessary to use a for-all in the interior of a statement. For example in the following statement expresses the Axiom of Extensionality which is one of the most fundamental principles in Omega: ((for-a// =d ((=d is =dr) * (=d is =d2))) 3 (=dl is =d2)) In the above statement, the scope of =dl and =d2 is the whole statetnent, while the scope of =d is only the statetnent ((=d is =dl) * (=d is =d2)). A form of existential quantification is implicit in the use of attributions+ For instance (Pat is (a Man (with father (an Irishman)))) says that there is an Irishman who is Pat’s father. Omega makes good use of the ability to place nontrivial descriptions on the left hand side of an inheritance relation. For example from the following statements (a Teacher (with subject =s)) iS (an Expert (W/f/l field ES)) (John is (a Teacher (With subject Music))) we get the following by transitivity of inheritance: (John iS (an Expert (Wifh field Music))) Note that statements like the following (is (and (a WarmBloodedAnimal) (a BearerOfLiveYoung)) (a Mammal)) are much more difficult to express in systems such as KRL and FRL which are based on generalized records and property lists. If it happens that two descriptions inherit from each other, we will say they are the same. For example if ((a Woman) is (a Human (Wifh sex female))) ((a Human (wifh sex female)) is (a Woman)) then we can conclude (a Woman) Same (a Human (wifh sex female)) We can express the general principle being used here in Omega as follows 158 ((=dl same =d2) <=> (A f=dl is =d2) (=d2 is =dl))) 4 -- Lattice Operators and Logical Operators The domain of descriptions constitutes a complemented lattice, with respect to the inheritance ordering is, meet and join operations and and or, complementation operation not, and Nothing and Something as the bottom and top respectively of the lattice Some axioms are required to express these relations For example all descriptions inherit from Something. (=d is Something) Furthermore Nothing is the complement of Something (Nothing same (not Something)) The usual logical operators on statements are A, V, -, => for conjunction, disjunction, negation, and implication respectively. The description operators not, and, or, etc. apply to all descriptions including statements. It is very important not to confuse the logical operators with the lattice operators in Omega. Note for example: ((A true false) is false) ((and true false) is /Vothingj ((and true true) is true) Unfortunately most “knowledge representation languages” have not carefully distinguished between lattice operators and logical operators leading to a great deal of confusion. Note that a statement of the form (p is true) does not in general imply that (p same true). For example ((Nixon iS (a UnindictedCoConspirator)) is true) (((a Price (with merchandise Tea) (with place China)) is 81) is true) does not imply (same (Nixon is (a UnindictedCoConspirator)) (a Price (with merchandise Tea) (with place China))) 5 -- Basic Axioms We will state some of the axioms for the description system. The axioms for a theory are usually stated in a metalanguage of the theory. However, since our language contains its metalanguage, we can here give the axioms as ordinary statements in the description system itself. 5.1 Extensionality Inheritance obeys an Axiom of Extensionality which is one of the most fundamental axioms of Omega. Many important properties can be derived from extensionality which can be expressed in Omega as follows: (W (=descriptionl iS =description2) (for-a// =d (=j (=d iS =descriptionl) (=d iS zdescription2)))) Note that the meaning of the above statement would be drastically changed if we simply omitted the universal quantifier as follows (W (=descriptionl iS =description2) (3 (=d iS =descriptionl) (=d is =description2))) The axiom extensionality illustrates the utility of explicitly incorporating quantification in the language in contrast to some programming languages which claim to be based on logic. From this axiom alone we are able to derive most of the lattice-theoretic properties of descriptions. In particular we can deduce that is is a reflexive and transitive relation. The following (=description iS =description) expresses the reflexivity of inheritance whereas the following (4 (A (=descriptionl /S =description2) (=description2 is =descriptions)) kdescriptionl is zdescriptiong)) expresses the transitivity of inheritance. 5.2 Commutativity Commutativity says that the order in which attributions Of a concept are written is irrelevant We use the notation that an expression of the form XC..>> is a sequence of 0 or more elements 159 (same (a =descriptionl <<=attributionsl>> =attribution2 <<=attributions3>> =attributionq <<=attributionsg>>) (a =descriptionl <<=attributionsl>> =attributionq <<=attributionsg>> =attribution2 <<=attributionsg>>)) (Susan is (a Mother (with child Jim) (with father Bill) (with child (a Female)))) 5.5 Monotonicity of Atributes Monotonicity of attributes is a fundamental property of instance descriptions which is close)y related to transitivity of inheritance. (=descriptionl is =description2) (is (a =concept (With =attribute =descriptionl)) (a =concept (With =attribute =descriptionp)))) For example ((a Father (with child Henry) (With mother Martha)) same (a Father (with mother Martha) (with child Henry))) For example if 5.3 Deletion (Fred is (an American)) (Bill is (a Person (with The axiom of Deletion is that attributions of an instance description can be deleted to produce a more general instance description. (is (a =descriptionl <<=attributions-l>> zattribution-2 <<=ettributiona-3 (a =descriptionl <<=attributions-I>> <<=attributions-3)))) father Fred))) then (Bill is (a Person (With father (an kkTWkaI7~~~~ Note that the complementation in monotonic. For example Omega is not For example ((a Bostonian) is (a NewEnglander)) (is (a Father (With child Henry) (With mother Martha)) (a Father (with mother Martha))) does not imply that ((not (a Bostonian)) is (not (a NewEnglander))) 5.6 Constraints 5.4 Merging Constraints can be used to restrict the objects will satisfy certain attributions. For example which One of the most fundamental axioms in Omega is Merging which says that attributions of the same concept can be merged. (a Human (withh?straint child (a Male))) describes humans who have Axiom for Constraints is only male children. The (=descriptionl is (a =description2 <<=attributions-l>>)) kdescriptionl is (a =description2 <<=attributions-2))))) (is =descriptionl (a =description2 <<=attributions-1>> <<=attributions-2))))) ((a =C (withconstraint =R =dl) (with =R =d2)) is (a =C (with =R (and=dl =d2)))) If For example if (Joan is (a Human (WithConstraint child (a Male)) (With child Jean))) (Susan is (a Mother (with child Jim))) (Susan is (a Mother (with father Bill))) (Susan is (a Mother (With child (a Female)))) 160 then 6.2 Projective Relations (Joan iS (a Human (with child (SfK! (a Male) Jean)))) Note that solely from the statement (Ann iS (a Human (with child (a Male)) (With child Jean))) no important conclusions can be drawn in Omega. It entirely possible that Jean is a female with a brother. is We have found the constrained attributions in Omega to be useful generalizations of the increasingly popular “constraint languages” which propagate values through a network of property lists. If (2 is (a Complex (with real-part (> 0)))) and (2 iS (a Complex (With real-part (an Integer)))) then by merging it follows that (z iS (a Complex (With real-part (> 0)) (with real-part (an Integer)))). However in order to be able to conclude that (z iS (a Complex (With real-part (and (> 0) (an Integer))))) some additional information is needed. One very general way to provide this information is by (rsalgart iS (a Projective-relation (with concept Complex))) 6 -- Higher Order Capabilities and by the statement In this section we present examples which illustrate power of the higher-order capabilities of Omega. the 6.1 Transitive Relations If (3 is ($7 Integer (with larger 4))) and (4 is (an Integer (with larger 5))), we can conclude by monotonicity that (3 is (an Integer (With larger (an Integer (with larger 5))))) From the above statement, we would like to be able to conclude that (3 is (an Integer (with larger 5))). This goal can be accomplished by the statement (larger is (a Transitive-relation (with concept Integer))) which says that larger is a transitive relation for the concept Integer. The Axiom for Transitive Relations states that if R is a transitive relation for a concept c and x is an instance of c which is R-related to an instance of c which is R-related to m, then x is R-related to m. (=> (=R iS (a Transitive-relation (With concept =C))) (is (a =C (with =R (a =C (with =R =m)))) (a =C (with =R m)))) The desired conclusion can be reached by using the above description with c bound to Integer, R bound to larger, and m bound to 5. (=R k (a Projective-relation (with concept =C))) (is (a =C (with =R =d)) (a =C (wifhConsfrainf =R =d)))) The desired conclusion is reached by using the above description with =C bound to Complex, =R bound to real-part, =descriptionl bound to (> 01, and =description2 bound to (an Integer). 6.3 Inversion Inverting relations for efficiency of retrieval is a standard technique in data base organization. Inversion makes use of the converse of a relation with respect to a concept which satisfies the following Axiom for Converse: (=R same (a Converse (With relation (a Converse (with relation =R) (With concept =C))) (With concept =C))) The Axiom of Inversion expresses how to invert inheritance relations for constrained instance descriptions (<=> (=dl is (a =C (withConstraint =R (an =d2)))) ((a =R (wifh (a Converse (with relation =R) (With relation =C) =dl)) is =d2))) 161 For example suppose ((a Converse (with relation son) (with concept Person)) same Parent) we can conclude (Sally iS (a Person (wifhConsfrainf son (an American)))) if and only if ((a Son (with parent Sally)) is (an American)) We have inversion to be a useful generalization of the generalized selection mechanisms in Simula, SmallTalk, and KRL as well as the generalized getprop mechanism in FRL. The interested reader might try to define the transitivity, projectivity, and converse relations in other “knowledge representation languages” 7 -- Conclusions Omega encompasses ‘?he capabilities of both the w-order quantification calculus, type theory, and pattern matching languages in a unified way. We have illustrated how Omega is more powerful than First Order Logic by showing how it can directly express important properties of relations such as transitivity, projectivity, and converse that . are not first order definable Omega is based on a small number of primitive concepts including inheritance, instantiation, attribution, viewpoint, logical operations (conjunction, disjunction, negation, quantification, etc.) and lattice operations ( meet, join, complement, etc.) It makes use of inheritance and attribution between descriptions to build a network of descriptions in which knowledge can be embedded. Omega is sufficiently powerful to be able to express its own rules of inference. In this way Omega represents a self-describing system in which a great deal of knowledge about itself can be embedded. Because of its expressive power, we have to be very careful in the axiom system for Omega in order to avoid Russell’s paradox. Omega uses mechanisms which combines ideas from the Lambda Calculus and Intutionistic Logic to avoid contradictions in the use of self reference We have found axiomatization to be a powerful technique in the development, design, and use of Omega Axiomatization has enabled us to evolve the design of Omega by removing many bugs which have shown up as undesirable consequences of the axioms. The axiomatization has acted as a contract between the implementors and users of the system. The axioms provide a succinct specification of the rules of inference that can be invoked. The development of Omega has focused on the goals of conceptual simplicity and power. The axiomatization of Omega in itself is a measure of our progress in achieving these goals. 8 -- Related Work The intellectual roots of our description system go back to von Neumann-Bernays-Goedel set theory [Goedel: 19401, the o-order quantificational calculus, and the lambda calculus. Its development has been influenced by the property lists of LISP, the pattern matching constructs in PLANNER-71 and its descendants QA-4, POPLER, CONNIVER, etc., the multiple descriptions and beta structures of .MERLIN, the class mechanism of SIMULA, the frame theory of Minsky, the packagers of PLAShlA, the stereotypes in [Hewitt: 19751, the tangled hierarchies of NETL, the attribute grammars of Knuth, the type system of CLU, the descriptive mechanisms of KRL-0, the partitioned semantic networks of [Fikes and Hendrix: 19771, the conceptual representations of [Yonezawa: 19771, the class mechanism of SMALL-TALK [Ingalls: 19781, the goblets of Knowledge Representation Semantics [Smith: 19781, the selector notation of BETA, the inheritance mechanism of OWL, the mathematical semantics of actors (Hewitt and Attardi: 19781, the type system in Edinburgh LCF, the XPRT system of Luc Steels, the constraints in [Borning: 1977, 1979 and Steele and Sussman: 1978] 9 -- Further Work We have also developed an Omega Machine (which is not described in this paper) that formalizes the operational semantics of Omega. Mike Brady has suggested that it might be possible to develop a denotational semantics for Omega along the lines of Scott’s model of the lambda calculus. This development is one possible approach to establishing the consistency of Omega. 162 10 -- Acknowledgments We are grateful to Dana Scott for pointing out a few axioms that were incorrectly stated in a preliminary version of this paper. Jerry Barber has been extremely helpful in aiding us in developing and debugging Omega. Brian Smith and Gene Ciccarelli helped us to clear up some important ambiguities. Conversations with Alan Borning, Scott Fahlman, William Martin, Allen Newell, Alan Perlis, Dana Scott, Brian Smith, and the participants in the “Message Passing Systems” seminar were extremely helpful in getting the description system ‘nailed down. Richard Weyhrauch has raised our interests in meta-theories. His system FOL is one of lthe first to exploit the classical logical notion of metatheory in AL systems Several discussions with Luc Steels have been the source of cross-fertilization between the ideas in our system and his XPRT system ‘Roger Duffey and Ken Forbus have served as extremely able teaching assistants in helping to develop this material for the Paradigms for Problem Solving Course at MIT. Comments by Peter Deutsch and Peter Szolovits have materially helped to improve the presentation. Our logical rules of inference are a further development of a natural deduction system by Kalish and Montague. Some of the axioms for inheritance were inspired by Set Theory. 11 -- Bibliography Barber, G. “Reasoning About Change in Knowledgeable Office Systems” 1980. Birtwistle, G. M; Dahl, 0.; Myhrhaug, B.; and Nygaard, K. “SIM..JLA Begin” Auerbach 1973. Bobrow, D. G. and Winograd, T. “An Overview of KRL-0 Knowledge Representation Language” ’ Co:nitive Science VoL 1 No. I. 1977. Borning, A. “ThingLab -- An Object-Oriented System for Building Simulations Using Constraints” Proceedings of IJCAI-77. August, 1977. Bourbaki, N. “Theory of Sets” Book I of Elements of Mathematics. Addison-Wesley. 1968. Church, A.. “A Formulation of the Simple Theory of Types”, 194 1. Dahl, 0. J. and Nygaard, K. “Class and Subclass Declarations” In Simulation Programming Languages J. N. Buxton (Ed.) North Holland. 1968. pp. 158-174. Fahlman, Scott “Thesis Progress Report” MIT AI Memo 331. May, 1975. Fikes, R. and Hendrix, G. “A Network-Based Knowledge Representation and its Natural Deduction System” IJCAI-77. Cambridge, Mass. August 1977. pp 235-246. Goedel, K “The Consistency of the Axiom of Choice and of the Generalized Continuum Hypothesis with the Axioms of Set Theory” Annals of Mathematics Studies No. 3, Princeton, 1940. Hammer, M. and McLeod, D. “The Semantic Data Model: A Modeling &Mechanism for Data Base Applications SIGMOD Conference on the Management of Data. Austin Texas. May 31-June 2, 1978. Hawkinson, Lowell “The Representation of Concepts in OWL” Proceedings of IJCAI-75. September, 1975. Tiblisi, Georgia, USSR pp. 107-114. Hewitt, C. “Stereotypes as an ACTOR Approach Towards Solving the Problem of Procedural Attachment in FRAME Theories” “Proceedings of Interdisciplinary Workshop on Theoretical Issues in Natural Language Processing” Cambridge, June 1975. Kalish and Montague Kristensen, B. B.; Madsen, 0. L.; Moller-Pedersen, B.; and Nygaard, K. “A Definition of the BETA Language” TECHNICAL REPORT TR-8. Aarhus University. February 1979. Moore, J. and Newell, A. “How Can MERLIN Understand?” CMU AM. November, 1973. Quine, W. K. “New Foundations of Mathematical Logic” 1952 Burstall, R and Goguen , J. “Putting Theories Together to Make Specifications”, Proceedings of IJCAI-77. August, 1977. 163 Rulifson, J. F.; Derksen, J. A.; and Waldinger, R. J. “QA4: A Procedural Calculus for Intuitive Reasoning” SRI Technical Note 73. November 1972. Schubert, L. K. “Extending the Expressive Power of Semantic Networks” Artificial Intelligence 7. Steele, G. L. and Sussman, G. J. “Constraints” MIT Artificial Intelligence Memo 502. November 1978. Steels, L. Master Thesis, MIT 1979. Weyhrauch, R. “Prolegomena to a Theory of Formal Reasoning”, Stanford AI Memo AIM-315, December 1978. Forthcoming in AL JournaL 164	1980	42
37	A Representation Language Language Russell Greiner and Douglas B. Lenat Computer Science Deptartment Stanford University ABSTRACT The field of AI is strewn with knowledge representation languages. The language designer typically has one particular application domain in mind: as subsequent types of applications are tried, what had originally been US&II j2urure.y become undesirable limitations, and the language is overhauled or scrapped. One remedy to this bleak cycle might be to construct a representational scheme whose domain is the field of representational languages itself. Toward this end, we designed and implemented RLL, a frame-based Representation Languange Language. The components of representation languages in general (such ti slots and inheritance mechanisms) and of RLL itself are encoded declaratively as frames. Modifvinn these frames can change the of the RLL environment. semantics of RLL, by altering the 1. MOTIVATION “One ring to rule them all... and in the darkness bind them. ” Often a large Artificial Intelligence project begins by designing and implementing a high-level language in which to easily and precisely specify the nuances of the task. The language designer typically builds his Representation Language around the one particular highlighted application (such as molecular biology for Units [Sfe$k], or natural language understanding for KRL [Bobrow & Winogrudj and OWL [Szolovifs, et al.]). For this reason, his language is often inadequate for any subsequent applications (except those which can be cast in a form similar in structure to the initial task): what had originally been useful features become undesirable limitations (such as Units’ explicit copying of inherited facts, or KRL’s sophisticated but slow matcher). Building a new language seems cleaner than modifying the flawed one, so the designer scraps his “extensible, general” language after its one use. The size of the February 1980 SIGART shows how many similar yet incompatible representation schemes have followed this evolutionary path. One remedy to this bleak cycle might be to construct a representation scheme whose domain is the field of representational languages itself. a program which could thc’n bc tailored to suit many specific applications. Toward this end, WC arc designing and implementing RLL, an object-centered’ Representation Languange I,anguage.2 This paper reports on the current state of our ideas and our implementation. 1 This “object-centering” does not represent a loss in geneiality. We will soon see that each part of the full system, including procedural information, is reified as a unit. * As RLL is itself a completely self-descriptive there is no need for an RLLL. representation language, 2. INTRODUCTION RLL explicitly represents (i.e. contains units-? for) the components of reprcscntation languages in general and of itself in particular. The programming language LISP derives its flexibility in a similar manner: it, too, encodes many of its constructs within its own formalisms. Representation languages aim at easy, natural interfacing to users; therefore their primitive building blocks are larger, more abstract, and more complex than the primitives of programming languages. Building blocks of a representation language include such things as control regimes (ExhaustiveBackwardChaining, Agendas), methods 3f associating procedures with relevant knowledge (Footnotes, Demons), tindamental access functions (Put/Get, Assert/Match), automatic inference mechanisms (InheritFromEvery2ndGeneration, InheritBut-PermitExceptions), and even specifications of the intended semantics and epistemology (ConsistencyConstraint, EmpiricalHeuristic). of the components RLL is designed to help manage these complexities, by providing (1) an organized library of such representation language components, and (2) tools for manipulating, modifying, and combining them. Rather than produce a new representation language as the “output” of a session with RLL, it is rather the RLL language itself, the environment the user sees, which changes gradually in accord with his commands. 3. HOWJS A REPRESENTATION LANGUAGE LIKE AN ORGAN? When the user starts RLL, he finds himself in an environment very much like the Units package [Slefik], with one major difference. If he desires a new type of inheritance mechanism, he need only create a new Inheritance-type of unit, initialize it with that desired property; and that new mode of inheritance will automatically be enabled. This he can do using the same editor and accessing functions hc uses for cntcring and codifying his domain knowledge (say, poultry inspection); only hcrc the information pertains to the actual Knowledge Base system itself, not turkeys. The Units package has Get and Put as its tindamcntal storage and retrieval functions; therefore RLL also begins in that state. But there is nothing sacred about cvcn these two “primitives”. Get and Put arc encoded as modifiable units: if they are altered, the nature of accessing iI slot’s value will change correspondingly. In short, by issuing a small number of commands hc can radically alter the character of the RLI. cnvironmcnt. molding it to his personal -----______________ 3 RLL is a frame-based system [Minsky], whose building blocks are called Units [Stefik], [Bobrow & Winograd]. Each unit consists of a set of Slots with their respective values. 165 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. preferences and to the specific needs of his application. RLL is responsible for performing the necessary “truth maintainence” operations, (e.g. retroactive updates) to preserve the overall correctness of the system as a whole. For example, Get and Put can be transformed into units which, when run as functions, resemble Assert (store proposition) and Match (retrieve propositions), and the user need never again mention “slots” at all. RLL is more like a stop organ than a piano. Each stop corresponds to a “pre-fabricated” representational part (e.g. a slot, inheritance, format, control regime, etc.), which resides in the overall RLL system. The initial RLL is simply one configuration of this organ, with certain stops “pulled out” to mimic the Units package. These particular stops reflect our intuitions of what constitutes a general, powerful system. Some of the units initially “pulled out” (activated) define more or less standard inheritance regimes, such as Inherit-Along-IS-A-Links, which enables Fido to gather default information from AnyDog. We chose to include a large assortment, of common slots. One hundred and six types of slots, including IS-A, SuperClass. BroaderHeuristics, and TypicalExamples, are used to hierarchically organize the units. That number grows daily, as we refine the organizing relationships which were originally “smeared” together into just. one or two kinds of slots (e.g. A-Kind-Ofl. An additional fifteen types of slots, including ToGerValue, ToPut Value, ToKillUnit. and To Add Value, collectively define the accessing/updating functions. This bootstrapping system (the initial configuration of “organ stops”,) dots not span the scope of RLL’s capabilities: many of its stops are initially in the dormant position. Just as a competent musician can produce a radically different sound by manipulating an organ’s stops, so a sophisticated RLL user can define his own representation by turning off some features and activating others. For instance, an FRL devotee may notice -- and choose to use exclusively -- the kind of slot called A-Kind-Ox which is the smearing together of Is- A, SuperSet, Abstraction, TypicalExampleOJ PartOf etc. He may then deactivate those more specialized units from his system permanently. A user who does not want to see his system as a hierarchy at all can simply deactivate the A-Kind-Of unit and its progeny. The user need not worry about the various immediate, and indirect, consequences of this alteration (e.g., deleting the Inherit-Along-IS-A-Links unit); RLL will take care of them. By selectively pushing and pulling, he should be able to construct a system resembling almost any currently used representational language, such as KRL, OWL and KLONE;4 after all, an organ can be made to sound like a piano. Unlike musical organs, RIL also provides its user with mechanisms for building his own stops (or even type of stops, or even mechanisms for building stops). With experience, one can use RLL to build his own new components. Rather than building them from scratch, (e.g., from CAR, CDR, and CONS,) he can modi& some existing units of RLL (employing other units which are themselves tools designed for just such manipulations.) 4 This particular task, of actually simulating various existing Representation languages, has not yet been done. It is high on our agenda of things to do. We anticipate it will require the addition of many new components (and types of components) to RLL, many representing orthogonal decompositions of the space of knowledge representation. The following examples convey the flavor of what can currently done with the default settings of the RLL “organ stops”. 4. EXAMPLES be 4.1. EXAMPLE: Creating a New Slot In the following example, the user wishes to define a Fafher slot, in a sexist genealogical knowledge base which containes only the primitive slots Morher and Spouse. As RLL devotes a unit to store the necessary knowledge associated with each kind of slot, (see Figure 1,) defining a new kind of slot means creating and initializing one new unit. In our experience, the new desired slot is frequently quite similar to some other slot(s), with but a few distinguishing differences. We exploited this iegularity in developing a high level “slot-defining” language, by which a new slot can be defined precisely and succinctly in a single declarative statement. 1 Name: IS-A 1 Description: Lists the classes I AM-A member of. I Format: List-of-Entries 1 Datatype: Each entry represents a class of objects. I Inverse: Examples 1 IS-A: (AnySlot) 1 UsedSylnhefifance: Inherit-Along-/S-A-Links 1 MyTimeOfCfeation: 1 April 1979, 12:Ol AM 1 MyCreator: D.B.Lenat Figure # 1 - Unit devoted to the “IS-A” slot. There are many other slots which are appropriate for this unit; whose value will be deduced automatically (e.g. inherited from AnySlot) if requested. Creating a new slot for Father is easy: we create a new unit called Father, and fill its HighLevelDefn slot with the value (Composition Spouse Mofher). Composition is the name of a unit in our initial system, a so-called “slot-combiner” which knows how to compose two slots (regarding each slot as a function from one unit to another). We also fill the new unit’s Isa slot, deriving the unit shown in Figure 2. Name: 1 IS-A: I HighLevelDefn: Father (AnySlot) (Composition Spouse Mother) Figure # 2 - Slots filled in by hand when creating the unit devoted to the”Father” slot. Several other slots (e.g., the syntactic slots MyCreator, MyTimeOfCreation) are filled in automatically at this time. The user now asks for KarlPhilippEmanuel’s father, by typing (GetVal ue 'KPE 'Father). GetValue first tries a simple associative lookup (GET), but finds there is no Fafher property stored on KPE, the unit representing KarlPhilippEmanuel. GctValue then tries a more sophisticated approach: ask the Father unit how to compute the Father of any person. Thus the call becomes [Apply* (GetValue 'Father 'ToCompute) 'KPE]. Notice this calls on GctValue recursivelv. and once again there is no value stored on the ToCompure slot-of the unit tilled Father. The call now has expanded into [Apply* (Apply* (GetValue 'ToCompute 'ToCompute) 'Father) 'KPE]. Luckily, there is a value on the ToCompute slot of the unit ToCompute: it says to find the HighLevelDefn, find the slot- combiner it employs, find its ToComput$, and ask it. Our call is now cxpandcd out into the following: [Apply* (Apply* (GetValue 'Composition 'ToCompute) 'Spouse 'Mother) 'KPE]. ’ Each unit which represents a function has a ToCompure slot, which holds the actual LISP function it encodes. Associating such a ToCompute slot with each slot. reflects our view that each slot is a function, whose argument happens to be a unit, and whose computed value may be cached away. 166 The unit called Composition does indeed have a ToCompute Slot; after applying it, we have: CApply* ‘(h (x) (GetValue (GetValue x ‘Mother) ‘Spouse)) ‘KPE]. This asks for the Mother slot of KPE, which is always physically stored in our knowledge base, and then asks for the value stored in her Spouse slot. The final result, JohannSebastian, is then returned. It is also cached (stored redundantly for future use) on the Fafher slot of the unit KPE. See [Lenat et al., 19791 for details. Several other slots (besides ToCompufe) are deduced automatically by RLL from the HighLevelDefn of Father (see Figure 3) as they are called for. The Format of each Fafher slot must be a single entry, which is the name of a unit which represents a person. The only units which may have a Fafher slot are those which may legally have a Mofher slot, viz., any person. Also, since Fafher is defined in terms of both Mother and Spouse, using the slot combiner Composition, a value stored on KPE:Fafher must be erased if ever we change the value for KPE’s M&her or AnnaMagdelena’s Spouse, or the definition (that is, 7’oCompufe) of Composition. Name: 1 IS-A: Father (AnySlot) I 1 HighLevelDefn: (Composition Spouse Mother) 1 Description: Value of unit’s Mother’s Spouse. 1 1 Format: SingleEntry 1 Datatype: Each entry is unit representing a person. I 1 MahesSenseFor: AnyPerson 1 DefinedlnTermsOf: (Spouse Mother) I 1 DefinedUsing: Composition 1 ToCompute: (A (x) (GetValue (GetValue x ‘Mother) ‘Spouse) i Figure # 3 - Later form of the Father unit, showing automatically. slots filled in Notice the the ease with which a use; can currently :‘extend his representation”, enlarging his vocabulary of new slots. A similar, though more extravagant example would be to define FavorifeAunf as (SingieMost (Unioning (Composition Sister Parenfs) (Composition Spouse Brofher Parenfs)) $$). Note that “Unioning” and “SingleMost” are two of the slot combiners which come with RLL, whose definition and range can be inferred from this example. It is usually no harder to create a new type of slot format (OrderedNonemptySet), slot combiner (TwoMost, Starring), or datatype (MustBePersonOverl6), than it was to create a new slot type or inheritance mechanism. Explicitly encoding such information helps the user (and us) understand the precise function of each of the various components. We do not yet feel that we have a complete set of any of these components, but are encouraged by empirical results like the following: The first two hundred slots we defined required us to define thirteen slot combiners, yet the lasf two hundred slots required only five new slot combiners. 4.2. EXAMPLE: Creating a New Inheritance Mode Suppose a geneticist wishes to have a type of inheritance which skips every second ancestor. He browses through the hierarchy of units descending from the general one called Inheritance, finds the closest existing unit, InheritSelectively, which he copies into a new unit, InheritFromEvery2ndGeneration. Editing this copy, he finds a high level description of the path taken during the inheritance, wherein he replaces the single occurrence of “Parenf” by “GrandParent” (or by two occurrences of Parenf, or by the phrase (Composition Parenf Parent)). After exiting from the edit, the new type of inheritance will be active; RLL will have translated the slight change in the unit’s high-level description into a multitude of low-level changes. If the geneticist now specifies that Organism # 34 is an “InheritFromEvery2ndGeneration offspring” of Organism#20, this will mean the right thing. Note that the tools used (browser, editor, translator, etc.) are themselves encoded as units in RLL. 4.3. EXAMPLE: Epistemological Status Epistemological Status: To represent the fact that John believes that Mary is 37 years old, RLL adds the ordered pair (SeeUnit AgeOfMaryOOOl) to the the Age slot of the Mary unit. RLL creates a unit called AgeOfMaryOOOl, fills its value* slot with 37 and its EpiSfa/us slot with “John believes”. See Figure 4. Note this mechanism suffices to rcprcscnt belief about belief (just a second chained SeeUnit pointer), quoted belief (“John thinks he knows Mary’s age”, by omitting the value 37 slot in AgeOfMaryOOOl), situational fluents, etc. This mechanism can also be used to represent arbitrary n-ary relations, escaping the associative triple (i.e. Unit/SZof/value) limitation. Name: ] IS-A: Mary I Description: (Person Female ContraryActor) 1 The grower of silver bells etc. I i Age: 1 Name AgeOfMaryOOOl il Name- 1 Isa (UnitForASlotFiller) I I lsa AgeOfMaryOO02 I I LiveslnUnit I I LiveslnUnit (UnitForASlotFiller) I Mary Mary I 1 LiveslnSlot Age- I ~Zs 37 John believes I Teleology: Epistemic . . I 1 LiveslnSlot Age- 1 1 value 1 I I Epistatus gring Wedding805 I 1 I Teleology: Historic I Figure 4 - Representing “John believes that Mary is 37, but she’s When she was married. she was 21”. really 39. 4.4. EXAMPLE: Enforcing Semantics Suppose that Lee, a user of RLL, is constructing HearSayXXIV, a representation language which contains cooperating knowledge sources (KSs). He specifies that each unit representing a knowledge source should have some very precise applicability criteria (he defines a FuZZRelevancy slot) and also a much quicker, rougher check of its potential relevance (he defines a PrePreCondifions slot). If HearSayXXIV users employ these two slots in just the way he intended, they will be rewarded with a very efficiently-running program. But how can Lee be sure that users of HearSayXXIV will use these two new slots the way he intended? He also defines a new kind of slot called Semantics. The unit for each type of slot can have a Semanfics slot, in which case it should contain criteria that the values stored in such slots are expected to satisfy. Lee fills the Semanfics slot of the unit called PrePreConditions with a piece of code that checks that the PrePreConditions slot of every KS unit is filled by a Lisp predicate, which is very quick to execute, which (empirically) correlates highly to the FullRelevancy predicate, and which rarely returns NIL when the latter would return T. This bundle of constraints captures what he “really means” by PrePreCondifions A user of HearSayXXIV, say Bob, now builds and runs a speech understanding program containing a large collection of cooperating knowledge sources. As he does so, statistics are gathered empirically. suppose Bob frequently circumvents the PrePreCondifions slot cntircly, by placing a’ pointer there to the same long, slow, complete criteria he has written for the FuNReZevancy slot of that KS. This is empirically caught as a violation of one of the constraints which Lee recorded in the Semanfics slot of the unit PrePreConditions. As a result, the Semanfics slot of the Semantics unit will be consulted to find an appropriate reaction; the code therein might direct it to print a warning message to Bob: “The PrePreCondifions slot of a KS is meant to run very quickly, compared with the F&Relevancy slot, but 70% of yours don’t; please change your PrePreCondifions slots, 167 or your FullRelevancy slots, or (if you insist) the Semantics slot of the PrePreConditions unit”.6 5. SPECIFICATIONS FOR ANY REPRESENTATION LANGUAGE LANGUAGE The following are some of the core constraints around which RLL was designed. One can issue commands to RLL which effectively “turn off’ some of these features, but in that case the user is left with an inflexible system we would no longer call a representation language language. Further details may be found in [Lena& Haye* Rorh, & Klahr] and in [Geneserelh & Lenaf]. Self-description: No part of the RLL system is opaque; even the primitive Get and Put and Evaluate finctions are represented by individual units describing their operation.2 Current status: complete (to a base language level). Self-modification: Changes in the high-level description of an RLL process automatically result in changes in the Lisp code for -- and hence behavior of -- RLL. Current status: this works for changes in definition, format, etc. of units representing slots and control processes. Much additional effort is required. Codification of Representation Knowledge: Taxonomies of inheritance, function invocation, etc. Tools for manipulating and creating same. These correspond to the stops of the organ, illustrated above. Current status: this is some of the most exciting research we foresee; only a smattering of representation knowledge has yet been captured. 6. INITIAL “ORGAN STOPS” The following characteristics pertain especially to the initirll state of the current RLL system, wherein all “organ stops” are set at their default positions. Each RLL user will doubtless settle upon some different settings, more suited to the representation environment he wishes to be in while constructing his application program. For details, see [Greiner]. Cognitive economy: Decision-making about what intermediate values to cache away, when to recompute values, expectation- f-iltcring. Current status: simple reasoning is done to determine each of these decisions; the hooks for more complex procedures exist, but they have not been used yet. Syntactic vs Semantic slots: Clyde should inherit values for many slots from TypicalElephant, such as Color, Diet, Size; but not from slots which refer to TypicalElephant qua data structure, slots such as NumerOjFilledInSlots and DaleCreated. Current status: RLL correctly treats these two classes of slots differently, e.g. when initializing a new unit. Onion field of languages: RLL contains a collection of features (e.g., automatically adding inverse links) which can be individually enabled or disabled, rather than a strict linear sequence of higher and higher level languages. Thus it is more like an onion field than the standard “skins of an onion” layering. Current status: Done. Three of the most commonly used settings are bundled together as CORLL, ERLL, and BRLL. ’ This work has led us to realize the impossibility of unambiguously stating semantics. Consider the case of the semantics of the Lisp function “OR”. Suppose one person believes it evaluates its arguments left to right until a non- null value is found: a second person believes it evaluates right to left: a third person believes it evaluates all simultaneously. They go to the Semanfics slot of the unit called OR, to settle the question. Therein they find this expression: (OR (Evaluate the args Iefi to right) [Evaluate the args right to leJ)J. Person #3 is convinced now that he was wrong, but persons 1 and 2 point to each other and exclaim in unison “See, I toId you!” The point of all this is that even storing a Lisp predicate in the Semanfics slots only specifies the meaning of a slot up to a set of fixed points. One approaches the description of the semantics with some preconceived ideas, and there may be more than one set of sucpi hypotheses which are consistent with everything stored therein. See [Genesereth & Lenar]. Economy via Appropriate Placement: Each fact, heuristic. comment, etc. is placed on the unit (or set of units) which are as general and abstract as possible. Frequently, new units are created just to facilitiate such appropriate placement. In the long run, this reduces the need for duplication of information. One example of this is the use of of appropriate conceptual units: Clarity of Conceptual Units: RLL can distinguish (i.e. devote a separate unit to each of) the following concepts: TheSetOfAllElephants, (whose associated properties describe this as a set -- such as #OjMembers or SubCaregories), TypicalElephant, (on which we might store Expected-TuskLenglh or DefauKolor slots), ElephantSpecies, (which EvoZvedAsASpecies some 60 million years ago and is CloselyRelaredTo the HippopatamusSpecies,) ElephantConcept, (which QuaZifiesAsA BeastOfBurden and a TuskedPackyderm,) ArchetypicalElephant (which represents an elephant, in the real world which best exemplifies the notion of “Elephant-ness”). It is important for RLL to be able to represent them distinctly, yet still record the relations among them. On the other hand, to facilitate interactions with a human user, RLL can accept a vague term (Elephant) from the user or from another unit, and automatically refine it into a precise term. This is vital, since a term which is regarded as precise today may be regarded as a vague catchall tomorrow. Current status: distinct representations pose no problem; but only an exhaustive solution to the problem of automatic disambiguation has been implemented. 7. CONCLUSION “‘...in Mordor, where the Shadow lies. ” The’system is currently usable, and only through use will direction.* for future effort be revealed. Requests for documentation and access to RLL are encouraged. There are still many areas for tirther development of RLL. Some require merely a large amount of work (e.g., incorporating other researchers’ representational schemes and conventions); others require new ideas (es., handling intensional objects). To provide evidence for our arguments, we should exhibit a large collection of distinct representation languages which were built out of RLL; this we cannot yet do. Several specific applications systems live in (or are proposed to live in) RLL; these include ELI-RISK0 (discoiery of heuristic rules), E&E (combat gamina). FUNNEL (taxonomv of Lisp objects, with an aim toward-autor&tic programming), ROGET (Jim Bennett: guiding a medical expert to directly construct a knowledge based system), VLSI (Mark Stefik and Harold Brown: a foray of AI into the VLSI layout area), and WHEEZE (Jan Clayton and Dave Smith: diagnosis of pulminary function disorders, reported in [Smifh & Clayton]). Experience in AI research has repeatedly shown the need for a flexible and extensible language -- one in which the very vocabulary can be easily and usefully augmented. Our representation language language addresses this challenge. We leave the pieces of a representation in an explicit and modifiable state. By performing simple modifications to these representational parts (using specially-designed manipulation tools), new representation languages can be quickly created, debugged, modified, and combined. This should ultimately obviate the need for dozens of similar yet incompatible representation languages, each usable for but a narrow spectrum of task. 68 ACKNOWLEDGEMENTS The work reported here represents a snapshot of the current state of an on-going research effort conducted at Stanford University. Researchers from SAIL and HPP are examining a variety of issues concerning representational schemes in general, and their construction in particular (viz., [Nii & Aiello]). Mark Stefik and Michael Geneseretb provided frequent insights into many of the underlying issues. We thank Terry Winograd for critiquing a draft of this paper. He, Danny Bobrow, and Rich Fikes conveyed enough of the good and bad aspects of KRL to guide us along the path to RLL. Greg Ilarris implemented an early system which perfomred the task described in Section 4.1. Others who have directly or indirectly influenced this work include Bob Balzer, John Brown, Cordell Green, Johan deKleer, and Rick Hayes-Roth. To sidestep InterLisp’s space limitation, Dave Smith implemented a demand unit swapping package (see [Smith]). The research is supported by NSF Grant #MCS-79-01954 and ONR Contract #NOOO14-80-C-0609. BIBLIOGRAPHY Aikins, Jan, “Prototypes and Production Rules: An Approach to Knowledge Representation from Hypothesis Formation”, HPP Working Paper HPP-79- 10, Computer Science Dept., Stanford University, July 1979. Bobrow, D.G. and Winograd, T., “An Overview of KRL; a Knowledge Representation Language”, SIJCAI, August 1977. a Brachman, Ron, “What’s in a Concept, Structural Foundations for Semantic Networks”. BBN report 3433, October 1976. Findler, Nicholas V. (ed.). Associafive Networks. NY. hcadcmic Press, 1979. Gencsereth, Michael, and Lenat, Douglas B.. SeIfDescripfion and -Modijicnrion in a Knowledge Representation Language, HPP Working Paper HPP-80-10, June, 1980. Greiner. Russell. “A Representation Language Language”, HPP Working Paper HPP-80-9, Computer Science Dept., Stanford University, June 1980. Lenat, Douglas B.. “AM: Automated Discovery in Mathematics”, SIJCAI, August 1977 Lenat, D. B., Hayes-Roth, F. and Klahr. P., “Cognitive Economy”, Stanford HPP Report HPP-79-15. Computer Science Dept., Stanford University, June 1979. Minsky, Marvin, “A Framework for Representing Knowledge”, in 77re I’s#rofogv OJ Computer Vision. P. Winston (ed.), McGraw-Ilill, New York, 1975. Nii, II. Penny, and Aiello, N., “AGE (Attempt to Generalize): A Knowledge- Based Program for Building Knowledge-Based Program”, 6IJCAI. August 1979. SIGART Newsletter, February 1980 (Special Representation Issue; Brachman & Smith, eds.). Smith. David and Clavton. Jan. “A Frame-based Production System Architecture”, AAAI Conference, .1980. Smith, David, “CORLL: A Demand Paging System for Units”, HPP Working Paper HPP-80-8, June 1980. Stefik, Mark J.. “An Examination of a Frame-Structured Representation System”, SIJCAI, August 1977. Szolovits, Peter, Hawkinson, Lowell B., and Martin, William A., “An Overview of OWL, A Language for Knowledge Representation”, MITLLCS/TM-86, Massachusetts Institute of Technology, June 1977. Woods, W. A., “What’s in a Link, Foundations for Semantic Networks”, in D. G. Bobrow & A. M. Collins (eds.), Representation and Understanding, Academic Press, 1975. , 169	1980	43
38	SPATIAL AND QUALITATIVE ASPECTS OF REASONING ABOUT MOTION Kenneth D. Forbus MIT Artificial Intclligcncc Laboratory 545 Technology Square Cambridge, Mass. 02139 ABSTRACT Reasoning about motion is an important part of common sense knowledge. The spatial and qualitative aspects of reasoning about motion through fret space arc studied through the construction of a program to perform such reasoning. An analog gcomctry rcprcscntation serves as a diagram, and descriptions of both the actual motion of a ball and envisioning are used in answering simple questions. 1 Introduction People reason fluently about motion through space. For example, WC know that if two balls arc thrown into a well they might collide, but if one ball is always outside and the other always inside they cannot. The knowl”cdgc involved in this qualitative kind of reasoning seems to be simpler than formal mechanics and appears to be based on our expcricnce in the physical world. Since this knowledge is an important part of our common SCIISC, capturing it will help us to understand how pcoplc think and enable us to make machines smarter. The issues involved in reasoning about motion through space wcrc studied by constructing a program, called FROB, that reasons about motion in a simple domain. 1 believe three important ideas have been illustrated by this work: 1. A quantitative geometric rcprescntation simplifies reasoning about space. It can provide a simple method for answering a class of gcomctric questions. The descriptions of space required for qualitative reasoning can bc defined using the quantitative representation, making it a communication device between several rcprcsentations of a situation 2. Describing the actual motion of an object can be thought of as creating a network of descriptions of qualitatively distinct types of motion, linked by descriptions of the state of the object before and after each of these motions. This network can bc used to analyze the motion and in some cases can be constructed by a process of simulation. 3. The description of the kinds of motion possible from some state (called the cnvisionment) is useful for answering certain questions and for checking an actual motion against assumptions made about it. The assimilation of assumptions about global properties of the motion into this description makes heavy use of spatial descriptions. FRO11 reasons about motion in a simplified domain called the “Bouncing Ball” world. A situation in the Bouncing IjJl world consists of a two dimensional scent with surfaces represented by line scgmcnts, and one or more balls which are mod&d as point masses. WC ignore the exact shape of balls, motion after two balls collide, spin, motion in a third spatial dimension, air resistance, sliding motion, and all forces other than gravity. The initial description of a situation is a diagram containing a description of the surfaces and one or more bails, as in ligure 1. Fig. 1. A typical sccnc from the Rouncing Ball world A silualion in the Bouncing Ml World consisls or a diagram that q.x~fies surfaces and one or more balls This drawing only shows lhe geometric aspects of the descriptions involved. When given a description of a situation in the Ijouncing hall world, I-ROB analyzes the surface geometry and computes qualitative descriptions of the free space in the diagram. ‘l’hc person using the program can describe bails. properties of their states of motion, rcqucst simulations, and rnakc global assumptions about the motion. FROl? incrcmcntally crcatcs and updates its descriptions to accommodate this information, complaining if inconsistencies are dctcctcd. Questions may bc asked by calling procedures that intcrrogatc thcsc descriptions. The four basic questions FROR can answer are: (1) What can it (a ball) do next?, (2) Where can it go next?, (3) Where can it end up?, and (4) Can thcsc two balls collide? ,Ll Spatial descriptions WC do not yet know why people are so good at reasoning about space. Thcorcm proving and symbolic manipulation of algebraic expressions do not seem to account for this ability. Arguments against the former mdy bc found in [l], while the sheer complexity of algebraic manipulations argues against the latter. I conjccturc that the fluency pcoplc exhibit in dealing with space comes mainly from using their visual apparatus. One example is the 170 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. USC of diagrams. Thcmarks in a diagram reflect the spatial relations bctwcen the things they rcprcscnt. which allows us to use our visual apparatus to interpret these relationships as WC would with real objects. In this case, perception provides a simple (at least for the processes that use it) decision procedure for a class of spatial questions. We do not yet understand the complcxitics of human vision, but the techniques of analytic gcomctry can be used to provide decision proccdurcs for geometrically simple cases. FROR uses a Metric Diagram. which is a representation of geometry that combines symbolic /and numerical information. The geometric aspects of a problem are rcprcscntcd by symbolic elements whose parameters are numbers in ;I bounded global coordinate system. The rcprescntation is used to answer questions about basic spatial relationships between elements, such as which side of a lint a particular point lies or whcthcr or not two lines touch. Calculation based on properties of the clcmen ts su ffkes to answer these questions. My conjecture about qualitative spatial reasoning is that it involves a vocabulary of PLACES whose relationships are described in symbolic terms. By PLACE, I mcan a piece of space (point, line, region, volume, etc.) such that ail parts of it share some property. The nature of a domain determines the notion of place appropriate to it. In FROII the Space Graoh provides the vocabulary of places. Since all balls arc point masses and arc subject to the same forces, the Space Graph is independent of them and dcpcnds only on the surface geometry in the Metric Diagram. Free space is divided into regions in a way that insures the description of qualitative state (dcscribcd below) will bc unique and simple, and these regions and the edges that bound them are the n_odcs of the Space Graph. These nodes are connected by arcs that are labclled with the name of the relationship between them (such as IXFT or UP). Any other place rcquircd for qualitative reasoning can be dcscribcd by composing these nodes, and the graph structure provides a framework for cfficicnt processing (see section 4). An example of the places in a scene and the graph structure they produce is contained in figure 2. Pointu- Fig. 2. Space Graph for a scene The free space in the diagram is broken up into regions in a way that simplifies the description of the kinds of motion possible. The labels on the pointers which indicate the spatial relationships between the nodes are not shown due to lack of space. III Describing a Particular Motion When we watch an object move, we generally couch OUT description in terms of a sequence of qualitatively distinct motion types. WC will call a network built from descriptions of motions linked by descriptions of the state of the object before and after each motion an Action SCQUCI~. ‘I’he knowledge associated with each type of motion allows it to IX further analy?cd, the consistency of the proposed description to be checked, and permits making predictions about what will happen next. A drawn trajectory of motion in the Bouncing Ml domain and the schema of its associated Action Sequence is illustrated in figure 3. Fig. 3. Action Sequence Schema for Bouncing Balls This schema describes the motion depicted in Figure 1 The PilYSOR constrarnt dcscribcs Lhc state of Ihc ball al some instant in time. and the ACT constraints dcscrlbe a piece of the ball’s history. l’hc two basic types of motion in the Bouncing Hall world are FLY and COLI.II)F.. The difference in computing boundary conditions between flying up and flying down rcquircs their consideration as separate acts in the sequence, and additional motion types are defined for transformations to motions outside the Bouncing I\all world (such as CONTINUE for leaving the diagram and SI,IDE/S’I‘OP when a ball is travclling along a surface). The description of a ball’s state includes such information as its velocity (quantitative if known, or just in terms of a rough heading like (LEFT UP)) and what it is touching. In FROB the Action Sequence descriptions are embedded in a constraint language (SW [2] for an overview), and include equations describing projectile motion to compute numerical values if numerical descriptions of the state parameters are obtained. The use of quantitative parameters in the qualitative description of motion makes possible a diffcrcnt kind of simulation from the usual incremental time simulations used in physics. When numbers are provided, an Action Sequence can bc produced by gcncrating a description of the next motion from the last known state of motion. The time to gcncrate the description, as well as the complexity of the result, depends on the qualitative complexity of the motion rather than some fixed increment of time used to evolve a set of state parameters. Simulation is not the only way an Action Sequence can be created. A network of constraints can bc built to describe some proposed motion and the knowlcdgc of the equations of motion can be used to analyze it to see if it is consistent. ‘The dependence on quantitative paramctcrs in FliOl~s analysis is a drawback. For example, FROB can detect that the situation -in figure 4 is 171 inconsistent only after being given some final height for the ball and a value for the elasticity. People can argue that this proposed motion is impossible with simpler arguments that require less information. Fig. 5. A Sequence Graph The arrows represent the Arcction of a quahtative state at the place the arrow is drawn Circles represent states without well defined direclions. The pomters expressing the possible temporal orderings of the states are not shown. Fig. 4. An inconsistent description of motion This motion is impossible hecause the ball could not get as high as it does after the second collision unless it had gone higher on the tirst. If it had gone higher after the first, the second collision would not even have happened ‘To discover that this description is inconsistent l:KOl3 requires a specific re!ocity a( the highest point and a specific value for the elasticity of the ball as well as the coordinates of the collision points. The basic idea of an Action Sequence stems highly suited as a target representation for parsing quantitative data about motion, perhaps gleaned by perception. For this purpose a more qualitative set of methods for analysis would havP to bc encoded. An example of such a rule for the Bouncing Ball domain would be “A ball cannot increase its energy from one act to the next”. E Describing Possible Motions The quantitative state of a ball consists of its position and velocity. A notion of qualitative state can be defined which generalizes position to be a PLACE, the velocity to be a symbolic heading (such as (RIGHT DOWN)), and makes explicit the type of motion that occurs. A set of simulation rules can be written to operate on qualitative state descriptions, but because of the ambiguity in the description the rules may yield several motions possible from any given state. Since there arc only a small number of places and a small number of possible qualitative states at each place, all the possible kinds of motion from some given initial qualitative state can easily bc computed. This description is called the cnvisionmcnf (after [3]) for that state. dcKleer used this description to answer simple questions about motion directly and plan algebraic solutions to physics problems. In FROB envisioning results in the Scclucncc Graph, which uses the Space Graph for its spatial framework (see Figure 5). It is ----lnll&lstate- (FLY (SREGIONl) (LEFT Knowing more about a ball than its state of motion at some time can restrict the kinds of motion possible to it. Energy limits the height a ball can reach, and knowing that a ball is pcrfcctly elastic or completely inelastic excludes certain results of a collision. Assumptions about whether a ball must or may not reach a particular place or qualitative state can restrict the possibilties as well. The Scqucncc Graph can bc modified by pruning states to reflect this information about the ball and its motion. of Fach of the constraints above directly rules out some states mot&n. The full consequences of eliminating such states are determined by methods that rely on specific properties of space and motion. Among these properties arc the fact that a mo:ion of an object must be “continuous” in its state path (which means that the active part of a Sequence Graph must be a single connected component) and that the space it moves in must bc connected (which is uscfi11 because there are many fcwcr places than qualitative states in any problem). Dependency information is stored so that the cffccts of specific assumptions may be traced. Conflicting assumptions, overconstraint. and conflicts bctwcen a description of the actual motion (as spccificd by an Action Sequence) and its constrained possibilties arc dctcctcd by FROB and the underlying assumptions arc offcrcd up for inspection and possible correction. v Answering Oucstions Many of the questions that could be asked of the Bouncing Ball domain can bc answered by direct examination of the descriptions built by FROB. These include questions (1) and (2) above. The three levels of description of motion in FROB (the Action Sequence, the Sequcncc Graph, and the path of qualitative states corresponding to the Action Sequence) allow some kind of answer to be given even with partial information. used for summarizing propcrtics of the long term motion of an object, evaluating collision possibilties, and,assimilating assumptions More complicated questions (such as (3) and (4) above) can about the global propcrtics of motion. Only the assimilation of be answer-cd with additional computation using these descriptions. Determining whcthcr or not a ball is trapped in a well (see figure 6) assumptions will be discussed here. can be done by examining a Scqucnce Graph for the last state in an 172 Fig. 6. Summarizing Motion ->>(motion-summary-for bl) FOR Bl TlIE BALL WILI. I:VI:NTUAI 1;Y STOP II‘ IS TRAPPED INSIDE (WELID) AND IT WILL S I’OP FLYING IZT ONE OF (SEGMENTll) Action Sequence to see if it is possible to bc moving outside the places that comprise the well. Often a collision between two balls can be ruled out because the two balls arc never in the same PI.ACE, as determined by examining their Scqucncc Graphs. With the Action Sequence description of motion it is possible to compute exactly where and when two balls collide if they do at all. Figure 7 contains the answers given by the progratn to collision questions in a simple situation. Fig. 7. Collision Problems ->>(collide? f g) (I’OSSIRI 1:. AT SEGMEN l-13 SREGlONl .,,) ->>(rannot-bc-at f segment31) (SIiGhlENT31) Ul’Dh I‘ING ASSUMPT10NS FOR (>> INlTlAl,-STATE 17) ClII:CKING l’ATli 01:MOIION ACAINSTASSUMPTIONS ->>(collide? f g) NO VI liclation to Other Work The focus of this work is very different from that of [4]!5][6], which arc mainly concerned with modcling students solving textbook physics problems. All of the problems dealt with in these programs were static. and the rcprcscntation of gcomctry expressed connectivity rather than free sp~c. lssucs such as gerting algebraic solutions and doing only the minimal amount of work rcquircd to answer a particular question about a situation were ignored here in order to bcttcr deal with the questions of spatial reasoning and the semantics of motion. The process of formalizing common sense knowledge is much in the spirit of the Naive Physics effort of Hayes (described in [7]). The Action Sequence, for example, may bc vicwcd as the history for a ball since it contains explicit spatial and temporal limits. However, this work is conccrncd with computational issues as well as issues of rcprcscntation. Unlike this work, Hayes (see [8]) explicitly avoids the use of metric rcprcscntations for space. 1 suspect that a metric representation will bc required to make the concept of a history useful, in that to compare lhcm rcquircs having a common coordinate frame. The Metric Dingrdm has much in common with the dcrcriptions used as targets for langu,tgc translation of [I] and the imagery theory of [9]. Arguments against the traditional “pure relational” gcomctric representations used in A1 and the “naive analog” rcprcscntations used by [lc)],[l l] may bc found in [12]. The concept of envisioning was first introduced in [3] as a technique for answering simple questions about a sccnc directly and as a planning device for algebraic solutions. ‘I’hc inclusion of dissipative forces, a true IWO dimensional domain, interactions of more than one moving object, and its USC in assimilation of global constraints on motion arc the envisioning advances incorporated in this work. vu Biblioprauhy [I] Waltz, D. and Boggess, L. “Visual Analog Representations for Natural Language Understanding” in Proc. IJCAI-79 Tokyo, Japan, August 1979 -- 121 Steele, G and Sussman. G “Constraints” Memo No. 502. MIT AI Iab. Cambridge, Massachusetts, November 1978 [3] deKleer, Johan “Qualitative and Quantitative Knowledge in Classical Mechanics” Technical Report 352. MIT AI lab, Cambridge, Massachusetts. 1975 [4) Bundy, A. et. al “MECllO:Year One” Research Report No. 22, Department of Artificial Intelligence. FAinburgh, 1976 (51 Novak, G “Computer Understanding of Physics Problems Stated in Natural Language” Technical Report NL-30. Computer Science Department, The University of Texas at Austin, 1976 [6] McDermott, J and Larkin, J. “Re-representing Textbook Physics Problems” in Proc of the 2nd National Conference of the Canadian Society for Computational Studies of Intelhgence, Toronto 1978 (71 llayes, Patrick J “The Naive Physics Manifesto” unpublished, May 1978 [8] Hayes, Patnck J “Narve Physln 1:Ontology for Liquids” unpublished, August 1978 191 Hinton. G “Some Demonstrations of the Effects of Structural kcription~ in Mental Imagery” Conmtive Science, Vol 3. No. 3, July-September 1979 [JO] Funt, B V. “Wl1ISPER:A Computer implementation Using Analogues in Reasoning” PI1.D. Thesis, Umvcrslty of British Columbia, 1976 [ll] Kosslyn & Schwartz, “A Slmuiation of Visual Imagery” Connitive Science. Vol. 1, No. 3, July 1977 [12] Forbus. K. “A Study of @ualitative and Geometric Knowledge in Reasoning about Motion” MIT Al Lab T~IJJJC~ report, in preparation. 173	1980	44
39	COMPUTER INTERPRETATION OF HUMAN STICK FIGURES Martin Herman Department of Computer Science Carnegie-Mellon University Pittsburgh, PA 15213 ABSTRACT A computer program which generates context-sensitive descriptions of human stick figures is described. Three categories of knowledge important for the task are discussed: (1) the 3-D description of the figures, (2) the conceptual description of the scene, and (3) heuristic rules used to generate the above two descriptions. The program’s representation for these descriptions is also discussed. 1. Introduction This paper describes a computer program, called SKELETUN, which generates context-sensitive descriptions of 2-D, static, human stick figures. The motivating interest is to study the process of extracting information communicated by body postures. Stick figures have been chosen to approximate the human form because they eliminate the problems involved in processing fleshed-out human figures (e.g., extracting them from the image, identifying and labeling body parts), yet they maintain the overall form conveyed by gross body posture. SKELETUN currently operates in two domains, emotions (figures may be sad, happy, depressed, etc.) and baseball (batting, catching, etc.) Its knowledge of baseball is much more complete, however. It can accept any figure and interpret it in terms of the following baseball activi ties: (1) batting, (2) throwing, (3) running, (4) catching a high ball with one or both arms, (5) catching a ball at torso height with one or both arms, (6) fielding a grounder with one or both arms. An example of how a figure is interpreted in the baseball domain is shown in Fig. 1, where hand-generated English Q 1. Two arms catching torso-high ball 2. (very good confidence) Batting (fair confidence) 3. Two arms fielding grounder (poor confidence) Fig. la PHYSICAL DESCRIPTION The figure is in a vertical orientation with the feet below the head. The figure is facing left and the face is pointing left. The torso is bent forward. The elbow of arm1 is in- middle and down. It can be considered either as partly or half bent. The elbow of arm2 is in- middle and down. It can be considered either as partly or half bent. The knee of leg1 is forward and partly bent. The knee of leg2 is down and partly bent. The lower body is in a configuration similar to “feet well planted.” The figure can also be consider- ed in a diagonal orientation with the feet to the lower right of the head (but with lower confidence than vertical). In this case, it is facing lower left with the face pointing low- er left, The following then changes from the previous description: the elbow of arm1 can be considered as either down or forward. The knee of leg2 is forward. MEANING-BASED DESCRIPTION The figure is catching a ball at torso height with two arms, with very good confidence. It may also be viewed as batting, but with only fair confidence. Finally, it may be fielding a grounder with two arms, but with only poor confidence. Fig. lb ((vertical orientation) verygood) ((feet to bottom of head) verygood) ((facing left) good) ((face pointing left) good) ((torso bent forward) good) ((elbow1 is in-middle) good) ((elbow1 is down) verygood) ((elbow1 partly bent) good) ((elbow1 half bent) good) ((elbow2 is in-middle) good) ((elbow2 is down) verygood) ((elbow2 partly bent) good) ((elbow2 half bent) good) ((kneel is forward) verygooc ((kneel partly bent) good) ((knee2 is down) verygood) ((knee2 partly bent) good) ((both legs “feet well planted” cfig) good) ((diagonal orientation) good) ((feet to lowerright of head) good) ((facing lowerleft) fair) ((face pointing lowerleft) fair) ((elbow1 is down) good) ((elbow1 is forward) good) ((knee2 is forward) verygood) ((two-arms-catching-torso high-ball) verygood) ((batting) fair) ((two-arms-fielding- grounder) poor) 174 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. descripticns are shown alongside the computer-generated output. SKELETUN’s primary purpose is to generate a description of what is communicated by body posture - the “meaning-based” description. In the process of generating this description, it also provides the 3-D configuration of the figures - the physical descripticn. Briefly, the notation in the example is as follows. If a figure is viewed from the front or back, each elbow or knee can be either QI& from the torso, b to the torso (i.e., crossing the torso) or If the figure is in-middle (i.e., along the same line as the torso). viewed from the side, each elbow or knee can be either u, forward. backward, or back-uo (i.e., backward and up). All assertions in the dessriptlons have discrete confidence values. The input to SKELETUN is a hand-encoding of the x, y coordtnates of the end points of the line segments of each figure, plus the center of the circle representing the head. SKELETUN assumes that all figures are complete and valid, and that no objects other than figures are in the scene (a scene may have two figures). This paper gives an overview of the types of information conveyed by gross body postures, SKELETUN’s representation for this information, and some inference rules used to generate this information from 2-D scenes. See [7] for details. 1 .l Backaround This work views vision as a medium of communication, recognizing that an important goal of the visual process is to provide the viewer with a “meaning” description of the external world. Most scene analysis systems are primarily concerned with identifying objects and other entities in a scene and specifying the spatial configuration of these entities [6, 3, 8, 41. Given a scene with human figures, such systems would tend to identify the individual figures, their body parts, and other objects, and then specify the spatial relationships of these entities [9, 10, 11. SKELETUN goes one step further in the interpretation process. It tries to determine what the people are doing, and perhaps why they are doing it. Although some previous work has taken the point of view of vision as communication [2, 14, 151, their primary purpose was to analyze and describe motion scenes, rather than to study how body posture conveys information. 2. Knowledqe cateqories Five categories of knowledge have been identified as important in the process of generating descriptions of 2-D scenes of stick figures. The first three represent important levels at which the scene should be described. o Two-Dimensional Descriotion - a low-level description involving the direction of each body part (each is a straight line segment), the angle of each joint (in the 2-D plane), and body parts which overlap (required for establishing touching relationships). e Phvsical Soace Descriotion - a 3-D description of the physical configurations of the figures. o Meanina Soace Descriotion - a description in terms of the information communicated by the figures (e.g., running, fighting, crying). The concepts here are said to be in Meaning Space since “meaning” (or “conceptual” information) is extracted from the scene. The next two categories involve knowledge used to extract the physical and meaning space descriptions from the 2-D description. @ Human Phvsical Structure - information dealing with the various parts of the stick figure body and components of these parts. Q Inference Rules - heuristic rules used to obtain the -- 3-D configuration of the figures from the 2-D scene, and to determine what the figures are doing based on the 2-D and 3-D configurations of the limbs. The following sections will further discuss the 2nd, 3rd, and 5th categories. More details than can be provided here on all of the categories may be found in [7]. 3. Phvsical Soace Descriotion In order to infer what is being communicated by a figure’s body posture, there must be knowledge of at least part of its 3-D configuration, for it is a 2-D figure interpreted as being in 3-D space to which meaning is applied. It is convenient to have two different levels of physical space descriptions. One, called the lower level phvsical soace descrintion, deals with the 3-D positions of the individual body parts. The second, called the hiaher level phvsical soace clescriotion, deals with frequently occurring positions of groups of body parts. Only the first description will be discussed in this paper (see [7] for more details). Although a figure’s 3-D configuration may be represented many ways, the representation to be described next was chosen for two reasons: 1. Its purpose is to describe the figure in a manner useful for generating meaning-based interpretations. If the resolution is too fine (as in [lo]), it will contain much information not significant for the task, thus burdening the system. If the resolution is too coarse, it will not contain enough information to perform the task. 2. It is convenient for SKELETUN to be able to express a figure’s 3-D configuration in a manner easily understood by humans. The current representation makes this kind of information explicit. 3.1 Descriotions relative to the torso The 3-D descriptions in SKELETUN are object-centered, as opposed to viewer-centered. That is, locations and directions of parts of the figure are indicated with respect to the figure, rather than the viewer. A viewer-centered description depends not only on the figure being described, but also on its orientation. An object-centered description, however, depends only on the figure being described, resulting in a smaller set of possible descriptions [111. Accordingly, the positions of the upper arms and legs are represented relative to the torso, and the shape of the torso is represented relative to the overall orientation of the figure. SKELETUN uses the predicates OUT, IN-MIDDLE, and IN to describe the position of each elbow or knee as viewed from the 175 front, and UP, FORWARD, DOWN, BACKWARD, and BACK-UP to describe the positions as viewed from the side. These predicates are adequate to completely specify (within the resolution of the representation) the 3-D position of any elbow or knee (i.e., upper arm or leg). SKELETUN uses the predicates BENT-FORWARD and BENT-BACKWARD to specify how the torso joints are bent. 3.2 Hierarchv of obiect-centered c&.criotions The positions of the lower arms and legs are represented relative to the upper arms and legs, respectively. Note that a representation of the lower limbs relative to the torso would result in a much larger set of possible descriptions than a representation relative to the upper limbs, since a different description of the lower limb would be required for each position of the upper limb relative to the torso, even if the position of the lower relative to the upper limb were to remain constant. Since similar arguments apply to describing positions of other body parts, such as hands, fingers, feet, etc., we conclude that each body part should be represented relative to the part it is connected to, resulting in a hierarchy of descriptions [lo]. SKELETUN represents the positions of the lower arms and legs by specifying the 3-D angle of the elbow and knee joints. The predicates used are PARTLY-BENT, HALF-BENT, FULLY-BENT, and NOT- BENT. 3.3 Orientation relative to viewer --- Thus far, all descriptions have been relative to parts of the figure. The whole figure must also be placed in 3-D space, relative to the viewer. The predicate ORIENTATION describes the overall orientation of the figure either as vertical, horizontal, or diagonal. Given one of these orientations, the predicate DIR-OF-FEET-TO-HEAD specifies the direction of the feet relative to the head. Finally, the predicates DIR-FACING and DIR-FACE-IS-POINTING specify the direction the figure is facing and the direction the face is pointing. 3.4 Phvsical space inference rules These rules generate the physical space description. They are domain-independent, for they depend only on the 3-D configuration of the figures. As an example of the knowledge in these rules, consider how SKELETUN determines the overall orientation of the figure. A figure is horizontal if both feet are east or west of the head (as in lying). A figure is diagonal if both feet are southeast, southwest, northeast, or northwest of the head. There are two types of vertical orientations, upright and upside-down. (SKELETUN currently cannot handle upside-down figures.) Fig. 2 shows three extremes of upright figures. In Fig. 2a, both feet are south of the head. In Fig. 2b, both feet are not south of the head; the point midway between the feet is south of the head. In Fig. 2c, the midway point is not south of the head; only one foot is south of the head. Rules which determine whether a figure is upright must examine these three types of cases. For more details on these and other inference rules, see [7]. 4. Meaninq soace descriotion 4.1 Reoresentation Meaning space concepts in SKELETUN are not represented explicitly in terms of simpler concepts and relationships between them (as in Conceptual Dependency [13]), since SKELETUN’s (a) (5) Cc) Three upright stick figures. 2 Fig. concern is not to extract all the details of each concept. Instead, they are represented as labels (RUNNING, CRYING,WALKING, etc.), where the meaning is represented implicitly in terms of the inference rules which may assert the concept and those which may use the concept to assert other concepts. This is because SKELETUN’s concern is to discover and make use of relationships among concepts [12]. Two important classes cf information that can be extracted from the body postures of stick figures deal with (1) the physical states of the figures (running, walking, throwing, standing, etc.) and (2) the mental or emotional states of the figures (weeping, happy, thinking, etc.). Two types of physical states can be distinguished, active and passive. Active physical states involve activities requiring motion, such as running, dancing, or hitting. Passive physical states involve no motion; examples are standing, pointing, and watching. Mental-emotional states can also be divided into two categories, negative and positive. Negative states generally involve feelings or tendencies such as painful excitement, destruction, dullness, loneliness, discomfort, tension, incompetence, dissatisfaction, and helplessness (e.g., anger, sadness, apathy, panic, hate, grief, disgust). Positive states generally involve feelings or tendencies such as vitality, empathy toward others, comfort, and self-confidence (e.g., cheerfulness, enjoyment, happiness, hope, love, pride) [5]. The negative and positive states can each be further subdivided into passive and active. These will not be pursued here (see [7]). 4.2 Meaninq soace inference rules These rules generate the meaning space description. They tend to be domain-dependent, since most meaning-space concepts are applicable only in limited domains. As an example of the knowledge in these rules, consider how SKELETUN determines that a figure is fielding a grounder (assuming that the domain is baseball). (See Fig. 3 for exampies.) First, one or both arms must be in a “fielding grounder” configuration (a higher level physical configuration described in [7]). In addition, the lower body should be in a configuration similar to “kneeling on one knee” (Fig. 3b), “kneeling on both knees”, “feet well planted” (Fig. 3c), or “crouching” (Fig. 3d) [7] and the figure should be vertical. If the figure’s orientation is diagonal, its lower body should be in a “crouching” configuration and it must be facing either lower-left or lower-right. Finally, if both arms are in a 176 “fielding grounder” configuration and the figure is running, it is also fielding a grounder, i.e., running after a ground ball (Fig. 3a). Acknowledoement This research is part of the author’s Ph.D. thesis done at the University of Maryland, under the guidance of Chuck Rieger and Azriel Rosenfeld. The support of the National Science Foundation under Grant MCS-76-23763 is gratefully acknowledged, as is Mike Shneier for valuable comments, and Ernie Harris for help in preparing this paper. > (a) (b) Cd) Each figure is fielding a grounder. Fig. 3 References 1. Adler, M. Computer interpretation of Peanuts cartoons. Proc. SIJCAI, Cambridge, MA, 1977. 2. Badler, N. I. Temporal scene analysis: conceptual descriptions of object movements. Tech. Rept. 80, Dept. of Computer Science, University of Toronto, 1975. 3. Bajcsy, R., and Joshi, A. K. A partially ordered world model and natural outdoor scenes. In Computer Vision Systems, Hanson and Riseman, Ed.,Academic Press, 1978. 4. Barrow, H. G., and Tenenbaum, J. M. MSYS: A system for reasoning about scenes. Artificial Intelligence Center Technical Note 121, Stanford Research Institute, 1976. 5. Davitz, 1969. J. R. The Language of Emotion. Academic Press, 7. Herman, M. Understanding body postures of human stick figures. Tech. Rept. 836, Computer Science Center, University of Maryland, College Park, MD, 1979. 8. Levine, M. D. A knowledge-based computer vision system. In Computer Vision Systems, Hanson and Riseman, Ed.,Academic Press, 1978. 9. Marr, D., and Nishihara, H. K. Spatial disposition of axes in a generalized cylinder representation of objects that do not encompass the viewer. AIM 341, MIT, 1975. 10. Marr, D., and Nishihara, H. K. Representation and recognition of the spatial organization of three-dimensional shapes. AIM 416, MIT, 1977. 11. Nishihara, H. K. Intensity, visible surface, and volumetric representations. Workshop on the Representation of Three-Dimensional Oblects, Univ. of Pennsylvania, Philadelphia, PA, 1979. 12. Rieger, C. Five aspects of a full-scale story comprehension model. In Associative Networks: The Representation and Use of Know/edge in Computers, N. Findler, Ed.,Academic Press, 1978. 13. Schank, R. C. Identification of conceptualizations underlying natural language. In Computer Models of Thought and Language, Schank and Colby, Ed.,W. H. Freeman and Co., 1973. 14. Tsuji, S., Morizono, A., and Kuroda, S. Understanding a simple cartoon film by a computer vision system. Proc. HJCAI, Cambridge, MA, 1977. 15. Weir, S. The perception of motion: actions, motives, and feelings. Progress in Perception Research Report No. 13, Dept. of Artificial Intelligence, University of Edinburgh, 1975. 6. Hanson, A. R., and Riseman, E. M. VISIONS: a computer system for interpreting scenes. In Computer Vision Systems, Hanson and Riseman, Ed.,Academic Press, 1978. 177	1980	45
40	RESEARCH ON EXPERT PROBLEM SOLVING IN PHYSICS Gordon S. Novak Jr. and Agustin A. Araya Computer Science Department University of Texas at Austin Austin, Texas 78712 ABSTRACT Physics problems cannot in general be solved by methods of deductive search in which the laws of physics are stated as axioms. In solving a real physics problem, it is necessary to treat the problem as a "nearly decomposable system" and to design a method of analysis which accounts for the salient factors in the problem while ignoring insignificant factors. The analysis method which is chosen will depend not only on the objects in the problem and their interactions, but also on the context, the accuracy needed, the factors which are known, the factors which are desired, and the magnitudes of certain quantities. Expert problem solvers are able to recognize many frequently occurring problem types and use analysis methods which solve such problems efficiently. Methods by which a program might learn such expertise through practice are discussed. I INTRODUCTION We are investigating the cognitive processes and knowledge structures needed for expert-level problem solving in physics. We believe that physics is a particularly fruitful area for the investigation of general issues of problem solving, for several reasons. The laws of physics are well formalized, and there is a large set of textbook physics problems (often with answers and example solutions) available for analysis and for testing a problem solving program. The application of physical laws is considered to be well defined, so that both the method of analysis of a problem and the final answer can be judged as either correct or incorrect. At the same time, physics is considered to be a difficult subject; students find problem solving especially difficult, even when the equations which express the physical laws are available for reference. Although the laws of physics are "well known", nobody has yet produced a program which can approach expert-level problem solving in physics. Such a program would potentially have great value for applications, since the types of reasoning used in computer science and engineering are closely related to those used in solving physics problems. Such a -------------------- * This research was supported by NSF award No. SED-7912803 in the Joint National Institute of Education - National Science Foundation Program of Research on Cognitive Processes and the Structure of Knowledge in Science and Mathematics. PB SCHWM PRGE 25 NUMBER 19 [THE FOOT OF R LRDDER RESTS RCAINST fl VERTICRL WFlLL RN0 ON fl HORIZONTAL FLOOR) (THE TOP OF THE LROOER IS SUPPORTEO FROM THE WLL BY 17 HORIZONTRL ROPE 30 FT LONG) (THE LfXKlER IS SO FT LONG . WEIGHS 100 LB WITH ITS CENTER OF GRAVITY 20 FT FROM THE FOOT . f4NO F1 150 L8 MAN IS 10 FT FROM THE TOPl[DETERMINE THE TENSION IN THE ROPE1 FINSWER : 120.00000 LB Figure 1: Example of Output of ISAAC Program program could also be of value in education, because many of the crucial skills used in solving physics problems are now taught only implicitly, by example; students who are unable to infer the skills from the examples do poorly in physics. The first author has previously written a program which can solve physics problems stated in English in the limited area of rigid body statics [1,21; an example of its output is shown in Figure 1. This program, which uses a general formulation of the laws of rigid body statics (similar to the form of the laws presented in modern textbooks), produces between nine and fifteen equations for simple textbook problems for which human problem solvers generate only one or two equations. This somewhat surprising result indicates that the expert human problem solver does not slavishly apply the general forms of physical laws as taught in textbooks, but instead recognizes that a 178 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. particular problem can be solved by applying a special case of the more general law and writes only the equations appropriate for the special case. By doing so, the expert greatly reduces the algebraic complexity of the problem. II THE NATURE OF PROBLEM SOLVING IN PHYSICS -__I__----- Most people, even experts, tend to identify the "content" of physics with the equations which express the physical laws. Bundy [3] has written PROLOG programs in which the laws of physics are expressed as Horn clauses and deductive search is used to find answers to problems posed as sets of predicates. As noted by Larkin, McDermott, Simon, and Simon r41, novice problem solvers do tend to use an "equation-driven search", working backwards from the desired quantity until the equations they have invoked can be solved; experts, however, usually work "forward" from the given information until the desired unknown has been found. Experts do not often verbalize the equations that they are using, but usually only verbalize intermediate "answers". The identification of physics knowledge with the equations which express physical laws and the notion that search is the primary mechanism used in solving physics problems are unsatisfying because they fail to account for several observed phenomena. Why is physics hard? If physics were "only" the equations for the laws, these equations could be collected in a reference book (along with tables of integrals and such), and the physics course could be dispensed with. Why does practice help? What is expertise in physics, i.e., what does the expert have that the novice lacks that enables the expert to perform so much better? Superior algebraic skills alone cannot account for the difference. What is the intellectual content of a physics course? Without continued practice, students forget the equations soon after taking the course; what is it that they retain that makes taking the course worthwhile? We believe that methods which employ deductive search and express the laws of physics directly as predicate calculus clauses (01: the equivalent) cannot account for expert-level problem solving ability when a variety of physical principles are involved (say, the principles covered in a first-year college physics course). Indeed, the best ways of solving certain problems are self-contradictory if examined too closely. Consider the following problem (from [5], p. 67): A rifle with a muzzle velocity of 1500 ft/s shoots a bullet at a target 150 ft away. How high above the target must the rifle be aimed so that the bullet will hit the target? An "expert" solution to this problem might proceed as follows : "It takes the bullet 0.1 second to reach the target. During this time, the bullet falls distance d = (1/2)gt*2 or (1/2)3270.1)2 ft, or 0.16 ft. So we aim'up by that amount to cancel the fall." In this solution, the "expert" has made several conflicting assumptions: first, that the bullet travels in a straight line; second, that the bullet falls from that path as it travels; and third, that the bullet is aimed upward to cancel the fall. Each succeeding assumption invalidates previous assumptions upon which it is based. In fact, the final answer is not exactly right; however, it differs from a more careful calculation for the parabolic path actually followed by the bullet by only about one part in a million. The "expert" has not solved this problem precisely, using all the applicable physical laws, but instead has treated the problem as a "nearly decomposable system" [61. Using qualitative knowledge that bullets move approximately in a straight line, the expert has decomposed the motion of the bullet into the dominant straight-line motion and the much smaller fall and upward motion. If we look harder, other decomposition assumptions can be found, viz. that air friction is negligible and that the earth is flat (i.e., that gravity is "straight down"). In fact, the laws of physics cannot be used directly in a deductive fashion to solve problems. For example, Newton's law, which we write compactly as "F = ma", relates the acceleration of a body (relative to an inertial reference frame) to the vector sum of all forces on the body; however, there are infinitely many such forces, and the frame of reference (e.go, the earth) isn't really inertial. Fortunately, in real problems most of the forces on the body are small and can be ignored, and the frame of reference is nearly inertial; using an appropriate decomposition of the problem, a good approximation of the answer to the problem can be found. Thus, solution of a real physics problem always involves treating a nearly decomposable system as if it were actually decomposable. Programs which use deduction to solve physics problems in a "microworld" are able to do so only because the decomposition decisions have been made, by the choice of the microworld and/or by the form in which the problem to be solved is encoded; however, this limits the extension of such a program to wider problem domains where other decompositions of "similar" problems are required. This view of problem solving in physics suggests answers to the questions posed earlier. Physics is hard because it is necessary to learn not only the equations, but also ways of decomposing actual problems so that application of the equations is tractable. The expert can solve problems better than the novice because the expert recognizes that a given (sub) problem is an instance of a class which can be decomposed in a particular way; this knowledge of how to decompose a real-world problem along lines suggested by the "fundamental concepts" of physics is a large part of what is sometimes called "physical intuition" r41 The knowledge of how to decompose problems may be retained by the student even though the formulas have been forgotten, and allows problems to be solved correctly even though the formulas have to be looked up again. The expert works forwards rather than backwards because the first 179 order of business is not to deduce the answer from the given information (which is likely to be grossly inadequate at.the beginning), but to find an appropriate decomposition or way of modeling the interactions of the objects in the problem. Once an appropriate decomposition has been found, solution of the problem is often straightforward. III RESEARCH ON PROBLEM SOLVING AND --~- LEARNING TO BE EXPERT --- We are currently writing a program to solve physics problems involving a variety of physical principles. Our work on this program is concentrating on the representation of problems, recognition of (sub) problem types which can be decomposed in particular ways, and learning of problem solving expertise through experience. Each of these areas is discussed briefly below. To insure that the problem solver is not told how to solve the problem by the manner in which the problem is stated to it, we assume that the input will be English or a formal language which could reasonably be expected as the output of an English parser such as the parser in [1,2]. For example, a car in a problem will be presented to the problem solver as "a car"; whether the car should be viewed as a location, a point mass, a rigid body, an energy conversion machine, etc. must be decided by the problem solver. We are developing a representation language which will allow multiple views of objects and variable levels of detail. For example, a block sliding on an inclined plane may be viewed as a weight, a participant in a frictional contact relation, and a location. A car might be viewed simply as a point mass, or as an object with its own geometry and complex internal structure, depending on the needs of a particular problem. The expert problem solver does not have a single, general representation of each physical law, but instead is able to recognize a number of special cases of each physical principle and use a rich set of redundant methods for dealing with them. For example, in addition to the general rigid body problem, the expert recognizes special cases such as a pivoted linear rigid body acted on by forces perpendicular to its axis. Such special cases often allow a desired unknown to be found using a single equation rather than many, or simplify analysis, e.g. by approximating a nonlinear equation with a linear one. Recognition of special cases is based on context, on what information is known, and on what answers are desired, as well as on the type of object or interaction. Our approach to recognition of special cases is to use a discrimination net, in which tests of features of the problem alternate with construction of new views of objects and collection of information into the appropriate form for a "schema" or "frame" representation of the view. Such a discrimination net can also be viewed as a hierarchical production system, or as a generalization of an Augmented Transition Network [71* If the recognition of the type of a (sub) problem is done by means of a discrimination net, special cases of a particular kind of physical system can be added to the net by adding discriminating tests for the special cases "above" the point in the net at which the more general case is recognized. We are investigating ways in which knowledge for handling such special cases could be learned automatically from experience in solving problems. One method of initiating such learning is data flow analysis of solutions using the more general solution. For example, if a problem involving a pivoted lever were solved using the general rigid body laws, data flow analysis would indicate that with a particular choice of the point about which to sum moments (the pivot), the "sum of forces" equations would not play a part in reaching the solution. A special case method could then be constructed from the general method by adding tests for the special case to the discrimination net and adding a corresponding "action" part which writes only the moment equation. Other opportunities for special case learning include elimination of zero terms (rather than eliminating them later algebraically), elimination of forces which turn out to be small when calculated, linearization of "almost linear" equations, use of small-angle approximations, and selection of simplified views of objects under appropriate conditions. REFERENCES 1. 2. 3. 4. 5. 6. 7. Novak, G. "Computer Understandi .ng of Physics Problems S tated in Na tural Language", American Journal of Computational Linguistics, Microfiche 53, 1976. Novak, G. "Representations of Knowledge in a Program for Solving Physics Problems", Proc. 5th IJCAI, Cambridge, Mass., Aug. 1977. Bundy, A., Byrd, L., Luger, G., Mellish, C. and Palmer, M. "Solving Mechanics Problems Using Meta-Level Inference", Proc. 6th IJCAI, Tokyo, 1979. Larkin, J., McDermott, J., Simon, D. P., and Simon, H. A. "Expert and Novice Performance in Solving Physics Problems", Science, vol. 208, No. 4450 (20 June 1980). Halliday, D. and Resnick, 5. Physics. ~- New York: Wiley, 1978. Simon, H. A. The Sciences of the Artificial, M.I.T. Press, 1969. - Woods, W. A. "Transition Network Grammars for Natural Language Analysis", comm. &J, vol. 13, no. 10 (Oct. 19701, PP. 591-606. 180	1980	46
41	KNOWLEDGE-BASED SIMULATION Philip Klahr and William S. Faught The Rand Corporation Santa Monica, California 90406 ABSTRACT Knowledge engineering has been successfully applied in many domains to create knowledge-based "expert" systems. We have applied this technology to the area of large-scale simulation and have implemented ROSS, a Rule-Oriented Simulation System, that simulates military air battles. Alternative decision-making behaviors have been extracted from experts and encoded as object- oriented rules. Browsing of the knowledge and explanation of events occur at various levels of abstraction. I. INTRODUCTION - Large-scale simulators have been plagued with problems of intelligibility (hidden embedded assumptions, limited descriptive power), modifiability (behaviors and rules buried in code), credibility (minimal explanation facilities), and speed (slow to build, to run, to interpret). The area of large-scale simulation provides a rich environment for the application and development of artificial intelligence techniques, as well as for the discovery of new ones. The field of knowledge engineering [2] is developing tools for use in building intelligent knowledge-based expert systems. A human expert communicates his expertise about a particular domain in terms of simple English-like IF-THEN rules which are then incorporated into a computer- based expert system. The rules are understandable, modifiable, and self-documenting. Knowledge-based systems provide explanation facilities, efficient knowledge structuring and sharing, and interfaces that are amiable for system building and knowledge refinement. Our approach to simulation views a decision- based simulator as a knowledge-based system. The behaviors and interactions of objects, the decision-making rules, the communiciation channels are all pieces of knowledge that can be made explicit, understandable, modifiable, and can be used to explain simulation results. For our work in simulation, we chose the domain of military air battles. Current large- scale simulators in this domain exhibit exactly the simulation problems we discussed above and thus provide a good area in which to demonstrate the feasibility and potential of knowledge-based simulation. II KNOWLEDGE REPRESENTATION -- .--- - Our domain experts typically centered their discussions of military knowledge around the domain objects. For example, a particular type of aircraft has certain attributes associated with it such as maximum velocity, altitude ranges, time needed to refuel, etc. Similarly, individual planes have given positions, speeds, altitudes, routes, etc. In addition, experts defined the behaviors of objects relative to object types or catagories. For example, they defined what actions aircraft take when they enter radar coverages, what ground radars do when they detect new aircraft (who they notify, what they communicate), etc. It became clear that an object-oriented (Simula-like [l]) programming language would provide a natural environment in which to encode such descriptions. We chose Director [6] for our initial programming language. Director is an object- oriented message-passing system that has been used primarily for computer graphics and animation. It offered considerable promise for our use in simulation, both in how knowledge is structured and how it is executed. Each object (class or individual) has its own data base containing its properties and behaviors. Objects are defined and organized hierarchically, allowing knowledge to be inherited, i.e., offsprings of objects automatically assume (unless otherwise modified) the properties and behaviors of their parents. In Director, as in other message-passing systems (e.g., Smalltalk [3] and Plasma [5]), objects communicate with each other by sending messages. The Director format for defining behaviors is (ask <object> do when receiving <message-pattern> <actions>), i.e., when the object receives a message matching the pattern, it performs the associated actions. In ROSS, we have added the capability of specifying IF-THEN rules of the form (IF <conditions> THEN <actions> ELSE <actions>) as part of an object's behavior. The conditions typically test for values (numeric, boolean) of data items while actions change data items or send messages to objects. Since Director is written in Maclisp, one may insert any Lisp s-expression as a condition or action. The following behavioral rule contains all of these options: 181 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. (ask RADAR do when receiving (IN RADAR RANGE ?AC) (SCRIPT (IF (lessp (length (ask MYSELF recall your OBJECTS-IN-RANGE)) (ask MYSELF recall your CAPACITY)) THEN (ask (ask (ask (ask ELSE (ask >>I HISTORIAN at ,STIME MYSELF detects ,AC) MYSELF add ,AC to your list of OBJECTS-IN-RANGE) ,(ask MYSELF recall your SUPERIOR) MYSELF detects ,AC) MYSELF monitor ,AC while in coverage) HISTORIAN at ,STIME MYSELF doesn't detect ,AC) This rule is activated when a radar receives a message that an aircraft (AC) is in its radar range. The radar tests whether the number of objects currently in its radar coverage is less that its capacity. If its capacity is not full, then the radar tells the historian that it detects the aircraft at time STIME (the current simulation time), it records the aircraft in its log, it notifies its superior that it detected the aircraft, and it continues to monitor the aircraft through its radar coverage. III -* BEHAVIORAL DESCRIPTIONS A difficult problem with large-scale simulators is that they contain complex behaviors that are hard to describe and understand. Each object has many potential actions it can take, and there may be hundreds of objects whose behavior the user may wish to examine and summarize at differing levels of generality and detail. To alleviate this problem, we organized the simulator's behavioral descriptions along two lines: 1. Static descriptions: descriptions of the major events simulated and the rules governing each object's potential actions. The events and rules are organized so that users can quickly peruse and understand the knowledge. 2. Dynamic descriptions: descriptions of each object's behavior as the simulator runs. Events are organized so the user is not overwhelmed by a mass of event reports. Events are reported as patterns with detail eliminated and selected events highlighted. To organize the descriptions, we constructed scenarios representing sequential event chains. Each major event in ROSS has an associated llevent descriptor" (ED). EDs are collected into chains, where each chain is a linear list of EDs. Each ED is causally associated with its immediate neighbors in the list: a preceding ED is necessary to cause its successor in that chain, but not necessarily sufficient. ([7] discusses the use of such chains in constructing proofs.) ED chains are further organized into Ilactivities," e.g., radar detection of an aircraft. Each activity is a tree of EDs. The root of the tree corresponds to the event that starts the activity. Each path from the root to a leaf is an ED chain. Logically, each ED chain corresponds to one possible scenario of events that could occur in a simulation. The scenario structure is used for both static and dynamic behavior descriptions. !!* BROWSING KNOWLEDGE Static behavior descriptions are given by ROSS's "browse" function, an on-line interactive facility with which users can examine ROSS's knowledge base. The user is initially given a list of all activities. He then selects an activity, and the browse function prints a list of the names of all EDs in that activity. The user can ask for a more complete description, in which case the browse function prints a tree of the EDs. The user can then select a particular ED to examine, and the browse function prints a simplified description of the event. If the user asks for a more complete description, the browse function prints the actual code for the corresponding behavior (as in the example above). At any point the user can select the next ED, or go up or down activity/event levels. The composition of ED chains, i.e., which EDs are members of which chains, is selected by the system developers for "naturalness" or appeal to a user's intuitive structure of the simulator. The system itself constructs the actual ED chains from the simulator's source code and a list of start and end points for each chain. The facility appears to be quite useful in our domain where the objects have interdependent yet self-determined behavior. B. EVENT REPORTING -- ROSS contains several objects that have been defined to organize, select, and report events to users. The Historian receives event reports from other objects (as exemplified in the behavioral rule above) and, upon request, supplies a history of any particular object, i.e., a list of events involving the object up to the current simulation time. The Historian also sends event reports to the Reporter, who selectively reports events to the user on his terminal as the simulator is running. The user can request to see all reports or only those involving a particular object or objects of a particular class. In addition, a Statistician accumulates statistics about particular fixed events (e.g., the total number of radar detections). (We have also interfaced ROSS to a color graphics system which visually displays simulation runs. The graphics facility has been an indispensable tool for understanding ROSS and the simulations it produces and for debugging.) C. EXPLANATION USING SCENARIOS - To explain "why" certain results occur, traditional rule-based systems typically show the 182 rules that were used, one by one, backchaining from the conclusions. We have taken a different approach to explanation in ROSS. Rather than displaying individual rules to explain events, ROSS presents higher-level behavioral descriptions of what happened, akin to the activity and event descriptions used for browsing. It is often the case that an event occurring in a simulaton run can be explained by specifying the chain of events leading up to the current event. This is accomplished simply by comparing event histories (gathered by the Historian) to the ED trees described above. The user can then browse the events and activities specified to obtain the applicable rules. What is perhaps more interesting in simulation is explaining why certain events do not occur. Often times simulations are run with expectations and the user is particularly interested in those cases where the expectations are violated. ([41 describes how such expectations can be used to drive a learning mechanism.) We have developed an initial capability for such "expectation-based" explanation. Explanations are given relative to specified ED chains. One event chain is designated as the "expected" chain. An "analyzer" reports deviations from this chain, i.e., it determines the point (event) at which the ED chain no lcnger matches the simulation events. Typical responses from the analyzer are: aircraft in radar range but not detected; radar sent message to superior but it was not received; command center requested aircraft assignment but none available. Such analysis can occur at any time within a simulation run to determine the current status of expected event chains. It is important to note that expectations need not be specifed prior to a simulation run (although this could focus the simulator's reporting activity). Users can analyze events with respect to any number of existing scenarios. Each analysis provides a simplified description and explanation of the simulator's operation from a different point of view (e.g., from radar's view, from aircraft's view, from a decision maker's view). This feature has also been extremely useful in debugging ROSS's knowledge base. IV* ---- SUMMARY AND FUTURE RESEARCH - ROSS currently embodies approximately 75 behavioral rules, 10 object types, and has been run with up to 250 individual objects. To show ROSS's flexibility, we have developed a set of alternative rule sets which encompass various strategies and tactics. Simulation runs using alternative rules show quite different behaviors and results. Future research will include scaling ROSS up both in complexity and in numbers of objects. Our goal is to turn ROSS into a realistic, usable tool. A more user-oriented English-like rule-based language will be required for users to express behaviors and strategies. We are looking toward ROSIE PI, or a hybrid ROSIE object-oriented language for this purpose. Scaling-up will necessitate enhancements in speed. We plan to explore parallel processing, abstraction (e.g., aggregating objects, adaptive precision), sampling, and focusing on user queries to avoid irrelevant processing. In summary, we have applied knowledge engineering to large-scale simulation and implemented ROSS, an interactive knowledge-based system that simulates the interactions, communications, and decision-making behavior within the domain of military air battles and command and control. We have shown the feasibility and payoff of this approach and hope to apply it to other domains in the future. ACKNOWLEDGMENTS We wish to thank Ed Feigenbaum, Rick Hayes-Roth, Ken Kahn, Alan Kay, Doug Lenat, Gary Martins, Raj Reddy, and Stan Rosenschein for their helpful discussions and suggestions during ROSS's development. We thank our domain experts Walter Matyskiela and Carolyn Huber for their continuing help and patience, and William Giarla and Dan Gorlin for their work on graphics and event reporting. REFERENCES 1. Dahl, O-J. and Nygaard, K. Simula -- an Algol-based simulation language. Communications ACM, 9 (9), 1966, 671-678. 2. Feigenbaum, E. A. The art of artificial intelligence: themes and case studies in knowledge engineering. Proc. IJCAI-77, MIT, 1977, 1014-1049. - -- 3. Goldberg, A. and Kay, A. Smalltalk- Instruction Manual, SSL 76-6, Xerox Palo Alto Research Center, 1976. 4. Hayes-Roth, F., Klahr, P., and Mostow, D. J. Knowledge acquisition, knowledge programming, and knowledge refinement. R-2540-NSF, Rand Corporation, Santa Monica, 1980. 5. Hewitt, C. Viewing control structures as patterns of passing messages. Artificial Intelligence, 8 (3), 1977, 323-364. 6. Kahn, K. M. Director Guide, AI Memo 482B, Arti .ficial Intel1 .igence Lab, MIT, 1979. 7. Klahr, P. Planning techniques for rule selection in deductive question-answering. In Pattern-Directed Inference Systems, D. A. Waterman -____ and F. Hayes-Roth (Eds.), Academic Press, New York, 1978, 223-239. 8. Waterman, D. A., Anderson, R. H., Hayes-Roth, F Klahr, P., Martins, G., and Rosenschein S. J. DeAign of a rule-oriented system for impleienting expertise. N-1158-l-ARPA, Rand Corporation, Santa Monica, 1979. 183	1980	47
42	I NTERACTI VE FRAME I NSTANTIATION Carl Engelman Ethan A. Scar1 Charles H. Berg* ABSTRACT This paper discusses the requirements that interactive frame instantiation imposes on constraint Verification. The representations and algorithms of an implemented software solution are presented. INTRODUCTION A number of frame representation languages or data access packages, seven of which are discussed in [STEFIK], have been developed as LISP extensions. In most applications of these languages, frame instantiation -- the creation of a new frame which represents an "instance", i.e., a more specific example, of a given "generic" frame -- occurs as a major theme. Yet, these languages do not really provide control structures sufficient to support interactive frame instantiation. Most of this paper will be concerned with constraint verification. A frame representation language typically provides the programmer with a way of attaching a constraint as a "facet" Of a given slot. It will reject any proposed values for the slot which fail that constraint, screaming a bit to the user. Such constraints attached to a slot in some generic frame also obtain automatically for slots of the same name occurring within its progeny. That is, they are "inherited". And that's about it. To explain why we felt the need for more control of constraint verification and, in fact, of the whole dynamics of interactive frame instantiation, we must present just a bit of our application. THE APPLICATION The KNOBS project [ENGELMAN] is directed towards the development of experimental consultant systems for tactical air command and control. We chose to focus first on what seemed to be a very simple type of aid. Imagine an Air Force officer is trying to plan a mission to strike some particular target. We are providing a program which interactively accepts the target, the airbase from which to fly the mission, the type of plane, the time of take-off, etc., and checks the input for inconsistencies and oversights. Such missions are stereotypes which itCurrent affiliation, AUTOMATIX Inc. The MITRE Corporation P.O.Box 208 Bedford, MA 01730 are represented naturally as frames and the checks are constraints among the possible slot values in such frames. DATA BASE/LANGUAGE SETTING k We first translated FRL [ROBERTSJULY] ROBERTSSEPT) f rom MACLISP to INTERLISP ERICSON]. Data bases of targets and resources ave been implemented as nets of individual and generic frames. An individual target frame, for example, contains information, e.g., location, specific to a particular target, while a generic target frame contains information true about classes of targets, for instance, the type of radar normally associated with a particular kind of surface-to-air missile. In all, the data base currently contains some 1400 frames. We have introduced several upwards compatible extensions to FRL, e.g., programmer controlled parallel inheritance along paths defined by a specified set of slot names and the controlled automatic invocation of "$IF-NEEDED" procedures during attempts to retrieve missing data. We split the concept of generic frame: those which we continue to refer to as "generic frames" contain information (defaults, attached demons, etc.) applicable to all their instances. Their slots are not necessarily in correspondence with those of their instances. What we refer to as a "template", on the other hand, is a prototypical representation of an instance of the associated generic frame. Its slots correspond to those to be found in an instance, but contain "$IF-NEEDED" procedures where the instance would contain values. It also contains constraints on the values that may appear in an instantiation. We also differentiate frames representing fully specified objects from those representing subclasses. The former is referred to as "instance" call it in [sTEFIK] and [SZOLOVITS], and we "individual". Such a frame is identified by having an "AIO" (An-Individual-Of) slot pointing back to its generic frame. AI0 corresponds to set membership, while AK0 (A-Kind-Of) corresponds to set inclusion. While the frame instantiation procedures are designed for use at any level, we have thus far employed them only in the creation of individual frames. DESIDERATA Our goal is a demonstration which, like a good shortstop, makes it look easy. Some requirements are: 184 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. 1. The system must know what information is Constraints needed and how-to ask for it. It must also know when the instantiation is complete and what to do then. 2. The user must be able to enter a value for'any slot in any frame at any time and the system must know what must be checked or rechecked and what remains to be done. 3. There should be a general facility to suggest choices for a slot which are consistent with the current values of the other slots. 4. The system should complain as soon as the slot values become inconsistent. It must show dynamic discretion in explaining enough, but not too much, of the difficulty. 5. The user must be able to ask questions about the data base and about the status of the instantiation. The form of a constraint is: (name domain expr), where "name" is an atom used to index a user-oriented explanation of the constraint, "domain" (the terminology is suggested in {STANSFIELD]) is the list of slot names whose interrelations are tested by the constraint, and "expr" is a predicate to be satisfied. "Expr" can refer to the current values of slots being instantiated simply by reference to their slot names. We define a bucket as an unordered list of constraints. The contents of the CONSTRAINTS slot in the template is a list of buckets, ordered to express priority. All constraints in the same bucket are of equal priority. The attachment of constraints to the template at the slot level, rather than the usual attachment to slots at the facet level, reflects our view that all the action is in the interaction of the slots and that it is presumptive to make a decision -- especially a static one -- as to which slot is "bad". 6. The instantiation of one frame must be able to initiate the instantiation of another. 7. Constraint satisfaction must be maintained after the original instantiation whenever a slot value is changed or a template is changed. EXECUTION Assuming a sequence of values is being suggested for the instantiated slots, the algorithm is as follows: Initially all constraints are unmarked. REPRESENTATION Templates Templates are represented by frames with slots whose names will be replicated in the instantiated frames and which contain either $IF-NEEDED or $VALUE facets. The $IF-NEEDED procedures are responsible for deciding whether the value can be computed or is to be requested from the user. It is our normal practice to have the $IF-NEEDED procedures also perform type checking. The presence of a SVALDE facet causes a recursive call to the instantiator. The template also contains two special slots, named CONSTRAINTS (discussed below) and BOOKEEEP. The BOOKKEEP slot contains procedures to be run upon completion of the frame instantiation, a sort of "IF-ADDED" mechanism at the frame level. The interpreter either steps through the template filling slots or fills those commanded by the user. Moreover, the user may interact at any time wi th LISP or with a natural language Q/A system for retrieving facts from the data base, including those inferred through "inheritance". The latter is implemented as an ATN parser, whose syntax and semantics are intimately related to the structure of the frame net. In add ition, there are a number of amenities: spelling correctors, synonym and word truncation recognizers, and facilities for viewing the current status of the instantiation process values, i.e. , a presentation of the current slot disti nguishing with explanation those At any time, there is a current slot name, the one for which a value has been most recently proposed. A constraint is considered timely if its domain includes the current slot name and if all the other slot names in its domain have values already assigned. The interpreter passes through the buckets in decreasing priority until it discovers a timely constraint. If the constraint fails, the interpreter marks the constraint and traps the slots in its domain, i.e., renders their values unknown to constraints in lower priority buckets, which are not tested. If these lower priority constraints are already marked, they become unmarked since they are no longer timely. Other failed constraints in the current bucket are marked and their domains trapped. If a previously marked constraint now succeeds, then it is unmarked. If all the constraints in a bucket having a given slot name become unmarked, the slot name is released, i.e., pushed down to lower priority buckets along with the current slot name. The process normally terminates when all the slots are filled and none of the constraints are marked. CONSISTENCY MAINTENANCE Should a slot value in an instantiated frame or a constraint in a template be changed, the system makes appropriate checks. COMPLEX CONSTRAINTS The discussion above deals with constraints on the related slots in a given frame. Such constraints are called simple in [STANSFIELD]. What he calls complex constraints, those which violate constraints. 185 involving slots in ditrerent frames, are of great importance to us. For example, we must be concerned with the timing constraints needed to synchronize a primary mission with its support missions (refueling, defense suppression, etc.). We are currently engaged in the design and implementation of suitable representations and algorithms. We believe that a purely recursive (depth first) sequence of frame instantiations is not acceptable, and that we shall have to provide flexible control of interleaved "co-instantiations". CHOICE GENERATION When cueing the user for a slot-value, we would often like to present a list of values consistent with those already chosen for other slots. This is, in general, computationally impossible. It turns out, however, to be in the nature of our application that we frequently can produce a list of consistent values. Furthermore, we can do this by a fairly general method, generating the choices, in fact, from the constraints. The key point -- and this is obviously application dependent -- is that many of our constraints are of the form (name (A Bl -- Bn) (MEMBER A (f00 Bl -- Bn))), where, if the code is to make sense, (foo Bl -- Bn) is a computable, finite list. So, for example, one constraint might mean: (1) The airbase is one of our airbases in Europe. and another might mean: the (2) The chosen chosen airbase. fighter wing is located at We sav a constraint enumerates a slot. S. + I I absolutely iff it is of the form: (name (S)(MEMBER S----)). Constraint (l), above, enumerates airbases absolutely. A constraint To make choice generation more efficient, we optimize the constraints within the current context by collecting those which are timely and "compiling" them into a function of only the current slot, essentially by pre-evaluating all subexpressions which do not contain this slot. CRITICISM AND FUTURE DIRECTIONS 1) We need to design and implement a comparable system for the complex constraints. 2) The only relative priorities we can express between constraints are static. This might, someday, prove inadequate. 3) There is danger of an existential trap. The simplest example occurs when the first slot filled is A and the candidates value satisfies every constraint whose domain is (A). There may, however, be a constraint whose domain is (A B) which cannot be satisfied with the proposed value of A and any value of B. Our interpreter does not see this until B is selected. The choice generation scheme discussed above could also be employed to test (perhaps very expensively), for such traps. ACKNOWLEDGEMENTS This work was supported by the Rome Air Development Center under Air Force contract F19628-80-C-0001. The FRL system was originally created at MIT by R. Bruce Roberts and Ira P. Goldstein. Roberts assisted us by defining and creating an "export nucleus" of FRL. The work reported here is built directly on the INTERLISP version of FRL translated and extended by Lars W. Ericson, while he was at MITRE. We are deeply indebted to him. REFERENCES LENGELMAN] Engelman, C., Berg, Charles H., and Bischoff, Miriam, "KNOBS: An Experimental Knowledge Based Tactical Air Mission Planning System and a Rule Based Aircraft Identification Simulation Facility", Proc. Sixth Inter. Joint Conf. Artificial Intelligence, Tokyo, 1979, pp. 247-249. [ERICSON] Ericson, Lars W., "Translation of Programs from MACLISP to INTERLISP", MTR-3874, The MITRE Corporation9 Bedford, MA, Nov. 1979. [ROBERTSJULY] Roberts, R. Bruce, and Goldstein, Ira P., "The FRL Primer", MIT AI Lab. Memo 408, July 1977. [ROBERTSSEPT) Roberts, R. Bruce, and Goldstein, Ira P., "The FRL Manual", MIT AI Lab. Memo 409, September 1977. [STANSFIELD] Stansfield, James L., "Developing Support Systems for Information Analysis", in Artificial Intelligence, An MIT Perspective, Winston, P. H., and Brown, R. H., (Eds.), The MIT Press, Cambridge, MA, 1979. [STEFIK] Stefik, Mark, "An Examination of a Frame-Structured System", Proc. Sixth Inter. Joint Conf. on Artificial Intelligence, Tokyo, 1979, pp. 845-852. [SZOLOVITS] Szolovits, P., Hawkinson, L. B., Martin, W. A., "An Overview of Owl, A Language for Knowledge Representation", MIT/LCS/TM-86, MIT, Cambridge, MA, June, 1977. 186	1980	48
43	DESCRIPTIONS FOR A PROGRAMMING ENVIRONMENT Ira P. Goldstein and Daniel G. Bobrow Xerox Palo Alto Research Center Palo Alto, California 94304, U.S.A Abstract PIE is an experimental personal information environment implemented in Smalltalk that uses a description language to support the interactive development of programs. PIE contains a network of nodes, each of which can be assigned several perspectives. Each perspective describes a different aspect of the program structure represented by the node, and provides specialized actions from that point of view. Contracts can be created that monitor nodes describing different parts of a program’s description. Contractual agreements are expressible as formal constraints, or, to make the system failsoft, as English text interpretable by the user. Contexts and layers are used to represent alternative designs for programs described in the network. The layered network database also facilitates cooperative program design by a wow, and coordinated, structured documentation. Int reduction In most programming environments, there is support for the text editing of program specifications, and support for building the program in bits and pieces. However, there is usually no way of linking these interrelated descriptions into a single integrated structure. The English descriptions of the program, its rationale, general structure, and tradeoffs are second class citizens at best, kept in separate files, on scraps of paper next to the terminal, or, for a while, in the back of the implementor’s head. Furthermore, as the software evolves, there is no way of noting the history of changes, except in some primitive fashion, such as the history list of Interlisp [lo]. A history list provides little support for recording the purpose of a change other than supplying a comment. But such comments are inadequate to describe the rationale for coordinated sets of changes that are part of some overall plan for modifying a system. Yet recording such rationales is necessary if a programmer is to be able to come to a system and understand the basis for its present form. Developing programs involves the exploration of alternative designs. But most programming environments provide little support for switching between alternative designs or comparing their similarities and d<fferences. They do not allow alternative definition,s of procedures and data structures to exist simultaneously in the programming environment: nor do they provide a representation for the evolution of a particular set of definitions across time. In this paper we argue that by making descriptions first class objects in’s programming environment, one can make life easier for the programmer through the life cycle of a piece of software. Our argument is based on our experience with PIE, a description-based programming environment that supports the design, development, and documentation of Smalltalk programs. Networks The PIE environment is based on a network of nodes which describe different types of entities. We believe such networks provide a better basis for describing systems than files. Nodes provide a uniform way of describing entities of many sizes, from small pieces such as a single procedure to much larger conceptual entities. In our programming environment, nodes are used to describe code in individual methods, classes, categories of classes, and configurations of the system to do a particular job. Sharing structures between configurations is made natural and efficient by sharing regions of the network. Nodes are also used to describe the specifications for different parts of the system. The programmer and designer work in the same environment, and the network links elements of the program to elements of the design and specification. The documentation on how to use the system is embedded in the network also. Using the network allows multiple views of the documentation. For example, a primer and a reference manual can share many of the same nodes while using different organizations suited to their different purposes. In applying networks to the description of software, we are following a tradition of employing semantic networks for knowledge representation. Nodes in our network have the usual characteristics that we have come to expect in a representation language--for example, defaults, constraints, multiple perspectives, and context-sensitive value assignments. There is one respect in which the representation machinery developed in PIE is novel: it is implemented in an object-oriented language. Most representation research has been done in Lisp. Two advantages derive from this change of soil. The first is that there is a smaller gap between the primitives of the representation language and the primitives of the implementation language. Objects are closer to nodes (frames, units) than lists. This simplifies the implementation and gains some advantages in space and time costs. The second is that the goal of representing software is simplified. Software is built of objects whose resemblance to frames makes them natural to describe in a frame-based knowledge representation. 187 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Perspectives Attributes of nodes are grouped into perspectives. Each perspective reflects a different view of the entity represented by the node. For example, one view of a Smalltalk class provides a definition of the structure of each instance, specifying the fields it must contain; another describes a hierarchical organization of the methods of the class; a third specifes various external methods called from the class; a fourth contains user documentation of the behavior of the class. The attribute names of each perspective are local to the perspective. Originally, this was not the case. Perspectives accessed a common pool of attributes attached to the node. However, this conflicted with an important property that design environments should have, namely, that different agents can create perspectives independently. Since one agent cannot know the names chosen by another, we were led to make the name space of each perspective on a node independent. Perspectives may provide partial views which are not necessarily independent. For example, the organization perspective that categorizes the methods of a class and the documentation perspective that describes the public messages of a class are interdependent. Attached procedures are used to maintain consistency between such perspectives. Each perspective supplies a set of specialized actions appropriate to its point of view. For example, the print action of the structure perspective of a class knows how to prettyprint its fields and class variables, whereas the organization perspective knows how to prettyprint the methods of the class. These actions are implemented directly through messages understood by the Smalltalk classes defining the perspective. Messages understood by perspectives represent one of the advantages obtained from developing a knowledge representation language within an object-oriented environment. In most knowledge representation languages, procedures can be attached to attributes. Messages constitute a generalization: they are attached to the perspective as a whole. Furthermore, the machinery of the object language allows these messages to be defined locally for the perspective. Lisp would insist on global functions names. Contexts and Layers All values of attributes of a perspective are relative to a context. Context as we use the term derives from Conniver [9]. When one retrieves the values of attributes of a node, one does so in a particular context, and only the values assigned in that context are visible. Therefore it is natural to create alternative contexts in which different values are stored for attributes in a number of nodes. The user can then examine these alternative designs, or compare them without leaving the design environment. Since there is an explicit model of the differences between contexts, PlEc can highlight differendes between designs. PIE also provides tools for the user to choose or create appropriate values for merging two designs. Design involves more than the consideration of alternatives. It also involves the incremental development of a single alternative. A context is structured as a sequence of layers. It is these layers that allow the state of a context to evolve. The assignment of a value to *a property is done in a particular layer. Thus the assertion that a particular procedure has a certain source code definition is made in a layer. Retrieval from a context is done by looking up the value of an attribute, layer by layer. If a value is asserted for the attribute in the first layer of the context, then this value is returned. If not, the next layer is examined. This process is repeated until the layers are exhausted. Extending a context by creating a new layer is an operation that is sometimes done by the system, and sometimes by the user. The current PIE system adds a layer to a context the first time the context is modified in a new session. Thus, a user can easily back up to the state of a design during a previous working session. The user can create layers at will. This may be done when he or she feels that a given groups of changes should be coordinated. Typically, the user will group dependent changes in the same layer. Layers and contexts are themselves nodes in the network. Describing layers in the network allows the user to build a description of the rationale for the set of coordinated changes stored in the layer in the same fashion as he builds descriptions for any other node in the network. Contexts provide a way of grouping the incremental changes, and describing the rationale for the group as a whole. Describing contexts in the network also allows the layers of a context to themselves be asserted in a context sensitive fashion (since all descriptions in the network are context-sensitive). As a result, super-contexts can be created that act as big switches for altering designs by altering the layers of many sub-contexts. Contracts and Constraints In any system, there are dependencies between different elements of the system. If one changes, the other should change in some corresponding way. We employ contracts between nodes to describe these dependencies. Implementing contracts raises issues involving 1) the knowledge of which elements are dependent; 2) the way of specifying the agreement; 3) the method of enforcement of the agreement; 4) the time when the agreement is to be enforced. PIE provides a number of different mechanisms for expressing and implementing contracts. At the implementation level, the user can attach a procedure to any attribute of a perspective, (see [2] for a fuller discussion of attached procedures): this allows change of one attribute to update corresponding values of others. At a higher level, one can write simple constraints in the description language (e.g. two attributes should always have identical values), specifying the dependent attributes. The system creates attached procedures that maintain the constraint. There are constraints and contracts which cannot now be expressed in any formal language. Hence, we want to be able to express that a set of participants are interdependent, but not be required to give a formal predicate specifying the contract. PIE allows us to do this. Attached procedures are created for such contracts that notify the user if any of the participants change, but which do not take any action on their own to maintain consistency. Text can be attached to such informal contracts that is displayed to the user when the contract is triggered. This provides a useful inter-programmer means of communication and preserves a failsoft quality of the environment when formal descriptions are not available. Ordinarily such non-formal contracts would be of little interest in artificial intelligence. They are, after all, outside the comprehension of a reasoning program. However, our thrust has been to build towards an artificially intelligent system through succcessive stages of man-machine symbiosis. This 188 approach has the advantage that it allows us to observe human reasoning in the controlled setting of interacting with the system. Furthermore, it allows us to investigate a direction generally not taken in Al applications: namely the design of memory-support rather than reasoning-support systems. An issue in contract maintenance is deciding when to allow a contract to interrupt the user or to propagate consistency modifications. We use the closure of a layer as the time when contracts are checked. The notion is that a layer is intended to contain a set of consistent values. While the user is working within a layer, the system is generally in an inconsistent state. Closing a layer is an operation that declares that the layer is complete. After contracts are checked, a closed layer is immutable. Subsequent changes must be made in new layers appended to the appropraiate contexts. Coordinating designs So far we have emphasized that aspect of design which consists of a single individual manipulating alternatives. A complementary facet of the design process involves merging two partial designs. This task inevitably arises when the design process is undertaken by a team rather than an individual. To coordinate partial designs, one needs an environment in which potentially overlapping partial designs can be examined without overwriting one another. This is accomplished by the convention that different designers place their contributions in separate layers. Thus, where an overlap occurred, the divergent values for some common attributes are in distinct layers. Merging two designs is accomplished by creating a new layer into which are placed the desired values for attributes as selected from two or more competing contexts. For complex designs, the merge process is, of course, non-trivial. We do not, and indeed cannot, claim that PIE eliminates this complexity. What it does provides is a more finely grained descriptive structure than files in which to manipulate the pieces of the design. Layers created by a merger have associated descriptions in the network specifying the contexts participating in the merger and the basis for the merger. Meta-description Nodes can be assigned meta-nodes whose purpose is to describe defaults, constraints, and other information about their object node. Information in the meta-node is used to resolve ambiguities when a command is sent to a node having multiple perspectives. One situation in which ambiguity frequently arises is when the PIE interface is employed by a user to browse through the network. When the user selects a node for inspection, the interface examines the meta-node to determine which information should be automatically displayed for the user. By appropriate use of meta-information, we have made the default display of the PIE browser identical to one used in Smalltalk. (Smalltalk code is organized into a simple four-level heirarchy, and the Smalltalk browser allows examination and modification of Smalltalk code using this taxonomy.) As a result, a novice PIE user finds the environment similar to the standard Smalltalk programming environment which he has already learned. Simplifying the presentation and manipulation of the layered network underlying the PIE environment remains an important research goal, if the programming environment supported by PIE is to be useful as well as powerful. We have found use of a meta-level of descriptions to guide the presentation of the network to be a powerful device to achieve this utility. Conclusion PIE has been used to describe itself, and to aid in its own development. Specialized perspectives have been developed to aid in the description of Smalltalk code, and for PIE perspectives themselves. On-line documentation is integrated into the descriptive network. The implementors find this network-based approach to developing and documenting programs superior to the present Smalltalk programming environment. A small number of other people have begun to use the system. This paper presents only a sketch of PIE from a single perspective. The PIE description language is the result of transplanting the ideas of KRL [2] and FRL [6] into the object oriented programming environment of Smalltalk [8], [7]. A more extensive discussion of the system in terms of the design process can be found in [l], and [4]. A view of the PIE description language as an extension of the object oriented programming metaphor can be found in [5]. Finally, the use of PIE as a prototype office information system is described in [31. References ill PI [31 [41 151 161 [71 WI WI [lOI Bobrow, D.G. and Goldstein, I.P. “Representing Design Alternatives”, Proceedings of the A/S6 Conference, Amsterdam, 1980. Bobrow, D.G. and Winograd, T. “An overview of KRL, a knowledge representation language”, Cognifive Science 1, 1 1977. Goldstein, I.P. “PIE: A network-based personal information environment”, Proceedings of the Office Semantics Workshop, Chatham, Mass., June, 1980. Goldstein, I.P. and Bobrow, D.G., “A layered approach to software design “, Xerox Palo Alto Research Center CSL- 80-5. 1980a. Goldstein, I.P. and Bobrow, D.G., “Extending Object Oriented Programming in Smalltalk”, Proceedings of the Lisp Conference. Stanford University, 1980b. Goldstein, I.P. and Roberts, R.B. “NUDGE, A knowledge- based scheduling program”, Proceedings of the Fifth International Joint Conference 0” Artificial Intelligence, Cambridge: 1977, 257-263. Ingalls, Daniel H., “The Smalltalk- Programming System: Design and Implementation,” Conference Record of the Fifth Annual ACM Symposium on Principles of Programming Languages, Tucson, Arizona, January 1978, pp 9-16. Kay, A. and Goldberg, A. “Personal Dynamic Media” /EEE Computer, March, 1977. Sussman, G., & McDermott, D. “From PLANNER to CONNIVER -- A genetic approach”. Fall Joint Computer Conference. Montvale, N. J.: AFIPS Press, 1972. Teitelman, W., The lnterlisp Manual, Xerox Palo Alto Research Center, 1978. 189	1980	49
44	A BASIS FOR A THEORY OF PROGRAM SYNTHESIS’ P.A.Subrahmanyam USC/Information Sciences Institute and Cepartment of Computer Science University of Utah, Salt Lake City, Utah 84112 1. Introduction and Summary In order to obtain a quantum jump in the quality and reliability of software, it is imperative to have a coherent theory of program synthesis which can serve as the basis for a sophisticated (interactive) software development tool. We argue that viewing the problem of (automatic) program synthesis as that (automatically) synthesizing implementations of abstract data types provides a viable basis for a general theory of program syntheeis. We brlefly describe the salient features of such a theory [5, Sj, and conclude by listing some of the applications of the theory. 1.1. Roquiromonts for en Accrptable Theory of Program Synthesis. We view some of the essential requirements of en acceptable theory of program synthesis to be the following: - the theory should be general; - it must adhere to Q coherent set of underlying principles, and not be based on an ad hoa collection of heuristics; - it must be based on a sound mathematical framework; - it must account for the “state of the art” of program synthesis; in particular, it must allow for the generation of “efficient” programs. Further, if a theory is to be useful, we desire that it possess the following additional attributes: - it should serve as the basis for a program development system which can generate provably correct non-trivial programs; - it should possess adequate flexibility to admit being tailored to specific application tasks1 - it should provide new and insightful perspectives into the nature of programming and problem solving, 1 This work warn rupportrd in prrt by l n IBM Followskip With these requirements in mind, we now examino the nature of the programming process in an attempt to characterize the basic principles that underly the conttructlon of “good” programs. 2 A basis for a Theory of Program Synthesis ---_ Intuitively, the abstraction of a problem can be viewed as consisting of an appropriate set of functions to be performed on an associated set of objects. Such a collection of objects and functions is an “abstract data typo” and has the important advantage of providing a representation independent characterization of a problem. Although the illustrations that most readily come to mind are commonly employed data structures such as a stack, a file, a queue, a symbol table, etc., any partial recursive function can be presented as an abstract data type. Programming involves representing the abstractions of objects and operations relevant to a given problem domain using primitives that are presumed to be already available; ultimately, such primitives are those that are provided by the available hardware. Various programming methodologies advocate ways of achieving “good” organizations of layers of such representations, in attempting to provide an effective means of coping with the complexity of programs. _ there exists, therefore, compelling evidence in favor of viewing the process of program synthesis as one of obtaining an implementation for the data type correponding to the problem of interest (the “type of interest”) in terms of another data type that corresponds to some representation (the “target type.“) This perspective is’ further supported by the following basic principles that we think should underly the synthesis process (if reliable programs are to be produced consistently): 1. The programming process should essentially be one of program synthesis proceeding from the spocificrtions of a problem, rather than being primarily analytic (e.g. constructing a program and then verifying it) or empirical (e.g. constructtng a program and then testing it). 74 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. 2. The specification of a problem should be representation independent. This serves to guarantee complete freedom in the program synthesis process, in that no particular program is excluded a priori due to overspecification of the problem caused by representation dependencies. 3. The synthesis should be guided primarily by the semantics of the problem specification. 4. The level of reasoning used by the synthesis paradigm should be appropriate to “human reasoning,” rather than being machine oriented (see [6]). In addition to making the paradigm computationally more feasible, this has two major advantages: a. existing paradigms of programming such as “stepwise refinement” can be viewed in a mathematical framework b. user interaction with the system becomes more viable, since the level of reasoning is now “visible” to the user. The above principles led us to adopt an algebraic formulation for the development of our theory [l, 21, [3, 43. An important consequence of this decision was that the synthesis paradigm is independent of any assumptions relating to the nature of the underlying hardware. In fact, it can even point to target types suited to particular problems of interest i.e. trrget machine architectures which aid efficient implementations. 3. the Proposed Paradigm for Prowam Synthesis We adopt the view that any object representing an instance of a type is completely characterized by its “externally observable behavior”. The notion of an implementation of one data type (the type of interest) in terms of another (the target type) is then defined as a map between the functions and objects of the two types which preserves the observable behavior of the type of interest. The objective, then, is to develop methods to automate the synthesis of such implementations based on the specifications of the type of interest and the target type. Intuitively, the crux of the proposed paradigm lies in “mathematically” incorporating the principle of stepwise refinement into automatic programming. This is done by appropriately interpreting both the syntactic and semantic structure inherent in a problem. An important distinction from most transformation based systems is that the refinem@rnt is guided by the semantics of the functions define on the type of interest, rather than by a fixed set of rules (e.g. [7]>. An formal characterization of some of the pivotal steps in the synthesis process is provided, and an attempt is made to pin-point those stages where there is leeway for making alternative choices based upon externally imposed requirements. (An example of such a requirement is the relative efficiency desired for the implementations of different functions depending upon their relative frequency of use.) This separation of the constraints imposed by (a) the structure inherent in the problem specification, (b) the requirements demanded by the context of use, and (c) the interface of these two, serves to further subdivide the complexity of the synthesis task -- it becomes possible now to seek to build modules which attempt to aid in each of these tasks in a relatively independent manner. In summary, our goal was to seek, in as far as is possible, a mathematically sound and computationally feasible theory of program synthesis. The formal mathematical framework underlying our theory is algebraic. The programs synthesized are primarily applicative in nature; they are provably correct, and are obtained without the use of backtracking. There is adequate leeway in the underlying formalism that allows for the incorporation of different “environment dependent” criteria relating to the “efficiency” of implementations. The objectives of the theory include that conventional programs be admitted as valid outcomes of the proposed theory, This is in consonance with our belief that any truly viable theory of synthesis should approximate as a limiting case already existing empirical data relevant to its domain. 4. An Example: The Synthesis of Block Structured Symbol Table Using an Indexed Array To illustrate some aspects of the paradigm for program synthesis we outline the synthesis of a block-structured SymbolTable using an indexed array as a target type (cf. [4].) The sorts of objects involved are instances of SymbolTable, Identifier, Attributes, Boolean, etc.; our primary interest here is on the manipulation of instances of SymbolTables. The functions that are defined for manipulating a SymbolTable include: NEWST (spawn a new instance of a symbol table for the outermost scope,) ENTERBLOCK (enter a new local naming scope,) ADDID (add an identifier and associated attributes to the symbol table,) LEAVEBLOCK (discard the identifier entries from the most current scope, re-establish the next outer scope,) ISINBLOCK (test to see if an identifier has already been declared in the current block,) and RETRIEVE (retrieve the attributes associated with the most recent definition of an Identifier.) The formal specifications may be found in [4] (see also [6].) Although implementations for more complex definitions of SymbolTables have been generated (which include tests for “Global” Identifiers), we have chosen this definition because of its familiarity, The overall synthesis proceeds by first categorizing the functions defined on the type of interest (here, the SymbolTable) into one of the flowing three categories: 0) Base 75 Constructor functions that serve to spawn new instances of the type (e.g. NEWST); (ii) Canstructor functions that serve to generate new instances from existing ones (e.g. ADDID, ENTERBLOCK, LEAVEBLOCKh and (iii) E&&or functions that return instances of types other than SymbolTable (e.g. RETRIEVE, ISINBLOCK). The next step is to identify a subset of these functions (termed kernel functions) which serve to generate all instances of SymbolTables: these are NEWST, ADOID, and ENTERBLOCK. A major step in obtaining an implementation for the TOI is to provide an implementation for the kernel functions. Since no model for the kernel functions is explicit in the specification of a type, a suitable model must be inferred from the behavior of the functions defined on the type. Such an inference follows from an examination of the axioms defining the extraction functions. Specifically, the domain of the terms of type SymbolTable is partitioned into its) equivalence classes by the extractors defined on the type -- and this is precisely what an implementation is attempting to capture. The defining equations of each function indicate how it “contributes” towards this partitioning, and therefore how this “semantic structure” imposed upon the terms of the SymbolTable is related to the syntactic structure of the underlying terms, Due to lack of space, we omit the details of how this is done. One of the implemetations generated (automatically) is shown in figure 1, wherein 0 denotes the “implementation map. We note that an auxiliary data type which is almost isomorphic to a Stack (of integers) was (automaticlly) defined in course of the implementation; this Stack can, in turn, be synthesized in terms of an indexed Array by a recursive invocation of the synthesis procedures. Other Implementations that are generated for the Symbol Table include an implementation using a “Block Mark” to identify the application Of the function ENTERBLOCK, and an implementation similar to a “hash table” implementation is suggested upon examining the semantics of the functions defined on the Symbol Table. --l----------cII--------------------~~~==~=-------~~~~~~~”~~-- We list below one of the implementations generated for a SymbolTable, using an Array as the initially specified target type. The final representation consists of the triple <Array,lnteger,AOT11 the integer represents the current index into the array, whereas AOTl (for &diary Eata lype-1) is introduced in course of the synthesis process, and is isomorphic to a Stack that records the index-values corresponding to each ENTERBLOCK performed on a particular instance of a SymbolTable.) We denote this by-writing t#s) - <a,i,edtl>. informally, ENTERBLOCK.ADTl serves to “push” the current index value onto the stack adtl, LEAVEBLOCK.ADTl serves to “POP” the stack, and D.ADTl returns the -topmost element in the Stack, returning a zero if the stack is empty. WCC and PREO are the successor and menus functions on Integers. B(NEWST) E <NEWARRAY,ZERO,NEWAOTl’ B(AOOIO(s,id,al)) - <ASSIGN(a,SUCC(i),<id,al>),SUCC(i),adt 1’ g(ENTERBLOCK(s)) - <a,& ENTERBLOCK.AOTl(adtl,i)> aLEAVEBLOCK(s)) - <a,D.AOTl(adt 1), LEAVEBLOCK.ADTl (adtl)> BIISINBLOCK(s,id 1)) = ISINBLOCKTT(<a,i,adt l>,id 1) bKRETRIEVE(s,idl)) - RETRIEVETT(<a,i,adt l>,ldl) lSlNE3LOCKTT and RETRlEVETT are defined as follows: ISINBLOCKTT(~a,i,adt 1~’ id 1) - if i - ZERO then FALSE else if D.ADTl(adt 1) < i then if proj(l, OATA = idl, then TRUE else ISINBLOCKTT(<a,PRED(i),adt l,id 1) else FALSE RETRIEVETT(<a,i,adt l>,id 1) = if i - ZERO then UNDEFINED else if proj(l,OATA)a,i)) - id1 then proj(2, OATA(a,i)) else RETRIEVETT(<a, PRED(i),adt I>, idl) Here, proj(i,exl..xn”) - Xi. Figure 1. A Symbol Table lmplemen!a!iOn Othw Examples of Applications of the Synthesis Paradigm. Several Programs have been synthesized by direct applications of the synthesis algorithms developed so far. These include implementations for a Stsck, a Queue, a Oeque, a Block Strut tured SymbolTable, an interactive line-orlented Text-Editor, o text formatter, a hidden surface elimination algorithm for graphical dispteys, and an execution engine for a data driven machine, eferencso [l] J.Goguen, J.Thatcher, E.Wagner, J.Wright. Initial Algebra Semantics and Continuous Algebras. JACM 24:68-95, 1977. (23 J.Goguen, J.Thatcher, E.Wagner. An Initial Algebra Approach to the Specification, Correctness, and implementation of Abstract Data Types, in Current Trends in Programming Methodology, Vol IV, Ed. R.Yeh, Prentice-Hall, NJ, 1979, pages 80-149. [33 J.Guttag, E.Horowitz, O.Musser. The Design of Data Type Specifications, in Current Trends in Programming Methodology, Vol IV, Ed. R.Yeh, Prentice-Hall, N.J., 1979. [43 J.Guttag, E.Horowitz, O.Musser. Abstract Data Types and Software Validation. CACM 21:1048-64, 1978. [53 P.A.Subrahmanyam. Towards a Theory of Program Synthesis: Automating Implementations of Abstract Data Types. PhO thesis, Dept. of Comp. SC., State University of New York a! Stony Brook, August, 1979. 16 J P.A.Subrahmanyam. A Basis for a Theory of Program Synthesis. Technical Report, Dept. of Computer Science, University of Utah, February, 1980. [7 3 OBarstow. Knowledge-Based Program Construction, Elsevier North-Holland Inc., NY., 1979. 76	1980	5
45	ABSTRACT Rule-Based Inference In Large Knowledge Bases * William Mark USC/Information Sciences Institute Having galned some experience with knowledge-based systems (e.g., [a), [9], [ 1 l]), our aspirations are growing. Future systems (for VLSI design, office automation, etc.) will have to model more of the knowledge of their domains and do more Interesting things with It. This means larger, more structured knowledge bases and inference mechanisms capabl? of manipulating the structures these knowledge bases contain. The necessarily large Investment In building these systems, and the very nature of some of the applications (e.g., data base query, cooperative interactive systems), also require these systems to be more adaptable than before to new domains within their purview (e.g., a new data base, a new Interactive tool). II RULE-BASED INFERENCE The need for adaptability argues strongly for perspicuity and modularity in the inference engine: the adapter must be able to see what must be changed (or added) and be able to make the alterations quickly. Inference mechanisms based on rules have these characterlstlcs. Unfortunately, most rule-based * Thus research was supported in part by the Defense Advanced Research Projects Agency under Contract No. DAHClS 72 C 0308, ARPA Older NO. 2223, and tn part by General Motors Research Laboratories. Views and conclustons contained in this paper are the author’s and should not be interpreted as representing the official opinion or policy of DARPA, the U.S. Qovsrnment or any person or agency connected with them. approaches rely on small, simply structured system knowledge bases (e.g., the rule-based formalism used In [lo] and [la] is dependent on representation In terms of triples). As rule-based systems grow to encompass a large number of rules, and as they are forced to work on complex knowledge structures to keep pace with modern knowledge base organizations, two major problems arise: o The inference mechanism becomes inefficient: It is hard to find the right rule to apply If there are many possibilities. o Rules begin to lose their properties of modularity and perspicuity: the “meaning” of a rule, especially In the sense of how It affects overall system behavior, becomes lost if the rule can interact with many other rules In unstructured ways. The remainder of this paper describes an approach to solving these problems based on a philosophy of doing inference that Is closely coupled with a principle of rule base organization. This approach is discussed in the context of two Implementation technologies. Ill PHILOSOPHY The inference methodology described here is a reaction to the use of rules as pattern/action pairs in a rule base that is not intimately related to the program’s knowledge baso. The basic philosophy [6] is to treat expert system inference as a process of redescription: the system starts with some kind of problem specification (usually a singlo user request) and redescribes It until It fits a known solution pattern (a data base access command or an implemented program function in the two examples described below). The control structure for this redescription depends on a knowledge representation that explicitly includes the Inference rules in the knowledge base. The redoscriptlon effort proceeds In two modes: a narrowing-down process that simply uses any applicable rules to redescribe the input as something more amenable to the program’s expertise; and a m-b process that 190 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. takes the description remaining from the first process (I.e., v&en no more applicable rules can be found), finds the most closely related solutlon pattern, and then uses that solution as a goal. In thls homlng-in mode, the inference procedure uses consequent rules to resolve the differences between the current description and the desked end. The narrowing-down phase focuses solely on the Input and transforms It lnto something closer to the system’s solution model. Depending on the input, the result of narrowlng-down might be an actual solution (meaning the Input request was quite close to the system expectations), or something that the system has no idea of how to deal with (meaning that it was very far from what was expected). Given that the user has some Idea of what the system can do, and that he wants to make himself understood, we can assume that the result of narrowing-down will often be only a minor perturbatlon of a system solution. This enables the homing-in process to find a closely related solution to use as a goal. Very speclflc consequent rules can then be used to resolve the few remalnlng differences. Assuming rule-based inference and a structured system knowledge base, the above phllosophy can be restated as follows: o Inference Is the transformation of one significant knowledge base structure into another. o The organization of the knowledge base organ&es the rule base. o The control structure supports both straight rule application anywhere in the knowledge base and consequent rule application In the context of a well defined goal. This approach makes rule application more efficient, even in a large knowledge base, because highly specific rules are not used--are not even looked for--until they have a good chance of success. That Is, the system does not have to look through everything it knows every time it must apply a rule. This efficiency is further enhanced by the fact that inference is modeled as redescriptlon and rules are tied closely to the system knowledge base. The inference mechanism will not even look at rules that do not relate directly to the kind of knowledge structures found in the description to be transformed. This makes the problem of finding the right rule to apply far less dependent on the number of rules In the system. The highly structured and detailed nature of the knowledge base now works for efficiency of rule application rather than against it. Separating the Inference process into modes and tying rules directly to the structures of the knowledge base also enhances the modularity and perspicuity of the rule Thls methodology has been used to Implement the inference components of two knowledge-based systems using quite different technologies. The first, a pat tern-match design, Is slmply sketched, while the second, a network based scheme, Is presented in more detail. . Inference In Asklt -- The Asklt system currently being developed at General Motors Research Laboratories (173, [S]) is a natural language data base query facility designed to be adaptable to new domains (e.g., a new project scheduling data base or a new inventory management data base). The inference task is to translate the user’s query Into the appropriate set of data base commands to provide the neodcd data. Askit’s knowledge base consists of case structures representing user request types, data base commands, individual English words, etc. Rules are expressed as transformations between system case frames. The condition part of the rule is a partially instantiated case frame that is treated as a pattern to be further Instantiated (Le., “matched”) by a description (a fully Instantiated case frame) from the program’s current state. The conclusion is a case frame to be filled on the basis of the instantiation of the condition part. When the rwlo is applied, the instantiating description is replaced In tho current state by the new structure generated from the conclusion part of the rule. Rules therefore represent allowable redescriptions of case frames for certain choices of case fillers. For example, the following is a rule for redescribing restrictions in certain user requests as SUBSET commands to the database system: 191 (REOUEST (OBJECT <table-data>:objI (RESTRICTION oobj (ccolumn-data>:property :relation :vaiue)) -> tSUBSt?T =obj WHERE -property w-elation =value) That is, the condition of the rule applies to any request which deals with something that is classified as “table data” in the system knowledge base, and which expresses a restriction of that table data in terms of “column data”. The conclusion states that such a request can be redescribed as a SUBSET command. When the rule is applied, the SUBSET command replaces the request In Askit’s current state. Rules are organized into “packets” based on the case frames In the knowledge base: all rules whose conditions aro partial instantiations of the same case frame are grouped in the same packet. For example, all rules that deal with restrictions In user requests would form a packet. The packet is represented by a packet pattern which states the common case structure dealt with by the rules of the packet, I.e., a generalization of their condition parts. The packet pattern for the ‘@restriction in rwquests” packet would be: (REClUEST (OBJECT :obj) (RESTRICTION nobj :rstr)) Packets play a key role in Askit’s rule application process. The process begins by matching a description in Asklt’s current state agains’t the packet patterns known to the system. If a match is found, the individual rules of the packet are tried. If one of the rules is successfully matched, the input structure is redescribed and placed back in the current state. If no rule matches, Askit goes into homing-in mode. It posits the discrepant part of a not-fully-matched rule from the packet as a new description in the current state, and looks for consequent rules to match this description. Consequent rules are also organized into packets (but based on their conclusions rather than their conditions). If Askit finds a matching rule in the consequent packet, and if the condition part of the consequent rule matches other descriptions in the current state, the discrepancy between the original description and the partially matched rule is considered to be resolved, and processing continues. Otherwise, other possible discrepancies or other partially matched rules are posited, and homing-in Is tried agaln. Thus, the narrowing-down process in this system Is represented by normal packet application, in which an lnltial description is matched against packets and successively redescribed until it can be seen as an instantiation of one or more system case structures representing data base commands. If this process breaks down, i.e., if a packet pattern is matched but no rule In tho packet can be matched, the input description is troated as a perturbation of one of the case structures rapresented by rule conditions in the packet. Consequent rules are then used to home-in the discrepant description on the deslred case structure. Successful resolution of the discrepancies allows normal processing (i.e., nerrowlng-down) to continue. t3, Inference in Consul -- The Consul system is being designed to support cooperative interaction between users and a set of online tools (for text manipulation, message handling, network file transmission, etc.). This “cooperative” interaction Includes natural language requests for system action and explanation of system activities. Since user requests may differ radically from the Input form expected by the tool designer, Consul’s inference task is to translate the user’s natural form of requesting system action into the system’s required input for actually performing that action. In the Consul system, the dependence of rule ropresentation and organization on the system knowledge base is carried much further than in Askit. On the other hand, the control structure does not have to be as complex. Consul’s knowledge base is a KL-ONE [2] topresentation of both tool-independent and tool-dependent knowledge. The major organizational framework of the knowledge base is set by the tool-independent knowledge, with the tool-dependent elements instantrating it. inference in Consul is a process of taking an exlstlng description in the knowledge base, redescribing it In accordance with an inference rule, and then reclassifying * It In the knowledge base. If the description can be classified as the invocation of an actual tool function In the system, then the description is “executable”, and inference is complete. A key aspect of Consul inference is that inference rules are also rspresented in KL-ONE in the knowledge base; the same process of classification that determines the status of a description also determines the applicability of a rule (compare 153). For example, let us examine Consul’s treatment of the user request “Show me a list of messages.” Parsing (using the PSI-KLONE system [l J) results In the structure headed by ShowAct. in figure 1. Consul classifies thls structure in its knowledge base, finding, as shown in figure 1, that it is a subconcept of the condition of Rulcl. This means that Consul can redescribe the request ShowAct. as a call to a “display operation” according to the conclusion of Rulel. The result is a new description, DisplayOperationlnvocationl .l, which Consul then classifies in Its knowledge base. If, via this classification process, the new descrlptlon Is found to be B subconcept of an executable function, inference Is “c The classification algorithm was written by Tom Lipkis. 192 complete--DlsplayOperationlnvocationl .I can simply be invoked by the Consul interpreter. Otherwise Consul will have to use additional rules to further refine the description until it can be seen as an actual call on 8ome other function or functions. (ab]ect) (mclpiant) Figure 1: Rule Application in Consul Figure 2 shows the classlflcatlon of DlsplayOperatlonlnvocatlonl .I in Consul’s knowledge base. Unfortunately, the system does not know very much about it In this posltion: it is not a subconcept of an executable function (shown shaded), nor Is It subconcept of any rule condition. Since no applicable function or rule can handle the current description, the system seeks a “closely related” function description to use as a mapping target. The deflnltlon of Wosely related” Is that the two descriptions must share a common ancestor that is a @basic” concept On Consul% actual ., the ancestor cannot be rule Itself, nor newly generated description. the function d P)lsplayMessageSummeryl ocation is closely related to the current description because they share an appropriate lsplayOperationlnvocation. Once a related tool function description Is found, It Is used as a goal for consequent reasoning. First, Consul must find the discrepancies between the current description and the desired result. These discrepancies re simply the differences that escrlption from being classified as desired description: In this example, DisplayOperationlnwocationI .I that prevent It from instantiating DlsplayMcssageSummerylwwoc ThW discrepancy is that the “input” role of the current ption is filled wlth list of messages ageLIst. f ), while the executable function yMessageSummerylnwocation requires list of mmarles (Summaryhist). Figure 2: Flndlng B Closely Related Function Description y to redes If It hope iatlon of P executable function. ere are two ways the current description: rules and executable functions (since functions produce new descriptions via output and side-effects). Rules are preferable because they save tool execution time; Consul therefore looks for rules first. In this case, there is a quite general rule that produces the desired effect. Users frequently ask to see a list of things (messages, files, etc.) when they really want to see a list of summaries of these things (surveys, directories, etc.). Consul therefore contains a rule to make this transformation, if necessary. The rule, shown as Rule2 in figure 3, says that if the current description is a display operation to be invoked on list of “summarizable objects”, then it can be redescribed as a display operation on a list of summaries of those objects. Consul finds this rule by looking In the knowledge base for rules (or executable functions) that produce the “target” part of the discrepancies found earlier. As shown in figure 3, Rule2 can be found by looking “up” the generalization hierarchy (not all of which is shown) from SummaryLIst. Consul must next be sure that the condition part of the rule is met in the current state of the knowledge base (this Includes the current description, descriptions left by previous rule applications and tool function executions, and the original Information in the knowledge base). if the condition is not satisfied, Consul will try to produce the needed state through further rule application and function 193 application of the execution--i.e., through recursive consequent reasoning process. In this example, the condition of Rule2 is entirely met in the current description (see figure 3). Furthermore, the conclusion of Rule2 resolves the entire discrepancy at hand. In general, a combination of rules and functions is needed to resolve discrepancies. Therefore, with the application of Rule2, Consul has successfully redescribed the initial user request as an executable function. The Inference process passes the resulting description on to the interpreter for execution, and the user’s request is fulfilled. Narrowing-down and homing-in are much the same In Consul as they are in Askit. In Consul, however, relationships between rules and notions such as “perturbation” and Iccfosely related structure” come directly from the existing knowledge base representation; a superimposed packet structure is not necessary. The control structure of rule application therefore need not consider the separate issues of matching packets, setting up packet environments, matching rules, etc. Instead, classification includes matching, as applicable rules are found “automaticallyl when a newly generated description Is put in its proper place in the knowledge base. Finding related knowledge structures and consequent rules are aI90 classification problems, representing only different use of the classification algorithms. a slightly b DisplryOporatlon Invocatlonl.1 summarized object) Rusty Bobrow and Bonnie Webber, “F’SI-KLONE: Parsing and Semantic Interpretation in the BBN Natural Language Understanding System,” in Proceedings of the 1980 Conference of the Canndlan Sociely for Computational Studies of /nte//igonco, CSCSI/SCEiO, 1980. Ronald Brachmen, A Structural Paradigm for Reprsscntlng Knowledge, Bolt, Beranek, and Newman, Inc., Technical Report, 1978. Bruca Buchanan, of al., Heuristic DENDRAL: A Program For Generating Explanatory Hypotheses In Organ/c Chemistry, Edinburgh University Press, 1960. Ranclall Davis, Applications of Meta Level Knowledge to the Construction, Maintenance and Use of Large Know/edge Bases, Stanford Artificial Intelligence Laboratory, Technical Report, 1976. Richard Fikes and Gary Hen&ix, “A Network-Based Knowledge Representation and it:; Natural Deduction System,” in Proceedings of the Fifth lnternatlonal Joint Conference on Artificial Intelligence, IJCAI, 1977. William Mark, The Reformulation Model of Expertise, MIT Laboratory for Computer Science, Technical Report, 1976. William Mark, The “Asklt” English Database Query Facility, General Motors Research Laboratories, Technical Report GMR-2977, Jim3 1!379. William Mark, A Rule-Based Inference System for Natural Language Database Query, General Motors Research Laboratories, Technical Report GMR-3290, May 1980. William Martin and Richard Fateman, “The MACSYMA System,” Proceedings of the Second Symposium on Symbolic and Algebraic Manipulation, 197 1. Edward Shortliffe, MYCIN: Computer-Based Medical Consultations, American Elsevier, 1976. William Swartout, A Digitalis Therapy Advisor with Explanations, MIT Laboratory for Computer Science, Technical Report, February 1977. Figure 3: Using Consequent Reasoning 194	1980	50
46	N PROCESS FOR EVALUATING TREE-CONSISTENCY John L. Goodson Departments of Psychology and Computer Science Rutgers University, New Brunswick, N.J. 08903 ABSTRACT General knowledge about conceptual classes represented in a concept hierarchy can provide a basis for various types of inferences about an individual. However the various sources of inference may not lead to a consistent set of conclusions about the individual. This paper provides a brief glimpse at how we represent beliefs about specific individuals and conceptual knowledge, discusses some of the sources of inference we have defined, and describes procedures and structures that ' can be used to evaluate agreement among sources whose conclusions can be viewed as advocating various values in a tree partition of alternate values. * I. INTRODUCTION Recent work by several investigators; [II, C31, C4l and C7l; has focused on the importance of augmenting deductive problem solvers with default knowledge. Their work provides some of the logical foundations for using such knowledge to make non- deductive inferences and for dealing with the side effects of such inferences. Currently, we are pursuing how conceptual knowledge about general classes of persons, locations, objects, and their corresponding properties can be represented and used by a planning process and a plan recognition process to make deductive and non-deductive inferences about particular persons, objects and locations in an incompletely specified situation (see C51 and C61). General knowledge about conceptual classes represented as a concept hierarchy provides a basis for various types of inferences about individuals. The definition of a conceptual class might be believed to hold for an individual, xl, that is believed to be a member of that class (Definitional Inference). The definitions of concept classes which include the class of which xl is a member might be believed to hold for x 1 (Inheritance Inference). The definition of some class that is a subset of the class to which xl belongs might be -------- * This research is supported by Grant RROO643-08 from the Division of Research Resources, BRP, NIH to the Laboratory for Computer Science Research, Rutgers University. used as a source of potential inferences (a kind of Plausible Inference). Additionally, information might be stored directly about xl (Memory Inference) and there may be other inference types based on different strategies of deductive or plausible inference (for a more detailed discussion see [61). However, these sources of inference may not lead to a consistent set of conclusions about the individual. For default theories in general, Reiter [4] has shown that when a default theory is used to extend a set of beliefs, determining the consistency of the extensions is an intractable problem. In this paper we are concerned with a very local and focused subproblem involved in evaluating the agreement or consistency of a set of conclusions and with a strategy for dealing with belief inconsistency. The focus arises from considering these issues in the context of a concept class hierarchy where the classes form a tree partition. Before we discuss this restricted case of belief consistency, we provide a brief glimpse at how we represent beliefs about specific individuals and at several types of class hierarchies used to represent concepts. II. REPRESENTATION a BELIEFS AND CONCEPTS Beliefs about specific objects, persons, etc. are represented as binary relations of the form ((x r y).CT F or Ql) where : r is a relation defined between two basic classes of entities X and Y; x is an instance of X; and y is either an instance of Y or is a concept that is part of the concept hierarchy with Y as its root. An example of the former relation is ((DON LOC NYC).F) where DON is an instance of PERSON and NYC is an instance of LOCATION. The latter form is exemplified by ((DON AGEIS YOUNG1.T) where YOUNG is a concept that is part of an AGE hierarchy. T, F or Q represents the truth value in the current situation. Concepts are organized into inclusion hierarchies which have as their root one of several basic classes, such as PERSON, AGE, LOCATION, OBJECT, which may have instances in a specific situation. A particular individual may be an instance of several basic classes, e.g. a person DON may be viewed as an OBJECT or a PERSON. Two simplified hierarchies are given below in graph form. Note that no claim is being made here about the adequacy or naturalness of the knowledge represented in these examples. 195 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. PERSON III. CONSISTENCY m INFERENCE SOURCES AGE I I ------------------- ------- We now consider structures that can be used t.o I I f t determine the consistency of a set of sources STUDENT ATHLETE YOUNG OLD contributing beliefs relevant to a proposition I I I I about an individual. The following paradigm -----mm---- ---------- m-w- ---- provides a more specific context in which to I I I I I I I I I I discuss the problems and mechanisms we have HIGH SCHOOL COLLEGE BASEBALL FOOTBALL I5 25 40 80 considered. Assume that the task specification is: STUDENT STUDENT PLAYER PLAYER YR YR YR YR 1) A goal proposition is given whose truth value is Hiearchies may contain subtree partitions as might desired by a higher level process. The goal is be the case for the PERSON hierarchy shown above. a statement about a particular individual A particular person may be a high school student represented as a binary relation between the &a football player. Some hierarchies, individual and a concept, called the goal- especially those representing properties, may form target. The goal-target is a member of a tree tree partitions, as is the case for the AGE partition of concepts called the target-tree. hierarchy shown above. This '1family11 of concepts is the set of potential targets defined for the relation There are several types of information that occurring in the goal. It should be noted that can be associated with a concept. One type, the there may be several trees rooted in the same concept definition, provides an intensional basic class. For now we limit our consideration characterization of the elements of the conceptual to the case where there is only one tree. class. A definition is a conjunctive set of descriptions of the form (<basic 2) Several sources of inference are consulted for class> (rl yl)(r2 zl)...) that are the necessary beliefs relevant to the goal, that is, beliefs and sufficient conditions for an instance of the relating the individual to concepts in the basic class to be considered an element of the target-tree. conceptual class. The relation that represents the hierarchical structure among concepts is termed 3) Since the target-tree is a partition, the "tree- COVERS and appears in the above example as a line consistency" of the beliefs can be evaluated and joining pairs of concepts. This relation implies used to determine a truth value for the goal that for each description in the higher level (one of T for true, F for false or Q for concept definition (e.g. (rl yl)), there is a question, determinate T or F not assignable). corresponding description -(e.g. (1-1 ~2)) in the This evaluation amounts to assessing the lower level concept definition and either yl=y2 or agreement among the sources on a truth value for (~1 COVERS ~2) in the Y (basic class) hierarchy and the goal. the lower or more specific concept definition must contain at least one description that is more 4) The structures created in 3 can be used to specific (where (yl COVERS ~2)) than the record the sources of inference drawn upon in corresponding description in the more general the attempt to achieve a conclusion about the concept definition. As an example, consider these goal. possible definitions for the concept ATHLETE and Many of the structures and procedures discussed in and a more specific concept, FOOTBALL-PLAYER. relation to this paradigm are implemented in the knowledge representation system AIMDS (see [ 83). (ATHLETE DEF I (PERSON (PHYSICALSTATE SOUND) The target-tree can be represented as a Truth I (PLAYSON TEAM)) Value Tree (TVT) where each node represents a COVERS1 concept in the target-tree. Each node has a slot, (FOOkBALL-PLAYER DEF TV, for one of the determinate truth values, T or F, and slots for two lists, a true-list and a (PERSON (PHYSICALSTATE SOUND> false-list. These lists consist of two inner (PLAYSON FOOTBALL-TEAM)) lists. The true-list contains one list for recording the sources that support the truth value T and one for recording nodes that require the node An example of a type of plausible inference to have the truth value T. The false-list has the can be given using these definitions. If the same structure and records the information relevant beliefs in memory about DON satisfy the to the truth value F. An example will serve to descriptions in the definition of ATHLETE, then clarify how this structure is used to evaluate there ' basis for believing (DON PLAYS;: FOOTBiLL-TEAM) with truth value T. tree-consistency. However, it is not the case that this inferred Assume that the goal is (DON AGEIS 40YR). The belief is consistent with those in memory. Thus TVT for the goal represents the AGE tree presented two sources of information about DON, memory and in a previous example. Initially each node in the plausible inference, may not agree. TVT has three empty slots shown below. 196 slots (TV NIL) Truth Value (T 1 True-list (F Ii 1 False-list supported required by by If ( (DON AGEIS YOUNG) .T) is contributed by memory, the following cycle of actions takes place. 1) The truth value slot of each node is made NIL. 2) T is entered as the TV of the node representing YOUNG. 3) Tree-consistency rules propagate truth value requirements to other nodes. Note that. these rules depend on interpreting the AGE hierarchy as a tree partition of a finite and closed set of values. One rule propagates T up the tree ( to all ancestors), thus the AGE node receives the truth value T. Another rule gives the node OLD the truth value F since the relation AGEIS has been defined such that only one path (leaf to root) of targets in the tree can be true for an individual at any given time, and all others must be false. Finally , a third rule propagates F down the tree (to all descendents) , thus the nodes 40YR and 80YR receive the truth value F. In the general case, these rules are looped through until no additional nodes can be given a truth value. Two other rules not applicable here are: propogate T from a parent to a daughter if all other siblings are F; and if all daughters are F, propogate F to the parent. 4) The source of the truth value T for the node YOUNG, i .e. Memory Inference (MI), is registered as support on its true-list. For each of the remaining nodes with a non-null TV, YOUNG is registered as a requirement on the true-list or false-list depending on the truth value required by YOUNG being T. At the end of this cycle, the nodes in the AGE TVT have the following form. YOUNG OLD TV T TV F (T (MI)()) (T 00) (F 00) (F O(YOUNG)) 15YR 25YR 4OYR 80~~ TV NIL TV NIL TV F TV F (T (10) (T 00) (T 00) (T 00) (F 00) (F 00) (F O(YOUNG)) (F O(YOUNG)) This cycle can be repeated for each of the relevant beliefs that have been contributed. Note that a cycle generates the deductive consequences of a truth value assignment, T or F, to a single node. We are not concerned with the propagating changes to previously assigned truth values based on a new assignment from a second relevant belief. Such a mechanism would be required for updating a model and might utilize antecedent and consequent propagation proposed by London [ 21. The TVT created by this cycle is inspected for a truth value for the goal. In our example, the node 40YR represents the goal-target and it has an entry for a single truth value, F. Thus F is returned as t,he truth value of the goal. However, the assignment of a truth value to the goal can be made contingent on the target-tree. Tree- consistency and thus agreement among the sources is easily determined. If any node has entries in both the true and false lists, then an inconsistency exists and the value Q (indeterminate truth value) should be returned even if a determinate truth value is indicated for the goal -target. Conversely, if no such node can be found, then the tree, and thus the set of beliefs contributed, are consistent and the truth value indicated for the goal-target may be returned. One way to find a det,erminate truth value for the goal when the tree is inconsistent is to seek a subtree of TVT such that : I) the subtree is rooted in the top node; 2) the subtree is tree-consistent; 3) the subtree contains the node representing the goal-target; and 4) this node has a true-list or false-list entry, then a determinate value can be assigned to the goal. An intuitive example of this case is where you can be fairly sure that Don is young even though you have conflicting information about whether he is ten or fifteen years old. If the consistent subtree does not indicate a determinate truth value for the node representing the goal-target, the tree-consistency requirement can be relaxed in order to find a determinate truth value. The consistent subtree can be extended sue h that the resulting subtree is tree-consistent, each node has a determinate trut,h value and the set of truth values is maximally supported by the sources contributing information. One extension procedure involves the following steps : 1) For each node in the consistent subtree assign the truth value indicated by its non-empty list. 2) From the set of nodes without a truth value (TV NIL), select the node with the maximum support for a truth value that is tree-consistent with the current subtree. 3) Assign this node the truth value indicated and apply the tree-consistency rules to further extend the consistent subtree. 4) Continue at 2 until all nodes have a truth value or the basis for deciding 2 does not exist (some 197 nodes may have empty support and cannot be assigned a determinate truth value by the rules). Several issues must be addressed in carrying out step 2. First, since ties are possible, a decision procedure must be provided. Second, the degree to which a node is required to have a particular truth value might be taken as a measure of indirect support for that truth value. Since the set of nodes that could possibly require a node to have a truth value is partially dependent on its position in the tree, positional bias must be taken into account in deciding degree of indirect support. This procedure can be applied even when a consistent subtree cannot be found. In this case the top node is given the truth value T to provide a trivial consistent subtree from which to extend. The staged relaxation of the tree-consistency constraint is particularly important when all of the beliefs relevant to a goal are drawn from default knowledge (arrived at through non-deductive inferences). There may be little basis for expecting this knowledge to be consistent yet it may be rich enough to suggest that one truth value is more plausible than the other. ACKNOWLEDGEMENTS. My collaborators, C.F. Schmidt and N.S. Sridharan, have contributed to the evolution and implementation of the ideas expressed here. REFERENCES. Cl] Doyle, J. A Truth Maintenance System. Artificial Intelligence 12 (1979) 90-96. [2] London, P. A Dependency-based Modelling Mechanism for Problem Solving. AFIPS Proc. Vol. X NCC-78, Anaheim, Ca., (June, 1978) 263-274. [3] McDermott, D. & Doyle, J. Non-Monotonic Logic I. Proc of the Fourth Workshoo on Automated Deduction. Austin, Texas, (Feb., 1979) 26-35. [4] Reiter, R. A Logic for Default Reasoning. Technical Report 79-8, Dept. of Computer Science, University of British Columbia, (July, 1979). 151 Schmidt, C.F. The Role of Object Knowledge in Human Planning. Report CBM-TM-87, Dept. of Computer Science, Rutgers University, (June, 1980). t.71 [81 c91 Shrobe, H.E. Explicit Control of Reasoning in the Programmer's Apprentice. Proc - oft&& Fourth WorkshoD on Automated Deduction. Austin, Texas, (Feb., 1979) 97-102. Sridharan, N.S. Representational Facilities of AIMDS: A Sampling. Report CBM-TM-86, Dept. of Computer Science, Rutgers University, (May, 1980). Sridharan, N-S., Schmidt, C.F., & Goodson, J.L. The Role of World Knowledge in Planning. Proc 1980). AISB-80. Amsterdam, Holland, (July, C63 Schmidt, C-F., Sridharan, N.S. & Goodson, J.L. Plausible Inference in a Structured Concept Space. Report CBM-TR-108, Dept. of Computer Science, Rutgers University, (May, lg8o). 198	1980	51
47	Reasoning About Change in Knowledgeable Office Systems Gerald R. Barber Room 800b Massachusem Ittslitute of Techtlologv 54.5 Techttologv Square Cambridge, Mass. 02139 (617) 253-5857 ABSTRACT Managing of and reasoning about dynamic processes is a central aspect of much activity in the office. WC present a brief dcscrintion of our view of office systems and why change is of central importance in the office. A description system used to describe the structure of the office and office activity is discussed. A viewpoint mechanism within the description system is prcscntcd and we discuss how this mechanism is used to dcscribc and reason about change in the offke. A gcncral scenario .is dcscribcd in which viewpoints arc illuslratcd as a meiins of describing change. Previous technologies for accommadating change in knowlcclgc embedding languages are charnctcrizcd. WC contrast the approach using viewpoints with previous technologies whcrc change is propagated by pushing and pulling information bctwccn slots of data structures. I. Int reduction The computer has been used in the office cnvironmcnt for many years with its application mainly limited to highly structured and repetitive tasks in a non-interactive mode. With the cvcr-dccrcasing cost of hardware computers can bc potentially used in the future to aid office workers in a wider variety of tasks. Indeed, the computer based office system is today seen as both a motivation for achieving a new understanding of office work and as a medium within which to intcgratc new tools and knowledge into a cohcrcnt system. This has led to the realisntidn that there is enormous potential in the use of the computer in the office in novel anti as yet unforcscen ways. These new uses will impact the way office work is done in fundamental ways demanding new ideas about how to manage information in an office and a new conceptualization of what off& work is in the presence of powerful computational capabilities. As a step toward developing computer systems that may effcctivcly support office workers in their tasks we employ two paradigms from Artificial Intclligencc, those of knowledge embedding and problem solving. We arc dcvcloping ;I descriptions system called OMEGA [IIcwitt SO] to bc used to cmbcd knowlcdgc about the structure of the office ilnd office work in an oflicc system. One of our objcctivcs is to support the problem solving activities of individuals in an office. Much of the problem solving activity within an office concerns reasoning about change. WC hnvc dcvclopcd mechanisms in OMEGA to dcscribc changing situations. In the following section we present a short description of our model of the office. Following this WC discuss the importance of ch,mgc in the office and the mechanism within OMEGA to deal with change. ‘I’hc approach WC take is compared with other approaches to the problem of accommodating and managing change. II. The Knowledgeable Office System We view an office system in terms of the two dominant structures in the office, the applicalion s(nlcfure and the organiza~iunal struclure. l’hc application structure of an office system concerns the subject domain of the office. It includes the rules and objects that compose the intrinsic functions of a particular office system. As an example, in an office concerned with loans the application structure includes such cntitics as loan applications. credit ratings and such rules as criteria for accepting or rejecting loans. In an insurance company the application structure is concerned with insurance politics, claims and actuarial tables. ‘Ihc application structure explains the scope of the functionality an offrcc system has on a subject domain as well as providing a model by which those functions arc charncteriir.cd. OVCI tly, the application structure is the primary reason for the existence of the office system. In contrast to the application structure is the social structure of the office system as an organi/.ation [Katz 781. Our concern with this aspect of an office system stems from the fact that the activity in the application domain of an ofl?cc system is IcalLed by people coopcrating in a social system. ‘I’hc structure of this social system involves such aspects of an organization as the roles of the individual participants. the interaction of roles, the social norms of the office and the various subsystems that make up the organization. WC bicw the office system as a functioning organism in an environment from which it extracts resources of various kinds and to which it dclivcrs the products of its mechanisms. OMEGA’s descriptions arc the hlndamental entities upon which the Knowlcdgcnblc Office System is based. ‘I’hc emphasis of our approach is on a description manipulation system for cmbcdding knowledge as opposed to a forms manipulation system. Descriptions are used to express the relationships bctwcen objects in the Knowledgcablc Office System. Descriptions are more fimdamental than electronic forms, in particular, electronic forms arc a way of viewing descriptions, a visual manifestations of descriptions. One of the goals of our work is to support office workers in their problem solving activity. Problem solving is a pervasive aspect of office work that has been ncglcctcd until recently [Wynn 79, Suchman 791. Office work is naturally characteri/.cd as goal oriented activity. ‘The office proccdurc is mcrcly a suggested way by which to accomplish a particular goal. WC believe that this is one reason why it has proved to bc difficult to dcscribc office work from a procedural point of view. ‘I‘hc formalism WC are developing allows us to describe and reason about the application and organizational structures of office systems as well as the inlcraction bctwccn these structures. ‘I’hc major bcncfits of OMEGA with relevance to our discussion here arc that a computational system can support problem solving in dynamic cnvironrncnts that are weakly structured, and knowledge rich. OMEGA also provides a prccisc Lmgu,igc within which to characterize the static and dynamic aspects of office systems. A central problem in an office system is reasoning about and managing change. This is a rccul rent thcmc at several lcvcls of 199 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. abstractions. Viewing the organization in relation to its environment, the organization must cvolvc in order to adapt to the changing cnvironmcnt. For example. an accounting off~cc must adapt to new tax laws or an office must adapt to new technology. Viewing the organization as producer of some product, the organization must adjust its production output to the demand for t.hc product which it produces in light of the rcsourccs available to the organitation and the constraints under which it must operate. The individuals that make up an organization arc faced with such tasks as reasoning about processes that have produced anomalous results, maintaining system constraints as the state of the constrained parts change and analyzing the implications of hypothesized processes. OMEGA has a viewpoint mechanism that is used to describe and reason about change. ‘I’hc viewpoint mechanism provides a means to present time varying proccsscs to office workers for analysis, bc these proccsscs historical, in progress or postulated. Cltanging cnvironmcntal dcpcndcncics and changing aspects of the organization can be captured in descriptions using the viewpoint mechanism. In the remainder of this pnpcr WC dcscribc OMEGA and its viewpoint mechanism. IV. The Viewpoint Mechanism OMEGA is a system with which a structure of descriptions is built. ‘1%~ system is dcsigncd to bc incremental; new knowlcdgc can be incorporated into the system as it is discovered or as the need for it arises. There is no minimal amount of information nccdcd bcforc the system is usable. The system is monotonic in the sense that nothing is lost when new information is added. As is cxplaincd in the following paragraphs, knowlcdgc is rclntivir.cd to viewpoints, information that is inconsistent with information in a particular viewpoint can be placed in a different viewpoint. ‘This ac&nmodntcs aspects of non-monotonic systems [McDermott 79]--where new information may invalidate previously held beliefs--without the need for a notion of global consistency. OMEGA’s fundamental rule of inference is mergitzg; new descriptions are merged with previous descriptions. Any new deductions as a result of the new information arc carried out during the merging operation. OMEGA is used to build, maintain and reason over a lattice of descriptions. Descriptions are related via an inhcritancc relation called the is relation. ‘I’he is relation is rclativized to a viewpoint that indicates the conditions under which the is rcl,ltion holds. Intuitively a viewpoint represents the conditions ulider which the inheritance relation holds. In this rcspcct it is rcminisccnt of McCarthy’s situational calculus [McCarthy 691 and the contexts of QA4 [Rulifson 721. A major differcncc between these approaches and viewpoints is that viewpoints arc descriptions and thus are subject to the full descriptive power of OMEGA. Viewpoints may be cmbcddcd in structures expressing complex inheritance relationships relating viewpoints to one another. Other aspects of OMEGA include higher order capabilities such as the ability to describe properties like transitivity for relations in the system and meta-description capabilities to talk about the parts of descriptions. Ill. Dealing With Change A key property-of viewpoints is that information is only added to them and is never changed. Consider, for example, a description which is the underlying representation of a form. I’hc description is rclativi7cd to a kicwpoint and information is added to this description increasing its specificity. JIcscriptions may contain constraints bctwecn attributes, as information is added further information may be deduced. Should the information in a field of a form be changed then the following scenario might occur: 1. A new viewpoint is created successor to the old viewpoint. and dcscribcd as being a 2. All information that was not derived from information is copied to the new viewpoint. the changed 3. The new information is added in the new viewpoint, deductions resulting from this information arc made. 4. ‘J’hc description in the new recent contents of the form. viewpoint rcprcscnt the most In this case the new viewpoint inherits all but the changed information and the information deduced from the changed information from the old viewpoint. What actions arc taken when information in a viewpoint is changed is specified via mcta-descriptions. Previous approaches to the problem of accommodating changing information have been to perform updntcs to the data structures in question. System based on property lists such as J,lSP have used pur and gel operations to update and read database information. ‘I’hcsc hate the disadvantage that deductions based on updated infonnation must bc handled explicitly leading to unacccptablc complexity and modularity problems. J.angungcs like J:RJ* [Goldstein 771 use triggers on data structure slots to propagate changes. ‘Ilrc disadvantage here is that thcrc is no support for keeping track of what was deduced and why. ‘This makes changes difficult because information dependencies arc not recorded. ‘lhc language K RI, has been used to implement a knowledge-based personal assistant called OlIYSSJ<Y [Fikes X0]. ODYSSlXY aids a user in the planning of trips. In this system pushers and pullers arc used to propagate deductions as a result of updates and to mnkc deductions on reads. A simple dependency mechanism is used to record information dcpcndcncics. In this cast it is necessary to be very careful ilbout the order in which triggers fire for as updates arc made there is both new and old information in the database making it difficult to prcvcnt anamolous results due to inconsistencies. OMEGA separates new and old information into diffcrcnt viewpoints. Information consistency is maintained within viewpoints. ‘l’hc propagation of information bctwccn viewpoints is controlled via mcta-description. An advantage of the approach using viewpoints is that the system has a historical character. This is an important step toward our goal of aiding office workers in problem solving about dynamic proccsscs. Viewpoints can bc used as historical records of past processes, as an aid in tracking ongoing proccsscs and as an aid to dctcrminc the implications of postulated actions. 200 VI. Conclusion We have presented the viewpoint mechanism of the descriptions system OMEGA along with some examples of its use to describe a changing form in an accounting office. The viewpoint mechanism has proved useful in describing objects whose properties vary with time as well as a means with which to interpret statements about the system’s description structure. The viewpoint mechanism presented here is related to that in IYHER [Kornfcld 791 and to the layers of the PIE system [Goldstein 801. Viewpoints arc a powerful unifying mechanism which combine aspects of McCarthy’s situational tags [McCarthy 691 and the contexts of QA4 [Rulifson 721. They serve as a replacement for update and pusher-puller mechanisms. Omega is monolottic using merging of descriptions as a fundamental rule of inference. It uses viewpoinls to keep track of different possibilities. This aspect causes it to differ substantially from systems based on property lists [IPL, Lisp, etc.] which arc based on operations to yul and get attributions in data structures. These differences carry over to more recent systems [SIR, SIMUI,A, FRL, KRL, etc.] based on record structures with attached procedures that execute when a plr( (rrytkt&) or ger (rettcl) operation is performed. References \Fikes 801 Richard Fikes. Odyssey: A knowledge-Based Assistant. To appear in Artificial Intelligence. [Goldstein 771 Goldstein, I. P. and Roberts, R.B. NUDGE, a Knowledge-Based Scheduling Program. Proceedings of the Fifth In~erna(ional Joint Cotlference on Artificial Intelligence. [Goldstein 801 Goldstein, Ira. PIE: A Network-Based Personal Information Environment. Presented at the Off~cc Semantics Workshop, Chatham, Mass. June 15-18 [Hewitt 801 Hewitt, C., Attardi, G., and Simi, M. Knowledge lkbedding wirh a Descriplion System. AI Memo, MIT, August, 1980. to appear [Katz 781 Katz, D. and Kahn, R. The Social Psychology of Organizations. John Wiley and Sons, 1978. [McCarthy 691 McCarthy, J. and Hayes, P. J. Some Philosophical Problems from the Standpoint of Artificial Intelligence. In Machine Inrelligence 4, pages 463-502. Edinburgh University Press, 1969. [McDermott 791 McDermott, D. and Doyle, J. Non-Alonoronic I,ogic I. AI Memo 486b, MIT, July, 1979. [Rulifson 721 Rulifson. J., Derksen, J. and Waldinger, R. QA4: A Procedural Calculus for lnluilive Reasoning. Artificial lntclligcncc Center Technical Note 73, Stanford Research Institute, November, 1972. [Suchman 791 Suchman, L. Office Procedures as Praclical .4ction: A Case Sludy. Technical Report, XEROX PARC, Scptcmber, 1979. [Wynn 791 Wynn, E. Office Conversaliott as an Infortnaiion Medium. Phi) thesis, Department of Anthropology, University of California, Berkeley, 1979. [Kornfeld 791 Kornfeld, W. Using Parallel Processing for Problem Solving. AI Memo 561, MIT, December, 1979.	1980	52
48	On Supporting the Use of Procedures in Office Work Richard E. Fikes and D. Austin Henderson, Jr. Systems Sciences Laboratory Xerox Palo Alto Research Center 3333 Coyote Hill Road Palo Alto, California 94302 Abstract In this paper, we discuss the utility of AI techniques in the construction of computer-based systems that support the specification and use of procedures in office work. We begin by arguing that the real work of carrying out office procedures is different in kind from the standard computer science notions of procedure “execution”. Specifically, office work often requires planning and problem solving in particular situations to determine what is to be done. This planning is based on the goals of the tasks with which the procedures are associated and takes place in the context of an inherently open-ended body of world knowledge. We explore some of the ways in which a system can provide support for such work and discuss the requirements that the nature of the work places on such support systems. We argue that the AI research fields of planning and knowledge representation provide useful paradigms and techniques for meeting those requirements, and that the requirements, in turn, present new research problems in those fields. Finally, we advocate an approach to designing such office systems that emphasizes a symbiotic relationship between system and office worker. Introduction We are interested in developing office systems that would make use of a knowledge base describing whal tasks are to be done, who is to do them, and how they are to be done. Such descriptions specify the functions of an office and how it is organized to achieve that functionality. We claim that such a knowledge base can form the basis for a broad range of system support in an office. In this paper, we discuss some of the ways in which AI paradigms and techniques are relevant to the support of office work by such computer-based systems. We begin by describing some of the support functions we have in mind, and then address what we consider to be the primary issue; namely: what is the nature and structure of the information in such a knowledge base? We are guided in addressing that issue by considering the nature of the work that occurs in an offrce and how such information is used in that work. We first argue that the work involved in carrying out office procedures is different in kind from the “execution” of a procedure that one might expect by drawing analogies with the behdvior of a computer executing a progmm. We illustrate and support this claim by presenting a typical case of office work and analyzing the actions that take place there. From this argument we derive a requirement for systems which support office work: namely, that they be flexible enough to support the variety of behavior occasioned by the unpredictable details of particular situations. We then turn to the relevance of AI for achieving this functionality. We develop the idea that the paradigms from the AI literature for automatic planning and execution monitoring of plans provide a more suitable alternative to the procedure execution model of office work: and furthermore that the demands of supporting office work require extensions to those paradigms. Second, we argue that the knowledge representation problems presented by the open-ended office domain are unsolved and challenging. We suggest that they can be attacked by the use of specialization-based representations and facilities for storing “semi-formal” structures in which uninterpreted text is intermixed with data whose semantics is understood by the system. Finally, we argue that the whole enterprise of supporting office work can only hope to succeed if we regard the office systems as functioning in a partnership with the office workers. Due to the open-endedness of the domain, the system cannot hope to “understand” the full import of the information which it is handling, and so must rely on human aid. Furthermore, to fully support the users, the system must be able to represent, although not necessarily understand, any of the infcrmation in the domain. We conclude by advocating an approach of “symbiotic processing” between system and office worker and the use of AI techniques in constructing systems to support office work. Supporting the Production and Use of Procedural Descriptions We begin by considering some of the whys in which computer-based office systems could facilitate the effective production and use of descriptions of what tasks are to be done, who is to do them, and how they are to bc done. There are two groups of people whom an office system dealing with such descriptions can support: the producers and the users. However, the production and LW phases are often tightly interwoven, with the same people often involved in both (despite what managers may choose to think). The producers of these “what, who, and how” specifications <typically managers and planners) are engaged in a process of organizing the work in the office so that the office’s goals and commitments will be met. That process involves defining the tasks to be done, designing procedures for doing those tasks, and assigning individuals to carry out those procedures. A system can support these specification processes by providing a descriptive namework in which to express the specifications and by helping to manage the complexity that arises from the interactions of the tasks. constraints, procedures, and policies being specified. The descriptive. framework would provide a guide as to what information needs to be specified (based on the intended purpose and uses of the specifications) and a terminology for expressing that information. For example, the system might provide a template for describing a task that would include fields for the task’s goals, inputs, outputs, responsible agent, activation conditions, etc., and a description language for filling those fields. The system could also indicate direct implications of a description, such as the subtasks implied by a task description of recognizing the task’s activation events, obtaining the task’s inputs, or communicating its outputs. The system could aid in managing the complexity of the specifications primarily by monitoring interface requirements atnong interacting components to help assure that those interfaces are well specified and the requirements are met. For example, if the description of a task included the source of each of the task’s inputs and the destination of each of its outputs, then the system could alert the specifier when those input-output connections between tasks are inconsistent (e.g., when some input is not an output of the 202 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. source task), and could input-outputs prompt for a specification are to be communicated. of how The other grotto that the system can support is the users of the “what, who. how” specifications. -. That support would include facilities for accessing the specifications in useful ways, for adding infarmal notes to the specifications, for monitoring the carrying out of tasks, and for doing some of the procedural steps. One useful way in which the system would act as an information source is in providing “how-to” information to a person who has a task to do and doesn’t know how to carry it out. For example, when a secretary is given a task that he is not familiar with (such as “order some business cards for me” or “obtain a consulting contract and arrange the travel plans for this person”), the system could provide him with a description of what needs to be done, how to do it, and who the people are who will play a role in getting it done. One could amplify the system’s usefulness in this role as a how-to resource by allowing its users to add informal notes to the descriptions. Then the system also becomes a repository for the accumulated societal wisdom concerning the realitics of actually carrying out the tasks. The functionality we have discussed thus far has only required knowledge of the procedures in general. The system’s usefulness can be further enhanced by providing it with the capability of knowing about specific instances. With this capability the system could participate in the work in one of two ways: by tracking the progress of tasks, and by carrying out tasks itself. A task tracking facility would allow the system to: * be a source of information regarding the task’s status, history, and plan: * send requests to the agents who are to do the next steps, and make available to them a description of what they are to do what has been done, and pojnters to the relevant documents: * send out reminders and alerts when expected completion times of task steps have passed; and * ask for intervention by the appropriate agent when problems arise. A system which is tracking tasks in this way is participating as a partner in doing the work. Once that symbiotic relationship has been established between system and office worker, there are many steps in office procedures that the system could do itself. Such tasks would certainly include communication activities (e.g., using electronic mail), and maintenance of consistency in structured information bases (e.g., automatically filling in fields of electronic forms, see [Fikes], 1980). Office Work and Office Procedures With this class of intended systems in mind, we now turn to the question of the nature of the office work that we hope to support. In so doing, our goal is to determine the nature and structure of the information needed by our intended systems to support that work. A Procedure Execution Model A common model of office work considers an office worker to be a processor with a collection of tasks to be done and a procedure for doing each task. The work, in this model, involves executing each procedure and “time sharing” among them. However, studies of office work reveal a complexity and diversity of behavior far beyond what would be predicted by this model (For example, see [Suchman], [Wynn], and [Zimmerman]). In this section we explore, the nature of this apparent discrepency as a way of exposing1 characteristics of office work that we think have importam implications in the design of systems to support that work. {Note: The potential usefulness of this discrepancy was suggested to us by [Suchman].} tasks assigned to him. His work involves the planning, scheduling, and context switching associated with time sharing among those tasks. However, he can exercise options in carrying out his scheduling task that are not available to the scheduler in a computerized time sharing system. In particular. he can modify tasks themselves. For example, the worker may choose to * ignore some of the requirements of a task, * reni:gotiate the requirements of a task, * get someone else to do a task, * create and follow a new procedure for doing a task. Iicnce, office work includes, in addition to the carrying out of tasks, the determination of when a task should be done, how the task is to be done, and whether the task will be done at all. {Note: The office worker also has goals other than the completion of assigned tasks. For example, he has career goals (try to get ahead in the company), company goals (maximize profit), personal goals (keep from being bored), social goals (be regarded as good company), and societal goals (be honest).] Second, we take it as obvious that the domain with which office systems must deal is open-ended: truly anything may become relevant to the workings of offices at one time or another. This fact inplies that a procedure which implements a task is necessarily an inadequate description of all the actions which must be done to achieve the task’s goals in all the various situations that can (and inevitably will) occur. That is, at the time the procedure is defined which implements a task, one cannot predict ‘either the range of situations that will be encountered in an office or the extent of the knowledge, activities, and considerations that will be needed to carry out the task. Hence, for any given procedure, situations may occur in which the procedure does not indicate what is to be done, or in which what is indicated in the procedure cannot be done. For example, situations may occur in which: a case analysis in the procedure does not include the current situation, assumptions about the form and availability of inputs for a step are not met. resources required to do a step are not available, the action described in a step will not have the intended effects. The procedures associated with each task serve as a guide in that they indicate one way of doing the task under a particular set of assumptions. The office worker has the responsibility of deciding in each particular situation whether the procedure’s assumptions are satisfied and whether he wants to carry out the task in the way specified by the procedure. Third. the office worker has the problem of interpreting abstract specifications of what is to be done. For example, it is not uncommon in procedure specifications to find phrases like “include any other pertinent data”, “send forms approximately six weeks in advance of the deadline”, and “arrange for employee to receive benefit checks at home”. What is “any other pertinent data”, when is “approximately six weeks in advance of the deadline”, and how is one to “arrange for the employee to receive benefit checks at home”? The specification of the procedure doesn’t say. Hence, a necessary part of the work of following office procedures is determining what the abstract specification implies is to be done in each particular case. We conclude from these observations that the standard model of procedure cxccution is inadequate for describing office work. The original procedure specification serves only as a guide in this process and can be thought of as the first approximation to a plan for the particular task at hand. It is the responsibility of the office worker in each particular case to determine the suitability of the procedure, fill in missing details, and modify it where necessary to achieve the goals of the task. First, consider the office worker’s ongoing (meta-)task of determining how to allocate his resources among the collection of 203 An Example Of Office Work. To make these points more tangible. of the everyday work which goes _ _ let us now look at an example on in an office (This is an elaboration of an actual case of office work reported by [Wynn], p. 49). This example exhibits the problematic nature of the work, and the need for reflecting upon the specifications of the procedures. Xerox sells supplies for its copiers - paper, toner, and such. Customer orders for supplies are taken over the phone by a “customer order entry clerk” (COEC). The COEC talks to the customer and fills out a form which records the order. This order form is used by other clerks to bill the customer and to deliver the supplies. The form has a field for rccordir,g the address at which the copier is located, and there is an underlying assumption that this is the address to which the supplies are to be delivered. In the particular incident of interest, the customer informed the COEC ihat he could not supply an address for the copier because it was located on an ocean-going bargc(!j. This situation, of course, raised the question of what should be put into the address field of the order form. The clerk realized that the intended use of the address was to specify where the supplies were to be delivered, and that because the copier was on a barge that the needed address was dependent upon when the delivery of the supplies was to be made. Since he could not predict that date, he obtained from the customer a telephone number that could be called when the delivery was about to be made to obtain the address of the current location of the barge. He entered that telephone number into the field of the form and added a notation indicating how the number was to be used. The story continues: When the billing clerk was making up the bill, the question arose as to whether or not to charge California sales tax. The answer depends on whether or not the supplies were to be delivered out-of-state. The address field of the order form was examined, as per the usual procedure for answering the question, and of course no information about the state was available. What now? The billing clerk read the notation, called the telephone number, and ask the respondent whether the delivery was to be made in or out of California. Again, the date of the delivery was crucial in determining the answer. However, the billing clerk knew approximately when the supplies would be available, and therefore was able to determine from the person called that the delivery would be made in California, even though the precise delivery address was still not known. An addition was made to the information in the address field of the order form indicating that the delivery was to be made in Califo&a, and the bill was prepared and sent. Finally, the shipping clerk, with the supplies in hand, repeated the telephone call when preparing the shipping label. The address was then known, the address was added to the form, and the supplies were delivered. Analysis of This Example. What we have here is a case of a blown assumption. The iprocedures in which all three of these clerks were playing a role were designed on the assumption that copiers do not move and therefore have a fixed address. The particular case violated that ‘assumption. The COEC ‘was confronted with a problem because he could not carry out a step of a procedure (i.e., he could not fill in an address for the copier). There are several things hc could have done at that point, including ignoring the step or telling the customer that unless he provided an address that the order could not be taken. Instead, he chose to stop “executing” the procedure and to step back and reason about it. In particular, he considered what were the intended uses for the problematical address; i.e.. what was the goal of filling in the form’s address field. Using that information, he created a plan involving both himself and the shipping clerk that was within the spirit, although not within the letter, of the established procedures. That is, he devised an alternative that would satisfy the goals of the intended users of the address, as he perceived them. Hence, those goals were the crucial information that the COEC needed in order to determine suitable alternative actions when the unexpected situation occured. Note that the COEC was apparently not aware of the billing clerk’s use of the address field to determine state sales tax. Hence, the COEC’s alternative plan did not indicate how the billing clerk was to deal with this situation. The billing clerk, like the COEC, was confronted with a problem of not being able to carry out a step in a procedure (because the address field of the order form did not contain an address). Again, as was the case with the COEC, he did not ignore the problem or reject the situation as unacceptable. Instead, he attempted to find suitable alternative actions that would satisfy his task goals and allow the billing to proceed. His planning involved understanding the alternative procedure for the shipping clerk that had been formulated by the COEC, and realizing that he could use the telephone number included in that formulation to satisfy his goals. Consider the nature of the information involved in this example. Note the unpredictabiliry at the time the form was designed of the kinds of information that would be put on the fom:. Note also that the information on the form regarding the address was changing throughout the procedure. First there was a note describing a procedure for obtaining the address, then a parfial address containing only the state added to that note, and finally a complete description of the address. Another form of partial description that played a role in the example was approximafim; in particular, the clerks’ knowledge of the approximate delivery date. The strength and certainty of those approximations determined when and to what extent the delivery address was obtained. Supporting the Work Requires Flexibility We have presented the idea that the work that actually goes on in offIces is not routine. It consists of many particular cases of applying the given procedures to the details of those cases. This work involves dealing with unsatisfied assumptions, doing planning, understanding goals, and using information that is partial, approximate, and changing. The illusion that office workers execute procedures in a manner that is analogous to the way computers execute procedures ignores these realities of the situation. Given that picture of office work, we now turn our attention to the requirements placed on the design of computer-based systems to support such work. A primary design challenge is to find ways of providing the flexibility that is needed to allow for the application of established procedures to the circumstances of particular cases. With respect to information being supplied to the system by users, this flexibility involves dealing with cases where information is missing, information is provided in unexpected forms, and/or information in addition to what was expected is supplied. With respect to the procedural steps being carried out, this flexibility involves dealing with cases where steps are omitted, steps are done in different order, and/or additional steps are done. When office systems lack the flexibility to deal with these contingencies, they severely restrict the options of their users and thereby become yet another bureaucratic barrier to be overcome in “getting the work done”. Consider. for cxamplc. an electronic fcrms system for supporting the work of the COEC. When the “copier on a barge” problem arose, the COEC would hake ncedcd that system to be flexible CI;OL@ to allow entries other than addresses in the form’s address ficid. In particular, the COEC needed to be able to say to the sy%tcm, in effect., “I can’t give you an address in this c&c. Insicad, 171 give you a note for the shipping clerk.” If the syr.tem aI50 used its descriptions of the procedures being followed to provide instmctions to the clerks regarding what is to be done, then the system would need to bc able to accept the COECs decision to omit the step of 204 providing an address in the form’s address field, and to incorporate into the shipping clerk’s procedure an instruction to read the COEC’s note the first time the address was needed. In addition to being able to accept such alternative inputs, any of the system’s facilities for doing computations based on those inputs (e.g., to compute the state sales tax on customer orders using the address on the order form) must be designed to deal with cases in which those inputs have some unexpected structure or are not available at all. The challenge in the design of such processing facilities is to provide ways for the system, in cooperation with the office worker who is being supported, to overcome the difficulty posed by the failed computation so that work on the task can continue in a productive manner. One often hears the argument that this need for flexibility and variation from established procedures could be overcome by doing a more thorough analysis of the office tasks and thereby producing complete procedures that would cover all of the cases that could occur. Our claim is that because of the open-ended nature of the office domain, one cannot anticipate all of the situations that will occur in the carrying out of a given task, and therefore cannot totally characterize the inputs that will be available or the actions that might be taken to satisfy the task’s goals. The Relevance of AI to Supporting the Work An Alternative Model: Planning and Plan Execution The observations we have presented on the characteristics of office work have led us to seek an alternative to the procedure execution model to guide us in building a knowledge base for office support systems. We have found what we think to be a suitable alternative in the paradigms from the AI literature for automatic planning and execution monitoring of plans. That is, we take the viewpoint that we are confronted not so much with the problems of representing and supporting the execution of procedures, but with the problems of representing plans and supporting the monitoring and replanning that occurs during their execution. This viewpoint provides us with a conceptual framework for understanding the use of procedures in an office, an understanding that we feel is critical to dealing with the problems of designing systems to actually support that work. In the following paragraphs we present some of the key aspects of this point of view and discuss the ways in which it suggests that a system could provide useful support. The basic requirement on a data base describing these plans are that they provide the information needed to monitor a plan’s execution and to do whatever replanning might be required. What information is needed during those operations? By referring to the planning paradigm used in the STRIPS systems ([Fikes], et al, 1972), we obtain the suggestions that execution monitoring requires descriptions of the expected results of each operator, the intended use of each operator result, the preconditions of each operator, and the assumptions made by the planner about the world at each step of the plan. Planning involves the use of descriptions of the current state of the world, the operators available as potential plan steps, and the goals to be achieved by the plan. This planning paradigm characterizes some of the information that might be useful in the doing of office tasks and therefore suggests what to include in the description of office tasks and their associated procedures. In particular, it suggests the inclusion of information about task goals, intended uses of operator results, and precondition assumptions of operators. For example, the COEC employed information regarding the intended use of the address by the shipping clerk to determine an alternative plan when the address was not available. If the COEC had also known about the billing clerk’s intended use of the address, then he would have tried to obtain the information needed for that use (i.e., the state in which the delivery would be made) and, if successful, would have eliminated the difficulties that the billing clerk had in the example. One of the major ways we see for a system to provide support is by serving as an information source for office workers -buring the execution of plans and during any replanning that may be required. Hence, the planning p?radigm suggests what information to include in the system’s representation of the tasks and procedures, and provides us with a basis for characterizing the questions that the user may ask of the system. The paradigm of hierarchical planning (e.g., see [Sacerdoti]) also applies here and can be used in our characterization of office work. That paradigm would suggest that we consider each individual step in a plan as being a task with its own inputs, enabling conditions, goals, etc. There may or may not be a plan associated with any given step’s task. In the cases where there are, these plans form a tree and then get combined in various ways to form a planning network. Such a network represents a hierarchical plan, where the top of the hierarchy decribes a top level task and each sucessive level of the hierarchy describes increasingly detailed subtasks. In the standard non-hierarchical planning case, there is a plan for each step and each plan consists of a single operator; hence, there is a one-to-one correspondance between plan steps and operators. In the hierarchical planning paradigm and in the office, that one-to-one correspondance need not exist. Hierarchical planning networks appear to be an important device for representing office plans for several reasons. They are a useM structure for representing the task-subtask and goal-subgoal relationships that need to be known about during execution monitoring and replanning, and they provide the basic descriptive framework for indicating how the work is to be organized. Also, since there are effectively no primitive operators in the office, there is a need for describing office plans at varying levels of detail, depending on the specific needs of the describer and users of the descriptions. That flexibility in the level of detail of specification is therefore needed in an office system’s representation facilities. The system can then be involved in the office work at varying levels of detail. For example, the system may know that a travel request needs to be authorized, but know nothing about the subtasks involved in obtaining the authorization. Such flexibility is also an important tool for enabling the system to participate in situations that it does not understand. For example, if a plan that the system is helping to monitor fails, the user may not describe to the system the alternative plan he decides to use. However, the system knows about the goals of the original plan that the alternative must also satisfy and can therefore monitor the accomplishment of those goals, ev;;evz~gh it now has no model of how those goals are being An important way in which office work motivates extension of current AI planning paradigms is that office work is done in a multi-processor environment, That is, one can consider each agent (person or system) in an office to be a processor that is accomplishing tasks by carrying out plan5 that have either been given to it or created by it. Any given plan may involve the participation of several agents, each asent acting as an independent processor executing one or more of the plan’s steps. The processors make commitments to each other regarding the goals th.ey will achieve during the execution of a plan [Flores]. Therefore, in creating a plan to be carried out by multiple agents, the commitments that those agents will make to each other are a crucial part of a multi-processor plan. Furthermore, the commitments an agent has made and that have been made to him form for him a set of constraints within which he must work. In particular, these commitments form the context in which replanning takes place. in that any new plan must satisfy those commitments. However, the agent also has the options during replanning of renegotiating the commitments he has made or of ignoring them altogether. 205 Any system that is to participate in the replanning process needs to support this commitment-based view of planning and take them into consideration. In particular, a system could help an agent keep track of the commitments he is involved with that relate to a particular task and indicate his options for changing them during replanning. To support this tracking, the system’s plan representation needs to include for each plan step both the commitments made by and to the agent responsible for doing the step. The multi-processing nature of office work also implies that steps of a plan can be done in parallel. Hence, representations for office plans need to allow specification of partial orderings for plan steps. That requirement and the pervasiveness of replanning that occurs during plan execution suggest that task descriptions should include a set of necessary and sufficient “enabling conditions” for beginning the task so that it can be determined when the task can be begun irrespective of the order or nature of the steps that achieved those conditions. Up to this point we have not considered perhaps the most immediate question that arises out of looking at office work from a planning point of view: to what extent can we expect a system to automatically do the planning and replanning that is needed for office tasks? The primary limitation on such automatic planning seems to be the open-endedness of the office domain. That is, the extent to which there are considerations relevant to the formation of the plan of which the system has no understanding will limit the system’s effectiveness at determining an appropriate plan. For example, the system may not know about possible operators in the situation, the costs or chances for success of renegotiating or ignoring existing commitments, other goals that interact with the task’s goals, or the implications of an unexpected situation. These limitations have lead us to focus on a symbiotic relationship between system and office worker during planning and replanning, ,where the system plays primarily a role of supporting the planning /being done by the users by helping represent., manage, and communicate the resulting plans. In conclusion, then, we are claiming that * a multi-processor hierarchical planning model is B useful one for understanding office work and therefore for structuring an offlice system’s knowledge base, and * the demands of supporting offlce work motivate new research in multi-processor commitment-based planning. Knowledge Representation Challenges We turn now to the demands that supporting office work makes on the representation of the knowledge which systems have of the ‘office domain. We then discuss two techniques that arise from work in the AI community that provide particularly promising starting points for confronting those demands. The single most salient demand of such representation schemes is that they be able to respond to the need for change in conceptualization of the work. As we have seen, the domain of office work is inherently open-ended (e.g., before the barge case there was no notion of addresses being time-dependent: afterwards there was). Consequently there is no way to anticipate the full range of subject matter with which the system will have to deal. In consequence, the representation scheme must be able to handle any conceivable conceptualization which, over the course of the life of the office, the users of the office system choose to enter into the system. Furthermore, as time passes, this conceptualization will change to meet the changing understanding of the office domain which the users of the system have. Sometimes these changes will be small; at other times, there will be major “re-thinkings” of the information. The system must not only be able to represent this changing pattern of thought, but must also be able to simultaneously represent the pattern of changing thought: to support office work, it will have to be able to support the history of what has happened previously, and consequently will have to be able to hold simultaneously the old conceptualizations - for supporting the understanding of the past, and the new. The second demand placed by offlice systems on the mechanisms for representing the knowledge which is within them is to support partial knowledge of their domain. This incompleteness comes in at least three forms: the support of a subset of some expected body of knowledge (e.g., the state in which an address is may be known, but nothing more): the support of an abstraction of the knowledge (e.g., the supplies ordered are a paper product, but which one is unknown); and the support of an approximation (e.g,. the date of delivery is between mid-April and mid-May). In particular, this ability to support partial knowledge will permit the entry into the system of all that one knows, even though that may only be part of what is desired in a complete description. A resultant demand is that the mechanisms which access information must be prepared for the expected information not to be there (e.g.,. the state portion of an address is missing from the available information when the billing clerk tries to bill the barge- ownei). This preparation involves the representation scheme in at least being able to detect the absense, and further, in having some means of coping with the resulting problems. The third demand results from the fact that knowledge of things /often accumulates over time. Sometimes such an accumulation of partial descriptions can be reformed into a coherent whole. But more often the pieces are better retained as independent, un- coordinated facts. Indeed, rather than think about OX description of an entity, it is often useful to view the object as having mulfiple descriptions. Thus, for example, the knowledge about the address of the copier might at some point include three distinct descriptions of the address: as having California as its state portion, as being a changing thing, and as being something which can be fkthur determined by carrying out the procedure “call this number and ask”. The final demand arises from the expectation that the system should provide ‘a general model of office work. This model could be crafted by experts on the organizational structuring of oflices, and would then be available as a conceptual framework to support the description of more particular details. In fact, these concepts become the zerrns in which the details are not only described, but understood. Thus, for example, the concepts of task, goal, procedure, plan, agent, post, authorization, commitment, and data repository might be provided as a very general framework for modeling offices. Particular offices would have their own particular tasks, goals, etc. These demands pose challenging research problems in knowledge representation which we are not claiming to have solved. However, we discuss in the following paragraphs two starting points for confronting these problems that look particularly promising to us and that we are using in our work. Our first starting point for responding to the demands of supporting office work is the use of a specialization-based knowledge representation formalism (see, for example, [Brachman]). This, and similar, schemes for formally and precisely representing knowledge take as their goals the first three of our needs: representation of a changing support for conceptual structure, description, and multiple description. partial The major structuring principle for their representations is defnilion 6~ specialization: any concept in a knowledge base can be taken as a basis for defiriing a new concept which is a special case of the old one. Thus, describing the details of a particular office can be done in this formalism by specializing more general descriptions of offices. This specialization can be done in steps, thus permitting the tailoring of the conceptualization in various ways to produce progressively less abstract descriptions. In the end, the most specific descriptions are understood through their place in a taxonomic lattice (a concept can specialize more than one abstraction) of more abstract concepts. Not only does this enrich the understanding of the domain, for one can understand similarities between concepts in terms of common abstractions, but it also provides locations for attaching knowledge about abstractions which will immediately apply to all special cases of those abstractions. Our second response to the knowledge representation needs presented by the offlice domain is the use of semi-formal descriptions. That is, WC are using the formal knowledge representation mechanisms in a style which permits us to capture information that the system dots not “understand” in such a way that it can be uscfmly employed b:i human users of the system. For example, paragraphs of English prose or diagrams can be associated with concepts: the only use the system will be able to make of them will be to present them to a human user, and permit him to read and modify them. We view these mixtures of formal (the structure is understood by the system) and informal (the structure is understood only by humans) descriptions as an essential “escape valve” for our knowledge representation systems: if it were required that the system had to understand the conceptualization underlying all the informaion in its knowledge structure, then the cost of entering information which is currently beyond the system’s understanding would be very high. Instead, by escaping into informal description, the system can still be used as a repository for all of the information about the situation at hand, and yet permit the work to proceed. The system becomes primarily a communications . device for supporting inteiaction between people carrying out office work. However, because the informal descriptions are represented in the same description formalism as the formal descriptions, they can be integrated into the knowledge base in a consistent manner. An Approach to System Building: Symbiotic Processing In the account given above of our vision of the behavior and properties of a system to support the real work of offices, there has been repeated reference to interaction between the system and its human users. This will be’ an important aspect of successfully completing the planning which must be done in office work. Also in representing knowledge within the system, we have argued that a “semi-formal” mixture of information will be important for achieving practical systems. That is, it takes both man and machine to understand the information held within the system. Changing the emphasis, we prefer to think of the humans and the computer as cooperating processing engines carrying out the “computations” of the system in partnership. The idea here is that both processing engines, each with its own processing capabilities, knowledge, and memory structures. are essential to getting the task done. Neither could effectively do the work without the other. In biology, such interdependence is called symbiosis, and a system composed of two or more interdependent organisms is called a symbiolic system. We therefore use the same term to refer to the sort of offrce systems envisioned here: there is a symbiosis of human and machine. Why is it that this quite obvious co-operation between man and machine has not so far been the dominating pattern of computer use in office (as well as other) systems? Our theory is this: when batch processing was the only economically feasible form of business (and therefore office) system, such interaction was impossible - the human partners were simply not around when they were needed to help the computer in its tasks. However, the pervasiveness of the belief among designers of computer systems that ale procedures in offices were “routine” obscured the need for truly cooperative interaction. A result of this belief, and of the introduction and widespread use of batch processing systems in the business environment, has been the establishment and buttressing of the pow-accepted notion that there is something fundamental about partitioning the world into routine cases and exceptions. We view this distinction between routine and exception as quite artificiai: routine cases are those cases where no action is required of the human partner in the symbiosis: exceptions are everything else. And, in fact, it is even worse than that: the distinction breeds viewing the world this way, which in turn enhances the distinction. To get the proportion of routine cases up to the point where batch systems could be justified, many cases which require some small amount of human processing are handled by forcing the human processing to be done before the cases are “entered into” the system. This is often done at the expense of some capacity of the system (as a whole) to handle not-quite-standard cases: these cases are wedged into the mold of the “routine”. And when exceptions are not even permitted - when everything has to go through the system as a routine case, the mold can well become a straight-jacket. We propose that office systems need not make this distinction between the routine and the exceptional. Instead, it should be possible, and is desirable, to return to the “good old days” when all cases were processed in the same way, some with more effort than others. We believe that, armed with the understanding of ofliccs presented here, and supported by studies in the AI fields of automatic planning and knowledge representation, a modem version of the “good old world” can be achieved through systems built around the notion that all cases are handled by the poweri% symbiosis of humans and machines. Acknowledgements We would like to thank Lucy Suchman and Eleanor Wynn for early contributions to the ideas in this paper. In particular, they made available to us transcripts of their interviews with office workers and observations of office work, and helped open our eyes to what those interviews and observations had to say about offlice work. We would also like to thank Lucy for her insightful participation in the continuing research that has led to the observations in this paper. References Brachmap, R. et al. KLONE Reference Manual, BBN Report No. 3848, July 1978. Fikes, R. E., Hart, P. E., and Nilsson, N. J. “Learning and Executing Generalized Robot Plans”. Artificial Intelligence, 3(4), winter 1972, pp 251-288. Fikes, R. E. “Odyssey: A Knowledge-Based To appear in ArtiJicial Intelligence. Personal Assistant”. Flores, F. Univ. of California at Berkeley, personal communication. Sacerdoti, E. D. A Structure for Plans and American Elsevier, 1977. Behavior, New York: Suchman, L. A. “Office Procedures as Practical Action: A Case Study”. SSL Internal Report, Xerox Palo Alto Research Center, Sept. 1979. WY% Eleanor Herasimchuk “Office Conversation as an Information Medium”. Department of Anthropology, University of California at Berkeley, May 1979. Zimmerman, D. H. “The Practicalities of Rule Use”. in J. D. Douglas (Ed.), Underslanding Everyday Life, Aldine Publishing Company, Chicago, pp 221-237. 207	1980	53
49	Metaphors and Models Michael R. Gcnesereih Computer Science Department Stanford University Stanford, California 94305 1. Introduction Much of one’s knowledge of a task domain is in the form of simple facts and procedures. While these facts and procedures may vary from domain to domain, there is often substantial &nilarity in the “abstract structure” of the knowledge. For example, the notion of a hierarchy is ‘found in biological taxonomy, the geological classification of time, and the organization chart of a corporation. One advantage of recognizing such abstractions is that they can be used in selecting metaphors and models that are computationally very powerful and efficient. This power and efficiency can be used in evaluating plausible hypotheses about new domains and can thereby motivate the induction of abstractions even in the face of partial or inconsistent data. Furthermore, there is a seductive argument for how such information processing criteria can be used in characterizing “intuitive” thought and in explaining the cogency of causal arguments. The idea of large-scale, unified knowledge structures like abstractions is not a new one. The gestalt psychologists (e.g. [Kohler]) had the intuition decades ago, and recently Kuhn [Kuhn], Minsky [Minsky], and Schank [Schank & Abelson] have embodied similar intuitions in their notions of paradigms, frames, and scripts. (See also [Bobrow & Norman] and [Moore & Newell] for related ideas.) The novelty hcrc lies in the use of such structures to select g6od metaphoa and models and in the effects of the resulting power and efficiency on cognitive behavior. This paper describes a particular formalization of abstractions -in a knowlcdgc rcprcsentation system called hNAr.OG and shows how abstractions can be used in model building, understanding and generating analogies, and theory formation. The prcscntation here is necessarily brief and mentions only the highlights. The next section defines the notions of abstraction and simulation structure. Section 3 describes the use of abstractions in building computational models, and section 4 shows how abstractions can bc used to gain power as well as efficiency. 2. Abstrrrclions and Sirrwlalion Struclures Formally, an abslrucfion is a set of symbols for relations, functions, constants, and actions together with a set of axioms relating these symbols to each other. Abstractions include not only small, simple concepts like hierarchies but also more complex notions like concavity and convexity or particles and waves. A model for an abstraction is essentially an interpretation fcr the symbols that satisfies the associated axioms. Different task domains can bc mod& of the same abstraction (as biological taxonomy, geological time, and organization charts are instances of hierarchies); or, said the other way around, each abstraction car have a number of different models. Importantly, there are multipk computational models for most abstractions. In order to distinguish computer models from the task domains they are designed tc mimic, they are hereafter termed simulufion slrucfures, following Weyhrauch [Weyhrauch]. There is a strong relationship between abstractions and metaphors, or analogies. Many analogies are best understood as statements that the situations being compared share a common abstraction. For example, when one asserts that the organization chart of a corporation is like a tree or like the taxonomy of animals in biology, what he is saying is that they are all hierarchies. With this view, the problem of understanding an analogy becomes one of recognizing the shared abstraction. Of course there are an infinite number of abstractions. What gives the idea force is that the simulation structures for certain abstractions have representations that arc particularly economical, algorithms that are particularly efficient, or theorems that are particularly powerful, e.g. hierarchies, grids, partial orders, rings, groups, monoids. Consequently, there is advantage to bL gained from recognizing the applicability of one of these special abstractions rather than synthesizing a new one. Even when the applicability of such special abstractions and simulation structures cannot be determined with certainty (say, in the face of incomplete or faulty information), there is advantage in hypothesizing them. Until one is forced to switch abstractions due to incontrovertible data, one has an economical representation and powerful problem solving methods. By biasing the early choice of abstractions in this way, these criteria can have qualitative effcctc on theory formation. 3. Models The importance of abstractions and their associated measures of economy, efficiency, and power is clearest in the context of a concrete implementation like the ANALOG knowledge reprcscntation system. l’hc interesting feature of ANALOG is that it utilizes a variety of simulation structures for representing different portions of the knowledge of a task domain. This setup is graphically illustrated in figure 1. The user asserts facts in the system’s uniform, domain-independent formalism, and the system stores them by modifying the appropriate simulation structure. Facts for which no simulation structure is appropriate are simply filed away in the uniform representation. (ANALOG currently uses a semantic network representation called Dl3 [Gcnesereth 761. The formalism allows one to cncodc scntcnccs in the predicate calculus of any order and provides a rich meta-level vocabulary.) Descriptions of each of ANALOG'S abstractions and simulation structures arc also cncodcd within the DU rcprescntation. I ASS1 Figure 1 - An Overview of ANALQC This approach departs from the custom in knowledge representation systems of using uniform, domain independent formalisms. While there are advantages to uniformity, in many cases the representations are less economical than specialized data structures, and the associated general procedures (like resolution) are less efficient or less powerful than specialized algorithms. For 208 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. example, a set in a small universe can be efficiently represented as a bit vector in which the setting of each bit determines whether the corresponding object is in the set. Union and intersection computations in this representation can be done in a single machine cycle by hardware or microcoded boolean operations. By contrast, a frame-like representation of sets would consume more space, and the union and intersection algorithms would have running times linear or quadratic in the sizes of the sets. The distinction here is essentially that between “Fregcan” and “analogical” representations, as described by Balzer [Balzer]. Note that ANALOG'S approach is perfectly compatible with uniform knowledge representation systems like DB and RLL [Greiner & Lenat]. The addition of abstractions and simulation structures can be viewed as an incremental improvement to such systems, and their absence or inapplicability can be handled gracefully by using the uniform representation. It’s important to realize that ANALOG is not necessarily charged with inventing these clever representations and algorithms, only recognizing their applicability and applying them. The approach is very much in the spirit of the work done by Green, Barstow, Kant, and Low in that there is a knowledge base describing some of the best data representations and algorithms known to computer science. This knowledge base is used in selecting good data representations and efficient algorithms for computing in a new task domain. One difference with their approach is that in ANALOG there is a catchall representation for encoding assertions when no simulation structure is applicable. Other differences include an emerging theory of representation necessary in designing new simulation structures (see section 3.2) and the use of the criteria of economy, efficiency, and power in theory formation. ANALOG'S use of simulation structures is in a very real sense an instance of model building. Architects and ship designers use physical models to get answers that would bc too difficult or too expensive to obtain using purely formal methods. ANALOG uses simulation structures in much the same’ way. In fact, there is no essential reason why the simulation structures it uses couldn’t be physical models. Furthermore, as VLSI dissolves what John Backus calls the vonNeumann bottleneck, the number of abstractions with especially efficient simulation structures should grow dramatically. 3.1 Building a Model As an example of modeling, consider the problem of encoding the ofganization chart of a corporation: ‘[‘he first step in building a model for a new task domain is finding an appropriate abstraction and simulation structure. The knowledge engineer may directly name the abstraction or identify it with an analogy, or the system may be able to infer it from an examination of the data. In this case, the hierarchy abstraction is appropriate, and there are several appropriate simulation structures. One of these is shown in figure 2. Each object in the universe is represented as a “cons” cell in which the “car” points to the object’s parent. The relation (here called Rel) is just the transitive closure of the Car relation, and Nil is the root. For the purposes of this example, the “cdr” of each cell may be ignored. ,++$$yj&$$ Figure 2 - A Simulation Structure for the Hierarchy Abstraction An important requirement for a simulation structure is that it be modifiable. Therefore, it must include actions that the model builder can use in encoding knowledge of the task domain. Usually, this requires the ability to create new objects and to achieve relations among them. In this case, the Neons subroutine creates a new object, and Rplaca changes an object’s parent. Part of the task of finding an appropriate abstraction and simulation structure is setting it up for use in encoding knowledge of the task domain. This includes three kinds of information. The first is an index so that the system can determine the simulation structure appropria‘tt: to a new assertion. (This index is necessary since several domains and simulation structures may be in use simultaneously). Secondly, there must be a procedure for mapping each assertion into its corresponding assertion about the simulation structure. And, finally, the system must have information about how to achieve the new assertion. Once the simulation structure is chosen and set up, the system builder can begin to assert facts about the task domain, and the system will automatically modify the simulation structure. As an example of this procedure, consider how the system would handle the assertion of the fact (Boss-of Carleton Bertram). First, it would use its index to determine that the simulation structure of figure 2 is being used and to recover the mapping information. ‘I’hcn it would map the assertion into the simulation domain. In this cast, let’s say that Arthur is the boss of Bcrtrm and lhdrice while Carleton has been installed in the model as Ileatrice’s employee. Then the new assertion would bc (He1 ((Nil)) (Nil)), where the first argument is the object rcprcsenting Carleton and the second represents Bcrtram. By examining the mcta-level information about the simulation structure, the system retrieves a canned proccdurc (Rplaca) for achieving this fact and cxecutcs it, with the result that Carleton’s “car” is redirected from Bcittr!.ce to Bcrtram. An intcrcsting aspect of model building is that complete information is often required. For example, in adding a node to the simulation structure of figure 2, the system must know the object’s parent in order to succeed. (It has to put something in the “car” of the cell.) This problem can sometimes be handled by the addition of new objects and relations that capture the ambiguity. For example, one could add the special token Unknown as a place filler in the simulation structure above. (Of course, the resulting structure would no longer be a hierarchy.) Another example is using the concept of uncle as a union of father’s brother and mother’s brother. Unfortunately, this approach increases the size of the model and makes deductions more difficult. Unless there are strong properties associated with such disjunctive concepts, it is usually better to carry the ambiguity at the meta-level (i.e. outside the model, in the neutral language of the knowledge representation system) until the uncertainty is resolved. Another interesting aspect of the use of simulation structures is the automatic enforcement of the axioms of the abstraction. For example, in the simulation structure of figure 2, it is impossible to assert two parents for any node simply because a “cons” cell has one and only one “car”. Where this is not the case (as when a simulation structure is drawn from a more general abstraction), thy axioms can still be used to check the consistency and completeness of the assertions a system builder makes in describing his task domain. For example, if the system knew that a group of assertions was intended to describe a hierarchy, it could detect inconsistent data such as cycles and incomplete data such as nodes without parents. 3.2 Designing a Simulation Sfructure for an Abslraclion The only essential criteria for simulation structures are representational adequacy and structure appropriate to their abstractions. For every assertion about the task domain in the language of the abstraction, there must be an assertion about the simulation structure; and the structure must satisfy the axioms of the abstraction. In creating a simulation structure, one good heuristic is to try to set up a homomorphism. Sometimes, the objects of the simulation structure can be used directly, as in the case of using “cons” cells to represent nodes in a hierarchy. In the example above, the mapping of objects from the corporation domain into the domain of list structure was one-to-one, i.e. the corporate objects were all rcprcsentcd by distinct pieces of list structure, and 209 the relations and actions all mapped nicely into one another. Of course, this need not always be the case. Consider, for example, the state vector representation of the Blocks World proposed by McCarthy, in which the Supports relation bctwccn each pair of blocks is represented by a distinct bit in a bit vector. (Think of the vector as a matrix in which the Ci, jxh bit is on if and only if block i is on block J). In this representation the fact (Supports A B) would translate into something like (On Bit-AR Vector-l), and theie would be no distinct representations of the blocks A and B. In other cases, more complex objects may be ncccssary in order to provide enough relations. When a domain does not provide an adequate set of relations, it’s a good idea to synthesize complex structures from simpler ones. For example, a simulation structure’ for an abstraction with three’ binary relations could be built in the list world by representing objects as pairs of “cons” cells in which the “car” represents the first relation, the “cd? points to the second cell, the “cad? represents the second relation, and the,“cddr” represents the third. This approach is facilitated by programming languages with extensible data structures. Obviously, it pays to economize by using the prcdefined relations of the simulation domain where possible. For example, a good representation for a univariate polynomiat is a list of its coefficients, and one gets the degree of the polynomial for free (the length of the list minus 1). One advantage is representational economy; another is automatic enforcement of the abstraction’s axioms, as described in the last section. In order to use a simulation structure, it may be necessary to transform objects into a canonical form. For example, one can represent a univariate polynomial as a list of coefficients, but the polynomial must be in expanded form. There is a large body of literature on canonical forms for specific algebraic structures, while [Genesereth 791 gives a general but weak technique for inventing such forms directly from an abstraction’s axioms. In using a simulation structure, there is a tradeoff betweeh the amount of work done by the model and the amount done by the knowledge representation system. For example, in the simulation structure of figure 2, one must loop over the parent relation to determine whether two objects are related. This can be done either by the knowledge representation system or by a L'ISP procedure in the simulation structure. Obviously, it’s a good idea to have the simulation structure do as much work as possiblq, 3.3 Interfacing Simulation Stniclures For most interesting task domains, the chances are that a single simulation structure is not sufficient. In such cases, it is sometimes possible to piece together several different simulation structures. The simplest situation arises when the objects of the task domain form a hierarchy under the “part” relation. Then one can choose one representation for the “topmost” objects and a different representation for the parts. The spirit of this approach is very similar to that of “object-oriented” programming in which each object retains information about how it is to be processed.’ One disadvantage of this approach is that each object must have explicit “type” information stored with it. Barton, Genesereth, Moses, and Zippel have recently developed a scheme that eliminates this need by separating the processing information from each object and passing it around in a separate “tree” of operations. ANALOG uses this schcmc for encoding the operations associated with each simulation struct’ure. Task domains with several relations are sometimes decomposable into several abstractions, and these relations can then be represented independently. More often the relations are interdependent; and, when this is the case, the interdependence must be dealt with in the uniform representation, Even when a single abstraction would fit the task domain, it may be advisable to use several. Consider, fw .example, a partial order that nicely decomposes into two trees. Furthermore, there are often advantages to multiple representations of objects, as argued by Moses [Moses]. 4. Thinking With Abstractions and Simulation S@uclures The USC of specialized simulation structures gives ANALOG an economy and efficiency not possible with a uniform representation. The economy can be expressed in terms of the space saved by representing assertions in the simulation structure rather than the uniform representation. This economy derives from the elimination of the overhead inherent in uniform formalisms and the use of relations implicit in the simulation structure (as the length of a list reflects the dcgrce of the polynomial it represents). The efficiency refers to the time involved in doing dcd:rtions and solving problems. This efficiency may be attributable to clever algorithms, or it may be the result of long familiarity with the domain from which the abstraction evolved (due to the memory of many special case heuristics). Lenat [Lenat et. al.] discusses how a computer might improve its own performance by self-monitoring. An interesting pos,bility suggested by this economy and efficiency is for the program to use these criteria in evaluating plausible hypotheses about a new domain. In the face of incomplete or contradictory data, the program should favor the more economical abstraction. Clearly, there is some evidence for this sort of behavior in human cognition. Consider, for example, Mendeleev’s invention of the periodic table of the elements. He was convinced of the correctness of the format in spite of contradictory data, for reasons that can only be identified as simplicity. These criteria of economy and efficiency are also of use in characterizing why it is easier to solve problems from one point of view than another, e.g. proving a theorem using automata theory rather than formal grammars. Part of what makes causal arguments (see [deKleer] for example) so compelling is that they are easy to compute with. The reason for this is that a causal argument is an instance of a cognitively efficient abstraction, namely a diicctcd graph. One is tempted, therefore, to generalize dcKleer’s notion of causal envisionment as finding economical and efficient abstractions (perhaps identified with analogies) in which the desired conclusions are reached via simple computations. The idea can be carried a bit fUrther and generalized to include the criterion of problem solving power. In particular, one should favor an abstraction for its ability to solve a pending problem despite insufficient data. The obvious difficulty is that the assumption may bc wrong or there may be scvcral abstractions that are equally probable and useful. Consider, for example, the following arguments for determining the distance between the observer and the middle vertex of a Necker cube. “Well, the lines form a cube, and so the middle vertex must be closer to me than the top edge.” “No, not at all, the figure is concave, and so the middle vertex must be filrther away.” Both arguments are consistent with the data and refer to a single abstraction, 2nd in each case the conclusion is deductively related to that view. A second example is evident in the particulate-wave controversy. The particulate view is a simple abstraction that accounts for much of the data and allows one to solve outstanding problems. Of cours6, the same can be said for the wave view. Unfortunately, the predictions don’t agree. A similar argument explains the inferential leap a child makes in declaring that the wind is caused by the trees waving their leaves. When the child waves his hand, it makes a breeze; the trees wave when the wind blows; so they must have volition and motive power; and that would account for the wind. The reasoning in these examples is usually termed “analogical”. The key is the recognition of a known abstraction common to the situations being compared. This conception of analogy differs markedly from that of Hayes-Roth and Winston. In their view two situations are analogous if there is any match between the two that satisfies the facts of both worlds. If the match is good, the facts or heuristics of one world may be transferred to the other. The problem is that these facts may have nothing to do with the analogy. Just because two balls are big and plastic, one can’t infer because one ball is red that the other is also red. Abstractions are ways of capturing the necessary interdependence of facts. For example, the size and material of a ball do affect its mechanical behavior, and so the skills usefU1 for 210 bouncing one should be of value in bouncing the other. Also notL that the match need not be close in order for there to be a useful analogy. Linnaean taxonomy and organization charts have few superficial details in common, but the analogy is nonetheless compelling, and as a result the algorithms for reasoning about one can bc transferred to the other. The work of Hayes-Roth and Winston is, however, applicable where no abstractions exist yet. Their matching algorithms and the techniques of Buchanan, Mitchell, Dietterich and Michalski, and Lenat should be important in inducing new abstractions. An important consumer for these ideas is the field of computer-aided instruction. There is a current surge of interest in producing a “generative theory of cognitive bugs” (see [Brown], [Genesereth 80a], and [Matz]). The use of abstractions and the criteria of economy, efficiency, and power in theory formation is very seductive in this regard. Unfortunately, there is no reason to believe that the hardware of a vonNeumann computer in any way resembles the specialized capabilities of the human brain. (Indeed, psychologists are still debating whether there are any analogical processes in the brain at all. See, for example, [Kosslyn & Pomerantz], [Kosslyn & Schwartz], pylyshyn], and [Shepard & Metzler].) Thus, the idea at present is not so much a model for human cognitive behavior as a metaphor. 5. Conclusion The ANALOG system was dcvclopcd over a period of time to test the ideas presented here. One program accepts an analogy and infers the appropriate abstraction; another builds a model of the task domain as assertions are entered; and a third uses the model to answer questions. There is a sketchy implementation of the simulation structure designer, but no effort has been made to build the theory formation program. In summary, the key ideas are (1) the role of abstractions in understanding metaphors and selecting good models for task domains, (2) the use of’ models to acquire economy, efficiency, and problem solving power, and (3) the-importance of these criteria in theory formation. Abstractions and simulation structures make for a knowledge representation discipline that facilitates the construction of powerful, efficient AI programs. The approach suggests a program for much future work in AI and Computer Science, viz. the identification of us&11 abstractions and the implementation of corresponding simulation structures that take advantage of the spccia! computational characteristics of the vonNeumann machine and its successors. Acknowledgemenfs The content of this paper was substantially influenced by the author’s discussions with Bruce Buchanan, Rick Hayes-Roth, Doug Lenat, Earl Sacerdoti, and Mark Stefik, though they may no longer recognize the ideas. Jim Bennett, Paul Cohen, Russ Greiner, and Dave Smith read early drafts and made significant suggestions to improve ‘he presentation. grants from ARPA, NLM, and ONR. The work was supported in part by References Balzer. R. Automatic Programming, Institute Technical Memo, Southern California/ Information Sciences Institute, 1973. University of Barstow, D. R. Knowledge Bused Program Construction, Elsevier North-Holland, 1979. Brown, J. S. & vanlehn, K. forthcoming paper on learning. Buchanan, B. & Feigenbaum. E. A. Dendral and Meta-Dendral: Their Applications Dimension, Arfificial InteUigence, Vol. 11, 1978, pp 5-24. . deKleer, J. The Origin and of the Sixth International 197-203. Resolution of Ambiguities in Causal Arguments, Proc. Joint Conference on Artificial Intelligence, 1979. pp Hayes-Roth, F. & McDermott An Interference Matching Technique for Inducing Abstractions, Comm of rhe ACM, Vol. 21 No. 5. May 1978. pp 401-411. Genesereth. M. R. A Fast Inf&ence Algorithm for Semantic Networks, Memo 5, Mass. Inst. of Tech. Mathlab Group, 1976. Genesereth, M. R. The Canonicality of Rule Systems, Proc. of the 1979 Symposium on Symbolic and Algebraic Manipulation, Springer Verlag, 1979. Genesereth, M. R. The Role of Plans in Intelligent Teaching Systems, in Inrclligent Teaching Systems, D. Sleeman, ed. 1980. Genesereth, M. R. & Lenat, D. B. Self-Description and Self-Modification in a Knowledge Representation System, IIPP-880-10. Stanford University Computer Science Dept., 1980. Green, C. C. The Design of the PSI Program Synthesis System, Proc. of the Second International Conference on Software Engineering, Oct. 1976, pp 4-18. Greiner, R. D. & Lenat, D. B. A Representation Language Language, submitted for inclusion in the Proc. of the First Conference of the American Association for ArtiIicial Intelligence, Aug. 1979. Kant, E. Efficiency Considerations in Program Synthesis: A Knowledge Based Approach, doctoral dissertation, Stanford Univ. Computer Science Dept., 1979. K_ohler. W. Ges!nJc P.weholnPv: An Introduction to New Concepts in Modem Psychology, Liveright, 1947. Kosslyn, S. M. & Pomcrantz. J. R. Imagery, Propositions, and the Form of Internal Representations, Cognifive Psychofogy, Vol. 9 No. 1, 1977, pp 52-76. Kosslyn, S. M. & Schwartz, S. P. A Simulation of Visual Imagery, Cosnitive Science, Vol. 1, 1977, pp 265-295. Kuhn, T. The Structure of Scientific Revolutions, Univ. of Chicago Press, 1962. Lenat, D. B. Automated Theory Formation in Mathematics, Proc. of the Fifth International Joint Conference on Artificial Intelligence, 1977, pp 833-842. Lenat, D. B., Hayes-Roth, F., Klahr. P. Cognitive Economy in Artificial Intelligence Systems, Proc. of the Sixth International Joint Conference on Artificial Intelligence, 1979. pp 531-536. Low, J. R. Automatic Coding: Choice of Data Structures, ISR 15. Birkheuser Verlag, 1976. Matz, M. A Generative Theory of High School Algebra Errors, in Intefligenf Teaching Systems, D. Sleeman. ed. 1980. McCarthy, J. Finite State Search Problems, unpublished paper. Minsky, M. A Framework for Representing Knowledge, in The Psychology o/ Computer Vision, P. H. Winston, ed., McGraw-Hill, 1975. Mitchell, T. Version Spaces: A Candidate Elimination Approach to Rule Learning, Proc. of the Fifth International Joint Conference on Artificial Intelligence. 1977. Moore, J. & Newell, A. How can MERLIN Understand, in L. W. Gregg, ed. Knowledge and Cognition, Lawrence Erlbaum, 1974. Moses. J. Algebraic Simplification: A Guide for the Perplexed, Comm of rhe ACM, Vol. 14 No. 8, 1971, pp 527-537. Pylyshyn. Z. W. What the Mind’s Eye Tells the Mind’s Brain: A Critique of Mental Imagery, Psychological Bullerin, Vol. 80, 1973, pp l-24. Schank. R. & Abelson, R. Scripts, Plans, and Knowledge, Proc. of the Fourth Internatylonal Joint Conference on Artificial Intelligence, 1975, pp 151-157. Shepard, R. N. & Metzler, J. Mental Rotation of Three-dimensional Objects, Science. Vol. 171, 1977, pp 701-703. Thorndyke, P. W. & Hayes-Roth, B. The Use of Schemata in the Acquisition and Transfer of Knowledge, Cognifive Psyckology Vol. 11, 1979, pp 82-106. Weyhrauch, R. Prolegomena to a Theory of Formal Reasoning, STAN-CS-78-687, Stanford Univ. Computer Science Dept., Dec. 1978. Winston, P. H., Understanding Analogies, Mass. Inst. of Tech. Artificial Intelligence Laboratory, Apr. 1979. Dietterich, T. G. & Michalski, R. S. Learning and Generalization of Characteristic Descriptions: Evaluation Criteria and Comparative Review of Selected Methods, Proc. of the Sixth International Joint Conference on Artificial Intelligence. 1979, pp 223-231. 211	1980	54
50	EVERYTHING YOU ALWAYS WANTED TO KNOW ABOUT AUTHORITY STRUCTURES BUT WERE UNABLE TO REPRESENT James R. Meehan uept. of Information and Computer Science university of California lrvine CA 92717 If we're ever to get programs to reason intelligently about activities that are governed by laws, we’ll need a clear understanding of such concepts as obliga- tions, politics, disputes, and enforce- ment. Our work on this topic started when we tried to build a program that could understand some legal correspondence from l'i'th-century England, and its current application is in a program that simulates social interaction among small groups such as families and schools. While the long-range applications are in natural language processing, this work is primarily concerned with fundamental issues of representation. we first attempted to adapt the work by Schank and Carbonell [I,21 but later chose alternative definitions and representations, because we were looking at a wider variety of legal systems, where legal authority, for example, did not always imply the power of enforcement, and while they focused on the resolution of disputes, we were more interested in how aecisions are influenced. Authority structures can be found everywhere, in law courts, AI conferences, bridge games, friendships, and even restaurants. The language is used meta- phorically even in such domains as everyday physics. We define an authority structure in terms of a group of people who participate in some set of activities. There are many structures one could impose on a group for purposes of analysis, such as the sequencing of events, or the plans and goals that motivate the participants, and different contexts will provide different answers to the same question. For example, if we ask, "Why did John order a hamburger?" we might answer with '-because he was hungry" or "So that the waitress would tell the cook" or "In order to initiate a contract with the restaurant" depending on the context. An authority structure, then, is associated with a aroun to pick a neutral term. A group has a se; of participants, connected by a social set which specifies the attitudes they have about their acquaintances in the group. Every group has a set of normal Drocedures or activ- ities, in which the participants take certain roles. For our present purposes, it doesn't matter whether those activities are highly predictable, goal-driven, or structured in any particular way. Some of those acts change the social net ("social" acts); others involve the exchange of goods and services ("economic" acts); and others are acts of authority (to be defined shortly). An individual belongs to many groups at tne same time. In fact, a pair of individuals may Delong to two groups and relate to each other in different ways (e.g., role conflict). Any group-associated act may have a legal status, an indication of its conformance with the laws. We define 6 types of legality: 1 . An act (in the past, present, or future) is explicitly legal, requiring no permission. [Example: free speech.] 2. An act is legal only . permission has been given. [YiG need a license to practice medicine.] 3. An act is legal only if it is commanded. [A six-year old child taking medicine.] 4. An act is legal only if YOU are acting as someone's agent (for whom the act may or may not be legal). LA judge authorizes a police officer to search a house, even though ne is not allowed to searcn it himse1f.j 212 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. 5. An act is legally required. LIn Autnority structures are often bridge, YOU must follow suit if emoeaaea in one anotner, you can.1 and if a group has no rules governing certain actions, then it may inherit the rules from an 6. An act is explicitly forbidden. embeaaing structure. r'or example, a [You may not submit more than one Gontract paper to this conference.] is an embedded system specifying little more than mutual obligations. The embedding system takes care of the rest. embedding is not universal, however. You Of the many possible states, some are can't sue your friends if they fail to of concern to the law, such as ownership, show up for an invited dinner. legal responsibility, and right of claim (debt or injury). eisputes are questions In general, your power within a group either about the legal status of some act is measured by your ability to "make LAm 1 required to file a tax return by April lbtn if l'm getting a refund?] or things happen," to cause social, economic, legal, or other kinds of actions. You a bout the trutn value of some law-relatea incur a debt when someone acts to increase state LThe defendant pleaded not guiltyJ. your power, and it is expected that you will reply, though not t'primitivesll necessarily in 'The acts of authority kind. A bribe, for example, is an are: exchange of economic power (money) for legal power (position of authority). 1. to define or decide the legal status of an act, according to many activities call for decisions to the six types listed above, such be made. Some of these decisions are as commanding or obliging someone based solely on evidence and are to do something [Clean up your simply evaluations, but in other cases, real room] choices must be made. A crucial part of the description of any participant in a 2. to enforce a decision [Papers group is the ability he has to influenca received after the program aecisiqns by whatever means, and we treat committee meets will be rejected tnis as a special kind of power, automatically] categor- lzed bY tne kind of decision being influenced. Politics is defined as the 3. to create and revise the rules influencing of authority-related themselves [The voters approved decisions, i.e., whether (or in what the ERA1 manner> to perform one of the acts of authority, the laws or 4. to resolve disputes [The such as revising jury admitting a new member to the acquitted the defendant] group. Attempts to influence aconomiq decisions range from friendly advice ["Try the 5. to change the position of someone cheeseburgerl'] to hard-sell in the group Lkou're hiredJ advertising techniques. Finally, you might attempt to influence a decision about a social act, such as arranging a blind date for breaking tne rules may entail someone. punisnment; non-compliance may entail enforcement; but then again, may be not. In some groups, the rules cannot be in Tudor Gngland, if a british shipping changed, in which case attempts to merchant were unable to obtain payment for influence decisions are absurd. [I'll let goods he had delivered to a Dutch YOU capture my rook if you let me ignore merchant, he could ask the Court of the the fact that I’m in check for the next Admiralty for a "Letter of Reprisal," a few m0ves.J On tne other hand, in a document that permitted him to seize any highly reticulated authority Dutch structure, ship and to share in the worth of where the rules can all be changed, the ship and its cargo. Two aspects are politics are likely. The important simplest here. First, the Court in no political acts are those that attempt to way enforced its decision; it was not the change tne rules by following the British Navy that went out to capture a appropriate procedures Lworking witnin the Dutch ship, but the merchant himself (or estaolisnmentJ. Another method is to more likely, his agents). Second, he was behave as if the rules had changed in the permitted to commit what would otherwise way you seek Idefiance], which may mean be a highly illegal act. that you've committed an illegal act. You might decide not to fill a role in a group whose rules YOU disagree with [Boycott 213 Coors and Carl's Jr.], and YOU can also exploit the embedding of inconsistent authority structures [sit-ins at lunch counters]. APPLICATIONS Perhaps the clearest applications of this work would be in a natural language understanding program, and would be visible in at least three places. First, acts would be categorized by their legal status. Some of these would be explicit LJohnny is allowed to stay up until 10 o’clockJ, but most are quantified to some degree La11 personnel actions require tne manager's approvalJ. we would nave a more accurate Ji!uzini xepresentation than we now do for examples such as 88Mommy, can 1 go to the movies?" l!or kids (i.e., in the family-group), going to the movies might be directly labeled as requiring permission, but we might also be able to infer that from knowing that it involves leaving home and spending money, both of which require permission. Second, understanding an authority structure would enable a program to make the inferences needed to connect and explain events in a story. Example: "Mary forgot to renew a book from the library. They sent her a bill." Without some understanding of the library rules, the second sentence is difficult to explain. Third, we can use authority structures to make gredictions about people's actions. If Mary orders her son Johnny to go to bed, we can make a set of reasonaole predictions about Wnat he might do ana hOW Mary will respond. If Johnny's little sister orders Johnny to go to bed, tne predictions are quite different since the authority relationsnip is obviously different. 1 f' Sue loans Tom some money, thUS increasing his economic power, we can inf'er a state of indebtedness and expect him to repay her in some way. If a student slips a $20 bill in an examination book, he probably intends to induce a state of indebtedness on the part of his professor. I;ONCLUSIONS Our goal here been to organize general informat authority struct ures, providing framework in which we can specific cultura 1 instances and d relevant infere rices.. We envis authority struct ures as a necess of the represent ation for aroubs i has to on abo ut comm on descri be fine t he on usi ng w pa rt f pew le ACKNuWLoUGhnENTs my thanks to Lamar Hill, Gene Fisher, Bob bechtel, Dave Keirsey, and Steve Heiss for their constructive (and lively) aiscussions on this topic. 11 21 RCr'tinBNChS Jaime G. Carbonell. Sub.jective Understanding: Computer Models sf Belief Svstemg, PhD dissertation, kale University, 1979. Research Report 150, Yale Computer Science Department. Roger C. Schank and Jaime G. Carbonell. He: The Gettysburg Address: Representing social and political acts. nesearcn Report 127, kale Computer science Department, January 1970. whose behavior pattern is shared. 214	1980	55
51	REAL TIME CAUSAL MONITORS FOR COMPLEX PHYSICAL SITES* Chuck Rieger and Craig Stanfill Department of Computer Science University of Maryland College Park, MD 20742 ABSTRACT Some general and specific ideas are advanced. about the design and implementation of cau;llmn;ltorlng s P stems for complex sites such as a p ant or NASA missions control room. Such 'c,P:% monitors" and are interesting from l;idth ;zEl;heoretical engineering greatly improve viewpoints, existing man-machine interfaces to complex systems. INTRODUCTION Human understandin of physical facility, a ~NASA mission control large complex sue as a room or a nuclear power plant, is often a hapk;;a:d affair. Traditionally, once a site is knoyledge about. it resides ,in technical manual; $hlch tend to sit on the-she1 ) and in the minds the experts. As technlcal staff turnover occurs over an extended period of time, this knowledge tends to get misplaced, rearranged, or altogether forgotten, - at least - as knowledge. The result is da--to-day usua P working ly that no single individhl fully appreciates or understands the system as a whole; cope with it to the operators learn simply to the extent required for daily operations and maintenance. Of course, thisa;;;;; that when an emergency occurs, or when need to perform some unusual human ex ertise maneuver, the required the she1 P is often hard to locate. Manuals on are virtually worthless in most contexts, especially those in which time is critical. Even ~JF the expert.is on the site, it may take him too per orm the synthesis of a lar e set of parameters P to perceive the context of an emergency and necessary to diagnose the prob em. 9 Given this state of produces systems too deep1 high technology, which or broadly complex for an individual or reasona ly 41 sized group of individuals to -there is a clear need for "intelligent" comprehend, secondary systems for monitoring the primary systems. To develop such intelligent systems, we need (1) flexible representations for physical causality, (2) good human interfaces that accept descriptions of the physical world andrnk;;ii $szrlptions In tkese representations, (3) ComDrehension that are capable of relating the myriad states of the site to the causal description, then passing along only relevant and ~~por%:t information to the human operators, and efficient symbolic computation environments to support items l-3 in real time. This paper brief1 CAM (Causal Monitor) Project, ?I desc;;kesthe w ose 1s to develop a framework for the construction of intelligent, causal1 -b;s;zs real-time physica P monitors for ar~;:;~;~ The project is under funding by NASA Goddard Space Flight Center, and will result in a causal monitor generator Our aim rototype ?%kEmabout K ere is to advance some specific the conceptual architecture of such a ~~~~~~for man-made devices can be found in [6]. Background ideas on the concept of causal GOALS An ideal causal monitoring system does not replace humans, F,ut rather 'x;;$$&~~$,b,'oad~;; their abi:;j-;;s process enhances ability to collect and synthesize relevant information in time critical situations. The ideal system would ap ear to the human controllers as a small number o I; color CRT displays keyboard. The system would and a 1. 2. 3. 4. 5. 6. 7. continually sense and all in symbolically characterize sensors reflected their relative a way ttCi importance acceptable ranges in the current operating context continually verify that causally related sensor groups obey symbolic rules ex ressin P Fi the nature of their causal re ations ip be aware of the human onerator's exoressed intent (e.g.; "runn'ing check 32", preventative maintenance "executin mornin power-up") and adiust causal rela ions an f 8 ex ectations P and- lower-level to erances accordingly parametric have a knowledge of component and sensor failure modes and limit violations, their &&e;table precursors, probable their i..iicators, corrective proceduresc8%%'in the their automatic correction algorithms forma;: recommendations to the huma:) have a knowledge of standard "maneuvers", expressed action cog?irmable sequences with ste wise f: (for bot automatic consequences execution and as an archival reference for the human) continually s nthesize K all aspects of important t e system synthesis, and from t;k,s identify which aspects of system to display to the human controllers (in the absence of specific requests from the humans) decide on the most appropriate screen allocation piece of and displayu;;c,nique for each information the current context Most procedural knowledge about the would be on-line, primar and in a form meaningfu P t:YE:etmh the hu;nTeando:he corn uter. P The s stem would play the inte ligent ifl watt er continually monitorin H hundreds of sensors and refating them to the causa model for the cur:;zt o crating context. The system would summarize s P te s state via well-managed CRT displays, and would be capable not Otll pro able 4: of detecting irregularities and suggesting causes, consequences and corrective measures, but also of automatically carrying out both routin: ;;2a;m;rge:;z c;;;z;tive measures w",;; given the physical s !? - te would ge "self aware"c~~t~~~~efimited sense. The research described here is funded by NASA. Their support is gratefully acknowledged. 215 ARCHITECTURE OF THE CAM GENERATOR SYSTEM --- In the CAM Project we are concerned not with modelin specific * but rather with develop ng z general-purp~~~e'model P that can be imported to any site. Once there, it will interact with the site experts as causal they define the site's structure system. From the v~~si.~t~nte;~cti;;~sframe-driven acquisition generated. phase, the specific site knowledge CAM is The CAM generator system consists of several pieces: 1. 2. 3. 4. 5. 6. A collection of frame-like knowledge uni;; that collectively virtually any hysica incorporate fj Actuator, ActionStep, Component, -* FailureMode; Dis laypacket, Maneuver, Operatingcontext, wil !f OperatorIntent. Others emerge as the project progresses. A collection of procedz;s for interact;;5 with site scientists will engineers interactively describe the site by filling in frames. interface is semi-active * in a methodical way from the site engineers. compiled by the The information frame-driven interface results in a collection of and production data objects rule-like knowledge about their interrelationships. A collection of display primitives and display handling techniques ,A,iollection of primitive sensor-reading actuator-driving schemata (i.e., code schemata) A system for corn iling R the production rule description of t e site onto an efficient lattice-like data structure that obviates most run-time pattern matching A run-time maintenance for coordinatin system causal mode 9 the real time of the symbolic monitor. The frames and the nature of the knowledge acquisition the interface are described in [3]. Since real-time efficiency of the generated system is a critical issue, we devote the remainder of discussion here to the architecture of the Ek real-time monitor s stem Propagation Driven ifi which we have termed achi;e (PDM). the PROPAGATION DRIVEN MACHINES from its runtime system. CAM will require an extre;Eky hi,p$ throughput cannot tolerate much reason, we the basic machine cyc f eneral e. pattern matchin f,i; To e iminate the nee d" attern mate hing, ependency lattice we base our machine on a form of similar in structure described in [l], [2],[4] and [5]. The e%enE?% such a scheme is to re reient antecedent-consequent relations forward and b~ac~wa~,ra~~~sohhi~~ is useful in both The central data structure of the graph whose nodes PDM is a represent the current value of some parameter or proposition (e. ., value of sensor 23 is X" "the current condition in K chamber 19"ih'rghesl~~e~~ef,p:~~s~~ network nodes are the primitive sensor readers real time clocks whose PD8 spontaneous se uences all propagation in the net. Nodes ticking are connected by Above and Below in the links which reflect computational dependencies among noies. For example, form i if node Nl represents knowledge of the A and B and C) causes D then in the PDM net E wtid E-o-1, a TZlT?lrwiil be Above each of A, , . Some event in the model operator start-up, ',,"L, ;zFputation at a no E'HA'the net) identi-ffieFhz 6 modeled , R, that deals with some aspect environment, the causal e.g., the ru~es~~;t describe relationships Pressure S stem. Rules in ti?s set are then Cht%dr on r; 3 even;t;ll; Fe ;p,;gtzd azd from which they wi 1 all evaluated. An important PDM concept is structural augmentation of a node's of the network as a natural byproduct evaluation: evaluation of a s ueued newly nto rule causes the rule to be structurally knit the PDM network. any ex ressions During a node's evaluation, B not alread as new net nodes, P present in the net are create expression's inked to the referencing node via Above links for expansion themselves. This and placed in "downward chaining" structurally process P introduces t:e rule into the PDM net. Once rule's instz$l.Fd by. downward chaining, the value evaluation via begin to receive constant PDM "upward chaining": changes in lower nodes' values in the net Will throu h this structure, ropagate upward reeva uated. f causing hig er nodes to be R Nodes subject to reevaluation by this mechanism are queued on Q. By using the same for both structural integration of rules in;~~~~~ current context and for the values, the PDM scheme P ropagation of a lows d namic i: a graczfil and mechanism. it 0 viates context-shifting the matching. need for most run-time pittern To illustrate, suppose some node's evaluation k$s called for zfrule set for monitoring the degree 0 enness x Relief-Valve-14 (RV-14) to be swappe and in. The PDM system will, say, then evaluate others) a R? ueue up for comparing the (among . Pipe-P"4sii$?14yf rule, RV-14 again;t the R ressure of This leads to nitting process: Rl looks for RV-14 and PP-lz determinin trigger a higher node, for communicating e.g., the a display ,~reirn;~;; results to R operators. for This dual-purpy;: use of the PDM queue, staking out net structure and nam;;; performing in net node co;~~:~~ion;ak~; pf-ygztion of changes ",,y~lz,ic system. R topologically upward sprouting, AE~ .&tsp;ft;heoBett;;e g;;pagating garbage collected). The system is e.sr:,tizi$g while others are atrophying sompi ed production rule system, but one in compilation" is values. as dynamic as the propagation of Rule sets can come and go as naturally as values can be changed. Semantically, the lowest nodes of the PDM network, and those that are present initially spontaneously ticking real-time clocks whose Lb are ove pointers gob;;iEensor-reading nodes. certain montoring Additionally, expressions are represented by nodes initially in the net. As the site's state changes, these 'caretaker" rules (net nodes) will make reference to other rules not ;kE;,turally part of the net references come to queuing them up.':,' bg evaluated. their evalua begin ----. tion draws to participa them into the PDM net, where they te in the monitoring process. 216 When the context shifts in a way that causes the ;a;Et&;r node to be uninterested in the paged-in , the caretaker node drops its reference to iktwe?!?tsE?f"and the top members of the rule set. effectively cutting the Below link Nodes with no Above pointers and which are not marked as "permanent" are subject to PDM collection. Garbage.collection is inve:;: oEu:E origmzl;ddownward effectivef;rbX chaining that knit in set, structurally removes the entire rule set by following Below pointers. SUMMARY A CAM generator system is a s ecial of knowledge acquisition R causality in large, do:z: l;,s sites. requires (1) ~ornpl~~"~~"hys~cZ~~~ models of knowled 3 e elicitation from the site experts, models of the conce ts causa T (2)alframe-li e common to physical sites and their topology, and (3) an efficient, yet flexible a!? en ineering solution to thEfproblems of controll+ng arge symbolic knowledge in rea?y~~~~. The production rule-like PDM model appears to be a realistic approach to real-time monitoring: it exhibits the breadth and thorou hness of classical production rule system, % ut without thz usual problems of pattern matching and database maintenance. The area of data driven real-time causal models of corn lex sites appears to be a very fertile an 8 I ph sical large y unexplored area of AI research. Research in this area will he1 brin following areas of AI closer together: E %a:: rame- systems, knowledge acquisition, causal modeling, and efficient implementation techniques for real-time symbol manipulation. The domain ' manageable because we are modeling s stems in whit; deep problem solving is not a % issue, and theoretically interesting becausEY of its breadth-wise complexity. REFERENCES ill [21 E31 [41 [51 [61 London, P. E., Dependency Networks as Representation for I?~?~~~~ Solvers,. Modelling in Genera? Department of Computer University Report ?R-698, 1978. of Maryland, Technical McDermott, D., Computer Pro AIM-402, 197 . 7 Flexibility and Efficiency hinTa ram for Designing Circuits, , Rieger, C. and Stanfill, C., A Causal Monitor Generator System, Forthcoming TR, Department of Computer Science, University of Maryland, 1980. Shortliffe, E. H., MYCIN: A Rule Based Computer Program for Advising Physicians Regardin Antimicrobial Therapy Selection, Memo ifi IM-251, Artificial Intelligence Laboratory, Stanford University, 1974. Sussman, G. J. and Knight, I. F. iZ:tal Holloway J. Computer Aided Evolutionar K Desi n for 1999. Integrated Systems, IT, AHM-526, Rieger, C., System of Representation $orACo$??~;E~~~~~ and Grinberg, M. Design, Artificial Intelligence and Pattern puteran~:d~~7D~gn, J. co 217	1980	56
52	GENERATING RELEKWC FXPLANATICNS: NATURAL LANGUAGERESPCNSES 'I0QJESTIONSAl3OUl?DATAHASE STRUCIWB* Kathleen R. McKeown Department of Ccmputer and Information Science The Moore School University of Pennsylvania Philadelphia, Pa. 19104 ABSTRACTT The research described here is aimed at unresolved problems in both natural language generation and natural language interfaces to database systems. How relevant information is selected and then organized for the generation of responses to questions about database structure is examined. Due to limited space, this paper reports on only one method of explanation, called 'kanpare and contrast". In particular, it describes a specific constraint on relevancy and organization that can be used for this response type* I INTRODUCTICN Following Thanpson [141, the process of generating natural language may be divided into two interacting phases: (1) determining the content, force, and shape of what is to be said (the "strategic aznponent") and (2) transforming that message from an internal representation into English (the "tactical ccmponent"). The decisions made in the strategic component are the focal point of the current work. These decisions are being investigated through the development of a system for answering questions about database structure that require sane type of explanation or ' description. This mrk, therefore, has two goals: (1) providing a facility that is lacking in many natural language interfaces to database systems, and (2) exercising theories about the nature of natural language generation. The system has been designed and implementation is in its beginning stages [12]. The decisions that the strategic component of a natural language generator must make are of two different types: decisions of a semantic/pragmatic nature and decisions that are structural in nature, Given a guestion, the strategic component must select only that information relevant to its answer (semantic/pragmatic decisions). What is selected must then be organized appropriately (structural decisions). These two types of decisions are the issues this work addresses. Not covered in this paper are the syntactic issues and problems of lexical choice that a tactical ccmponent must address. * This work was partially supported by NSF Grant MCS 79-08401 and an IBM Fellowship. Structural issues are important since the generation of text and not simply the generation of single sentences is being considered. Anumber of organizational principles that can be used for structuring expository text have been identified WI. These are termed ccmpare and contrast, description, top-down 7 illustration through example, definition, bottcm-up description, and t3IELLogy. In this paper, discussion is limited to compare and contrast and its effect on the 7 organization and selection processes. II THE APPLICATION Current database systems, including those enhanced by a natural language interface (e.g. [61), are, in most cases, limited in their responses to providing lists or tables of objects in the database.* Thus, allowable questions are those which place restrictions upon a class of objects occurring in the database. To ask these kinds of questions, a user must already knaw what kind of information is stored in the database and must be aware of how that information is structured. The system whose design I am describing will answer questions about the structure and organization of the database (i.e. - meta-questions)*. The classes of meta-questions which will be accepted by the system include requests for definitions, requests for descriptions of information available in the database, questions about the differences between entity-classes, and questions about relations that hold between entities. Typical of such meta-questions are the following, taken from Malhotra [7]: What kind of data do you have? What do you know about unit cost? What is the difference between material cost and production cost? What is production cost? -------------------------------------------------- Note that in sane systems, the list (especially in cases where it consists of only one object) may be embedded in a sentence, or a table may be introduced by a sentence which has been generated by the system ( e.g. - [43). ** I am not addressing the problem of deciding whether the question is about structure or contents. 306 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. III KNWLEDGIZ BEBBESENTATICJN In order for a system to answer meta-questions, it requires information beyond that normally encoded in a database schema. The knowledge base used in this system will be based on a logical database schema similar to that described by Mays 191. It will be augmented by definitional information, specifying restrictions on class membership, and contingent information, specifying attribute-values which hold for all members of a single class. A generalization hierarchy, with mutual exclusion and exhaustion on sub-classes, will be used to provide further organization for the information. For more detail on the knowledge representation to be used, see WI. Iv SAMf?LEQTJESTIONS Textual responses to meta-questions must be organized according to sane principle in order to convey information appropriately. The compare and contrast principle is effective in answering questions that ask explicitly about the difference between entity-classes occurring in the database. (It is also effective in augmenting definitions but this would require a paper in itself.) In this paper, the following two questions will be used to illustrate how the strategic ccmponent operates: (1) What is the difference between a part-time and a full-time student? (2) What is the difference between a raven and a writing desk? V SELECTION OF RELEVAW IEFOBMATION - Questions about the difference between entities require an assumption on the part of the speaker that there is sane similarity between the items in question. This similarity must be determined before the ways in which the entities differ can be pointed out. Entities can be contrasted along several different dimensions, all of which will not necessarily be required in a single response. These include: attributes super-classes subclasses relations related entities For sane entities, a comparison along the lines of one information type is more appropriate than along others. For example, ccmparing the attributes of part-time and full-time students (as in (A) belaw) can reasonably be part of an answer to question (1) r but a comparison of the attributes of raven and writing desk yields a ludicrous answer to question (2) (see (B) below). (A) A part-time student takes 2 or 3 courses/semester while a full-time student takes 3 or 4. (B) A writing desk has 4 legs while a raven has only 2. One factor influencing the type of information to be described is the "conceptual closeness" of the entities in question. The degree of closeness is indicated by the distance between the entity-classes in the knowledge base. Three features of the knowledge base are used in determining distance: the generalization hierarchy, database relationships, and definitional attributes. A test for closeness is made first via the generalization hierarchy and if that fails, then via relationships and definitional attributes. A successful generalization hierarchy test indicates the highest degree of closeness. Usually, this will apply to questions about two sub-types of a common class, as in: What is the difference between production cost and material cost? What is the difference between a part-time and a full-time student? In the generalization hierarchy, distance is determined by two factors: (1) the path between the entity-classes in question and the nearest ccnnmon super-class; and (2) the generality of the common super-class (path between the common super-class and the root node of the hierarchy). The path is measured by considering its depth and breadth in the generalization hierarchy, as well as the reasons for the branches taken (provided by the definitional attributes). Entities are considered close in concept if path (1) is sufficiently short and path (2) sufficiently long. If the test succeeds, a discussion of the similarity in the hierarchical class structure of the entities, as well ,as a comparison of their distinguishing attributes, is appropriate. Although the entities are not as close in concept if this test fails, sane similarities may nevertheless exist between them (e.g. - consider the difference between a graduate student and a teacher). A discussion of similarities may be based on relationships both participate in (e.g. - teaching) or entities both are related to (e.g. - courses). In other cases, similarities may be based on definitional attributes which hold for both entities. For both cases, a discussion of the similarities should be augmented by a description of the difference in hierarchical class structure. Entities that satisfy none of these tests are very different in concept, and a discussion of the class structure which separates them is informative. For example, for question (2) above, indicating that ravens belong to the class of animate objects , while writing desks are inanimate results in a better answer than a discussion of their attributes. 307 VI TEXT ORGATYIZATION There are several ways in which a text can be organized to achieve the canpare and contrast orientation. One approach is to describe similarities between the entities in question, followed by differences. Alternatively, the response can be organized around the entities themselves; a discussion of the characterizing attributes of one entity may be followed by a discussion of the second. Finally, although the question may ask about the difference between entities, it may be impossible to compare them on any basis and the ccanpare and contrast must be rejected. The determination of the specific text outline is made by the structural processor of the strategic oamponent. On the basis of the input question, the structural processor selects the organizing principle to be used (for the two sample questions, compare and contrast is selected). Then, on the basz of information available in the knowledge base, the decision is reevaluated and a ccxrnnitment made to one of the outlines described above. Because of this reliance on semantic information to resolve structural problems, a high degree of interaction must exist between the structural processor and the processor which addresses semantic and pragmatic issues. One type of semantic information which the structural processor uses in selecting an outline is, again, the distance between entity-classes in the knowledge base. For entities relatively close in concept, like the part-time and the full-time student, the text is organized by first presenting similarities and then differences. By first describing similarities, the response confirms the questioner's initial assumption that the entities are similar and provides the basis for contrasting them. Two entities which are very different in concept can be described by presenting first a discussion of one, followed by a discussion of the other. Entities which cannot be described using the ccmpare and contrast organization are those which have very little or no differences. For example, if one entity is a sub-concept of another, the two are essentially identical, and the compare and contrast organizing principle must be rejected and a new one selected. VII STRATEGICPRCCESSING Although dialogue facilities between the structural processor (SIR) of the strategic ccmponent and the semantic/pragmatic processor (S&P) have not yet been implemented, the following hypothetical dialogue gives an idea of the intended result. Question (1): What is the difference between a part-time and a full-time student? STR: notes form of query and selects COMPARE AND -T S&P: queries knowledge base: DISTANCE(part-time,full-time) --> very close (same immediate super-classes) SIR: retains COMPARE AND CONTRAST selects outline: SIMIL,ARITIES DIEFERENCFS: ATTRIBUTE-TYPE1 . . ATTRIBUI'E-TYPEn CONSEQUENCES* S&P: queries knowledge base and fills in outline: SIMILARITIES super-classes(part-time,full-time) --> graduate student attribute/value(part-time,full-time) --> degree-sought = MS or PhD DIFFERENCES attribute/value(part-time,full-time) -> courses-required = part-time: 1 or 2/semester full-time: 3 or 4/sanester -> source-of-inme = part-time: full-time job full-time: unknown CONSEQUENCES none STR: further organizational tasks, not described here, include determining paragraph breaks (see [12]). Here there is 1 paragraph. The tactical canponent, with additional information frcrn the strategic component, might translate this into: Both are graduate students going for a masters or Phd. A full-time student, however, takes 3 or 4 courses per semester, while a part-time student takes only 1 or 2 in addition to holding a full-time job. After engaging in similar dialogue for question (2) p the strategic mponent might produce outline (C) belaw, which the tactical ccanponent could translate as (D): (C) RAVEN FACTS: super-classes(raven) = raven E bird E animate object WRITING DESK FACTS: super-classes(writing desk)= writing desk E furniture E inanimate object CONS~S: bird and furniture inaznpatible 2 different objects (D) A raven is a bird and birds belong to the class of animate objects. A writing desk is ------------------------------------------------- * CDNS~S here involve OdY minimal inferences that can be made about the class structure. 308 a piece of furniture and furniture belongs to the class of inanimate objects. A bird can't be a piece of furniture and a piece of furniture can't be a bird since one is animate and the other isn't. A raven and a writing desk therefore, are 2 very different things. VIII RELATEDRESEARCH INGHNFPATICN - Those working on generation have concentrated on the syntactic and lexical choice problems that are associated with the tactical component (for example, DOI, [31, D-31, D11) - Research on planning and generation ([l], 121) comes closer to the problems I am addressing although it does not address the problem of relevancy and high-level text organization. Mann and Moore [81 deal with text organization for one particular domain in their generation system, but avoid the issue of relevancy. The selection of relevant information has been discussed by Hobbs and Robinson [5] who are interested in appropriate definitions. IX CGNCLus1oNs The effects of a specific metric, the "conceptual closeness" of the items being ccmpared, were shown on the organization and selection of relevant information for meta-question response generation. Other factors which influence the response, but were not discussed here include information about the user's knowledge and the preceding discourse. Further research will attempt to identify specific constraints from these two sources which shape the response. The research described here differs frcm previous work in generation in the following ways: 1. Previous work has concentrated on the problems in the tactical ccmponent of a generator. This work focusses on the strategic cunponent: selecting and organizing relevant information for appropriate explanation. 2. While previous work has dealt, for the most part, with the generation of single sentences, here the emphasis is on the generation of multi-sentence strings. When implemented, the application for generation will provide a facility for answering questions which the user of a database system has been shown to have about the structure of the database. In the process of describing or explaining structural properties of the database, theories about the nature of text structure and generation can be tested. ACKNB I would like to thank Dr. Aravind K. Joshi, Dr. Bonnie Webber, and Eric Mays for their detailed comments on the content and style of this paper. -S [l] Appelt, D. E. "Problem Solving Applied to Language Generation" In Proc. of the 18th Annual Meeting of the ACL. 1980, pp.-59x.- Philadel$icEz [2] Cohen, P. R. "On Knowing What to Say: Planning Speech Acts," Technical Report # 118, University of Toronto, Toronto, Canada, 1978. I31 Goldman,N. "Conceptual Generation" In R. C. Schank led.1 , Conceptual Information Processing. North-Holland Publishing Co. Amsterdam, 1975. [41 Grishman, R. "Response Generation ' Question-Answering Systems" In Proc. of the 17:: Annual Meeting of the ACL. --- La Jolla, Ca., August, --- 1979, pp. 99-101. [5] Hobbs, J. R. and J. J. Robinson "Why Ask?", Technical Note 169, SRI International, Menlo Park, Ca., October 1978. [63 Kaplan, S. J. "Cooperative Responses Fran a Portable Natural Language Data Base Query System", Ph.D. Dissertation, Ccmputer and Information Science Department, University of Pennsylvania, Pa., 1979. [71 Malhotra, A. "Design Criteria for a Knowledge-based English Language system for Management: an Experimental Analysis", MAC TR-146, MIT, Cambridge, Ma., 1975. [81 Mann, w. c. and J. A. Moore "Computer as Author - Results and Prospects", ISI/RR-79-82, ISI, Marina de1 Rey, Ca., 1980. [91 Mays, E. "Correcting Misconceptions About Database Structure" In Proc. of the Conference of the CSCSI. Victoria, May 1980, pp. 123-128. BritisrCaia, Canada, I101 McDonald, D. "Steps Toward a Psycholinguistic Model of Language Production", MIT AI Lab Working Paper 193, MIT, Cambridge, Ma., April 1979. [ll] McKeown, K. R. "Paraphrasing Using Given and New Information in a Question-Answer System", In Proc. of the 17th Annual Meeting of the ACL. La e--p --- Jolla, Ca., August, 1979, pp. 67-72. 1121 McKeown, K. R. "Generating Explanations and Descriptions: Applications to Questions about Database Structure," Technical Report # MS-CIS-80-9, University of Pennsylvania, Philadelphia, Pa., 1979. [131 Simns, R. and J. Slccum "Generating English Discourse from Semantic Networks." CACM. 15:lO (1972) pp. 891-905. [14] Thompson, H. "Strategy and Tactics: A Model for Language Production" In Papers fran the 13th Regional Meeting, Chicago Linguistic S&ieFlm 309	1980	57
53	THE SEMANTIC INTERPRETATION OF NOMINAL COMPOUNDS * Timothy Wilking Finin Coordinated Science Laboratory University of Illinois Urbana IL 61801 ABSTRACT This paper briefly introduces an approach to the problem of building semantic interpretations of nominal ComDounds, i.e. sequences of two or more nouns related through modification. Examples of the kinds of nominal compounds dealt with are: "engine repairs", "aircraft flight arrival", ~aluminum water pump", and "noun noun modification". I INTRODUCTION This paper briefly introduces an approach to the problem of building semantic COmDOundS, interpretations of nominal i.e. sequences of two or more nouns related through modification. The work presented in this paper is discussed in more detail in [3] and [4]. The semantics of ncminal compounds have been studied, either directly or indirectly, by linguists and AI researchers. impressive In an early study, Lees [8 3 developed an taxoncm of the forms. II81 and Downing f 21 More recently, Levi have attempted to capture the linguistic regularities evidenced by nominal ccmpounding. Rhyne explored the problem of generating canpounds from an underlying representation in [JO]. Bra&man Cl] used the problem of interpreting and representing nominal compounds as an example domain in the development of his SI-Net representational formalism in [l]. Gershman [6] and McDonald and Hayes-Roth [9] attempt to handle noun- noun modification in the context of more general semantic systems. In this work, the interpretation of nominal compounds is divided into three intertwined subproblems: lexical internretation (mapping words into concepts), modifier narsina (discovering the structure of compounds with more that two nominalsj and concert modification (assigning an interpretation to the modification of one concept by another), this paper. This last problem is the focus of The essential feature of this form of modification is that the underlying semantic relationship which exists between the two concepts is not explicit. Moreover, a large number of relationships might, in principal, exist between the two concepts. The selection of the most appropriate one can depend, in general, on a host of semantic, pragmatic and contextual factors. been As a part of this research, a computer program has written which builds an appropriate semantic interpretation when given a string of nouns. This program has been designed as one component of the natural language question answering system JETS 151, a successor to the PLANES query system [133. The interpretation is done by a set of semantic interpretation rules. Some of the rules are very specific, capturing the meaning of idioms and canned phrases. Other rules are very general, representing fundamental case-like relationships which * Ailthor's current address: Department of Ccmputer and Information Science, Philadelphia, PA 19104. University of Pennsylvania, This work was supported by the Office of Naval Research under Contract NOOOlll-75-C-0612. can hold between concepts. A strong attempt has been made to handle as much as possible with the more general, highly productive interpretation rules. The approach has been built around a frame-based representational system (derived from FRL [ill) which represents concepts and the relationships between them. The concepts are organized into an abstraction hierarchy which supports inheritance of attributes. The same representational system is used to encode the semantic interpretation rules. An important part of the system is the concert matcher which, given two concepts, determines whether the first describes the second and, if it does, how well. II THE PROBLEM Let's restrict our attention for a moment to the simplest of ccmpounds - those made up of just two nouns, both of which unambiguously refer to objects that we know and understand. What is the fundamental problem in interpreting the modification of the second noun by the first? The problem is to find the underlying relationship that the utterer intends to hold between the two concepts that the nouns denote. For example, in the compound "aircraft engine" the relationship is part of, in "meeting room" it is location, in "salt water" it is dissolved in. There are several aspects to this problem which make it difficult. First, the relationship is not always evident in the surface form of the compound. What is it about the compound GM cars which suggests the relationship made bv? The correct interpretation of this compound depends on our knowledge of several facts. We must know that a is the name of an organization that manufactures things, and in particular, automobiles. Another fact that helps to select this interpretation is that the identity of an artifact's manufacturer is a salient fact. It is even more important when the artifact is an automobile (as opposed to, say, a pencil). A second source of difficulty is the general lack of syntactic clues to guide the interpretation process. The interpretation of clauses involves discovering and making explicit the relationships between the verb and its "arguments", e.g. the subject, direct object, tense marker, aspect, etc. Clauses have well developed systems of syntactic clues and markers to guide interpretation. These include word order (e.g. the agent is usually expressed as the subject, which canes before an active verb), prepositions which suggest case roles, and morphemic markers. None of these clues exists in the case of nominal compounds. Third, even when the constituents are unambiguous, the result of compounding them may be multiply ambiguous. For example, a woman doctor may be a doctor who is a woman or a doctor whose patients are women. Similarly, Chicano fliQhtS may be those bound for Chicago, coming from Chicago or even those making a stop in Chicago. A fourth aspect is that compounds exhibit a variable degree of lexicalization and idiomaticity. In general, 310 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. the same compound form is used for lexical items (e.g. duck soup, hanger queen) and completely productive expression (e.g. engine maintenance, faculty meeting). Finally, I point out that it is possible for any two nouns to be ccmbined as a compound and be meaningful in some context. In fact, there can be arbitrarily many possible relationships between the two nouns, each relationship appropriate for a particular context. III m INTERPRETATION RULES The implemented system contains three components, one for each of the three sub-problems mentioned in the introduction. The lexical internreter maps the incoming surface words into one or more underlying concepts. The concert modifier takes a head concept and a potential modifying concept and produces a set of possible interpretations. Each interpretation has an associated score which rates its "likelihoodm. Finally, the modifier parser applies a parsing strategy which compares and combines the local decisions made by the other two components to produce a strong interpretation for the entire compound, without evaluating all of the possible structures (the number of which increases exponentially with the number of nouns in the ccmpound). The remainder of this paper discusses some of the interpretation rules that have been developed to drive the concept modifier. Three general classes of interpretation rules have been used for the interpretation of nominal compounds. The first class contains idiomatic rules - rules in which the relationship created is totally dependent on the identity of the rule's constituents. These rules will typically match surface lexical items directly. Often, the compounds will have an idiomatic or exocentric (R) meaning. As an example, consider the Navy's term for a plane with a very poor maintenance record - a "hanger queentl. The rule to interpret this phrase has a pattern which require an exact match to the words *hanger" and lqueenl. The second class consists of productive rules These rules attempt to capture forms of modifian which are productive in the sense of defining a general pattern which can produce many instantiations. They are characterized by the semantic relationships they create between the modifying and modified concepts. That is, the nature of the relationship is a property of the rule and not the constituent concepts. The nature of the concepts only determines whether or not the rule applies and, perhaps, how strong the resulting interpretation is. For example, a rule for dissolved in could build interpretations of such compounds as lsalt water" and (tsugar water" and be triggered by compounds matching the description: (a NominalCompound with Modifier matching (a ChemicalCompound) Modified matching (a Liquid) preferably (a Water)) The third class contains the structural rules. These rules are characterized by the structural relationships they create between the modifying and modified concepts. The semantic nature of the relationship that a structural rule creates is a function of the concepts involved in the modification. Many of these rules are particularly useful for analyzing compounds which contain ncminalized verbs. IV STRUCTURAL RULES I have found this last class to be the most interesting and important, at least from a theoretical * An exocentric compound is one in which the modifier changes the basic semantic category of the head noun, as in hot dog and ladv finger. perspective. This class contains the most general semantic interpretation rules - precisely the ones which help to achieve a degree of closure with respect to semantic coverage [5]. Similar structural rules form the basis of the approaches of Bra&man Cl] and McDonald and Hayes-Roth C91. This section presents some of the structural rules I have catalogued. Each rule handles a compound with two constituents. RULE: RoleValue + Conceot. The first structural rule that I present is the most common. It interprets the modifying-concept as specifying or filling one of the roles of the modified concept. Some examples of compounds which can be successfully interpreted by this rule are: engine repair (a to-repair with object = (an engine)) January flight (a to-fly with time = (a January)) F4 flight (a to-fly with vehicle = (an F4)) engine housing (a housing with superpart = (an engine)) iron wheel (a wheel with raw-umterial = (a iron)) Note that when the compound fits the form *subject+verb" or "object+verbll this works very nicely. The applicability of this rule is not limited to such compounds, however, as the last two examples demonstrate. To apply this rule we must be able to answer two questions. First, which of the modified concept's roles can the modifier fill? Obviously some roles of the modified concept may be inappropriate. The concept for the to-repair event has many roles, such as an agent doing the repairing, an object being repaired, an instrument, a location, a time, etc. The concept representing an engine is clearly inappropriate as the filler for the agent and time roles, probably inappropriate as a filler for the location and instrument roles, and highly appropriate as the object's filler. Secondly, given that we have found a set of roles that the modifier may fill, how do we select the best one? Moreover, is there a way to measure how well the modifier fits a role? Raving such a figure of merit allows one to rate the overall interpretation. The process of determining which roles of a concept another may fill and assigning scores to the alternatives is called role fitting. This process returns a list of the roles that the modifier can fill and, for each, a measure of how "good" the fit is. Each possibility in this list represents one possible interpretation. Not all of the possibilities are worthy of becoming interpretations, however. A selection process is applied which takes into account the number of possible interpretations, their absolute scores and their scores relative to each other. Making a role fit into an interpretation involves making a new instantiation of the modified concept, and filling the appropriate role with modifier. Details of this process are presented in the next section. RULE: Conceot + RoleValue. This rule is similar to the first, except that the concepts change places. In interpretations produced by this rule, the modified concept is seen as filling a role in the modifier concept. Note that the object referred to by the compound is still an instance of the modified concept. Some examples where this rule yields the most appropriate interpretation are: drinking water (a water which is (an object of (a to-drink))) washing machine (a machine which is (an instrument of (a to-wash))) maintenance crew (a crew which is (an agent of (a to-maintain))) Again, the application of this rule is mediated by the role fitting process. 311 RULE' Concent + RoleNominal. This rule is app1i.G when the modified concept is in the class I call role nominals, nouns that refer to roles of other underlying concepts. &glish has but one productive system for naming role ncminals: the agent of an verb can commonly be referenced by adding the -er or -or suffix to the verb stem. This should not hide the possibility of interpreting many concepts as refer@ to a role in another related concept. Some examples are: a student is the recipient of a teaching, flowing, a pipe is the conduit of a a pump is the instrument of a pumping, and a book is the obiect of a reading. This rule tries to find an interpretation in which the modifier actually modifies the underlying concept to which the role nominal refers. For example, given IF4 Pilot', the rule notes that ~pilotl is a role nominal refering to the agent role of the to-fly event and attempts to find an interpretation in which lF4" modifies that to-fly event. The result is something like "an F4 pilot is the agent of a to-fly event in which the vehicle is an F4". Some other examples are: cat food (an object of (a to-eat with agent = (a cat))) oil pump (an instrument of (a to-pump with object = (an oil))) dog house (a location of (a to-dwell with agent = (a dog))) Viewing a concept as a role nominal (e.g. food as the object of eating) serves to tie the concept to a characteristic activity in which it participates. It is very much like a relative clause except that the characteristic or habitual nature of the relationship is emphasized. RULE: RoleNominal + Concert. This rule is very similar to the previous one exceot that it anolies when the modifying concept is a role -nominal. Thgaction is to attempt an interpretation in which the modification is done, concept not by the first concept, but by the underlying to which it refers. For example, given the compound "pilot school", we can derive the concept for "an organization that teaches people to fly". This is done by noting that pilot refers to the agent of a to-fly event and then trying to modify "school" by this "to- fly". This, in turn, can be interpreted by the Conceot + RoleNominal rule if school is defined as 'an organization which is the agent of a to-teach". This leads to an attempt to interpret to-fly modifying to- teach. The RoleValue + Conceot rule interprets to-fly as filling the object (or discipline) role of to-teach. Some other examples of ccmpounds that benefit from this interpretation rule are newspaper glasses (glasses used to read a newspaper), driver education (teaching rfTple to drive), food bowl (a bowl used to eat food out . Other Structural Rules Other structural interpretation rules that-?-have identified include SDecific+Generic which applies when the modifier is a specialization of the modified concept (e.g. F4 planes, boy child), Generic+SDecific tiich applies when the the modifier is a generalization of the modified concept (e.g. Building NE43, the integer three), Eauivalence in which the resulting concept is descendant frcm both the modifier and modified concepts (e.g. woman doctor) and Attribute Transfer in which a salient attribute of the modifier is transferred to the modified concept (e.g. iron will, crescent wrench). V ROLE FITTING The process of role fitting is one in which we are given two concepts, a RoleValue-and a Host, and attempt to find appropriate roles in the Host concept in which the RoleValue concept can be placed. Briefly, the steps carried out by the program are: [ll Collect the local and inherited roles of the Host concept; [2] Filter out any inappropriate ones (e.g. structural ones); 133 For each remaining role, compute a score for accepting the RoleValue concept; [4] Select the most appropriate role(s). In the third step, the goodness-of-fit score is represented by a signed integer. Each role of a concept is divided into an arbitrary number of facets, each one representing a different aspect of the role. In computing the goodness of fit measure, each facet contributes to the overall score via a characteristic scoring function. The facets which currently participate include the following: Requirements descriptions candidate value &match. Preferences descriptions candidate value should match DefaultValue a default value. TypicalValues other very ccmmon values for this role. Modality one of Optional, Mandatory, Dependent or Prohibited. Multiplicity maximum and minimum number of values. Salience a measure of the role's importance with respect to the concept. For example, the scoring function for the reauirements facet yields a score increment of +I for each requirement that the candidate value matches and a negative infinity for any mismatch. For the preferences facet, we get a +4 for each matching preference description and a -1 for each mismatching description. The salience facet holds a value frcm a 5 point scale (i.e. Verylow, Low, Medium, High, VeryHigh). Its scoring function maps these into the integers -1, 0, 2, 4, 8. VI SUMMARY This paper is a brief introduction to an approach to the task of building semantic interpretations of nominal compounds. A nominal compound is a sequence of two or more nouns or nominal adjectives (i.e. non-predicating) related through modification. The concepts which the nouns (and the compound) denote are expressed in a frame- based representation system. The knowledge which drives the interpretation comes from the knowledge of the concepts themselves and from three classes of interpretation rules. Examples of the most general class of interpretation rules have been given. L-11 L-21 [31 141 151 C61 [71 181 r.91 II101 [ill [I21 El31 REFERENCES Tennant, H., i%perience with the Evaluation of Natural Language aestion Answerers, Proc. IJCAI- $Ltpkp Japan, 1979. IAn English language Question Answering S steh fob a Large Relational Data Base", CACM, vol. 2Y , PP 526-539, 1978. 312	1980	58
54	TOWARDS AN AI MODEL OF ARGUMENTATIGA. Lawrence Birnbaum, Margot Flowers, and Rod McGuire Yale University Department of Computer Science New Haven, Connecticut Abstract This paper describes a process model of human argumentation, and provides examples of its operation as implemented in a computer program. Our main concerns include such issues as the rules and structures underlying argumentation, how these relate to conversational rules, how reasoning is used in arguments, and how arguing and reasoning interact. implemented, and using them the program is capable of participating in the following argument fragment, concerning the question of who was responsible for the 1967 Arab-Israeli war. The program can assume either the Israeli or the Arab point of view. [ll Arab: Who started the 1967 War? [ 21 Israeli: The Arabs did, by blockading the Straits of Tiran. Introduction [31 Arab: But Israel attacked first. Engaging in an argument is a task of considerable complexity, involving the coordination of many different abilities and knowledge sources. Some questions that arise in trying to construct a process model of argumentation include: What sub-tasks comprise argumentation? What are the argumentation ? rules underlying What representations of argument structure are necessary to support these argument rules? What are the conversational rules required for dealing with arguments as a specific type of dialogue? How is the ability to reason about the world used in the argumentation process? How do reasoning interact? and argumentation To address these questions, we are in the process of building a system, called ABDUL/ILANA, that can adopt a point of view and engage in a political argument, supporting its beliefs with argument techniques using appropriate facts and reasoning. The program takes natural language input and produces natural language output. All of the rules and mechanisms described in this paper have been This work was supported in part by the Advanced Research Projects Agency of the Department of Defense, monitored by the Office of Naval Research under contract N00014-75-C-1111, and in part by the National Science Foundation under contract IST7918463. [41 Israeli: According to international law, blockades are acts of war< [51 Arab: Were we supposed to let you import American arms through the Straits? [6l Israeli: Israel was not importing arms through the Straits. The reason for the blockade was to keep Israel from importing oil from Iran. No matter which point of view the program adopts, it uses the same principles of argumentation, abstract knowledge, and historical facts. In this way, the argument is not over the facts but instead over the interpretation of the facts, slanted by the sympathies of each arguer. Arpument tasks What tasks must an arguer perform when faced with his opponent's responses? Before anything else, each input must be transformed into a meaning representation for use in further processing (see Birnbaum and Selfridge (1980)). But much more is necessary. An arguer must understand how new input relates to the structure of the argument as a whole, and this task usually requires making non-trivial inferences. For example, consider what the program (in the role of an Israeli) must do to relate the Arab's input [31, "But Israel attacked first", to the utterances which preceded it. The Israeli must realize that [31 constitutes evidence for the unstated claim that Israel started the 1967 War, which is contrary to his claim in [21. An arguer must also relate new input to what he knows about the domain, for two reasons. First, this allows the arguer to determine whether or not 313 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. he believes that some claim put to him is true. knowledge of how to look through this network for a Second, in doing this he can uncover relevant point which can be attacked or defended. As an information which may be useful in forming a example of the use of such rules, consider how the response. program (operating as the Arab) generates a Once the understanding sub-tasks sketched out response to input 141: above have been accomplished, an arguer must decide [41 Israeli: According to international how to respond. This involves several levels of law, blockades are acts of war. decision. First, he must make large-scale strategic choices, such as whether to attack his opponent's claims, defend his own, change the This point stands in a support relation to the previous Israeli claims of utterance [21: subject, and so forth. Second, he must determine which points should be taken up, and how evidence [2al The Arabs started the war. can be provided for or against those points. A A Third, he must use his knowledge and reasoning I I ability in order to actually produce such evidence. attack This final step ends with the generation of a response in natural language (see McGuire (1980)). I support I ---[3al The Israelis ArPument structures and rules -- I started the war. A There are basically two ways that propositions in an argument can relate to each other. One point [2bl The Arab blockade led to the war. 6 I support can be evidence for another point, in which case I the argument relationship is one of support, or it I [3bl The Israeli attack can challenge the other point, in which case the support led to the war. relationship is one of attack. Thus, for example, the ultimate analysis of the Israeli's statement: I 141 Blockades are acts of war. [ 21 Israeli: The Arabs did, by blockading the Straits of Tiran. Once the program has decided to follow an attack requires not only understanding strategy, the it needs to find a weak point in two separate the opponent's argument. In this case a search propositions: rule suggests traversing up the support links in the graph, starting with the most recent [2al The Arabs started the war. input. The first item considered is 141, the Israeli's claim that blockades are acts of war. However this [2bl The Arab blockade led to the war. proposition was already checked during the understanding phase and found to be one of. the but also realizing that the second proposition [2bI program's beliefs. stands in a relation to the first Thus, it is not a good support candidate for attack. Traversing the support link proposition 112al. The argument relations are leads to [2bl, the proposition that the Arab themselves complex structures, including such blockade led to the war. However, this too was f information as which inference rules enabled the establishment of verified during understanding. Following one more the relation, (See de Kleer et link leads to (1977) and Doyle (1979) for [2al, the claim that the Arabs al. a discussion of related techniques.) started the war, which the Arab does I& believe to be true. Hence, this is a good candidate for attack. One motivation for having these argument relations is that their local "geometry" can be Now the program considers the three tactics used by argument tactics in determining how to for attacking this proposition's respond to the input. simple support These are rules that relation. describe Tactic (a>, attacking the options as the main point, to how to go about attacking or defending a proposition based on the has already been used once, as can be determined by inspecting the argument graph. Tactic argument relations in which it (b), takes part. For attacking the evidence, can't be used in this case example, one such rule coordinates the three ways to attack a simple support relationship: since the evidence has already been rejected as a candidate for attack by the argument graph search rule. (a> Attack the main point directly; This leaves tactic cc>, attacking the claim that the evidence is adequate support for the main (b) Attack the supporting evidence; point. (c) Attack the claim that the evidence Having decided to attack this support relation between proposition [2al, that the Arabs started gives support for the main point. the war, and [2bl, that the Arab blockade led to the war, The representation of the entire history of the program now attempts to do that by inferring a justification for the blockade, using the argument, as a network of propositions its abstract knowledge of blockades and its factual connected by argument relations, forms an arpument knowledge of U.S. military aid to Israel. This graph. Argument graph search rules - - embody justification is ultimately generated as: 314 151 Arab: Were we supposed to let you import American arms through the Straits? An interesting point about question [51 is that it has the form of the standard argument gambit of asking one's opponent to support or justify a position. What makes the question rhetorical is the assumption that, in this case, there is no justification for demanding that arms importation be allowed, Reasoning & memory &I arguments We have been particularly concerned with investigating how reasoning and memory search interact with the argument process. Reasoning in an argument is not simply blind inference: requirements imposed by the structure of the argument constrain when and how inferences should be made. For example, consider how the program, when adopting the Israeli point of view, responds to question [ll: El1 Arab: Who started the 1967 War? [2al Israeli: The Arabs did, [2b] Israeli: . . . by blockading the Straits of Tiran. In this case, the generation of [2al does not require the use of argument rules or episodic memory retrieval. a "gut reaction" Instead it is derived by use of rule that always assigns blame for "bad" events to some other participant. However, as soon as such a claim is put forth in a serious argument, it must be supported. The support [2bl is then produced by a more complex use of inferential memory, activated only in response to the argument goal of providing support. Conversely, reasoning and memory guide the argument process by discovering information that affects subsequent argument choices. In particular, good rebuttals may often be found in the course of understanding the input. Consider how the program (in the role of the Arab) processes utterance [21: [21 Israeli: The Arabs did, by blockading the Straits of Tiran. To understand this input, the Arab must relate it to what he knows in order to verify its truth, and perhaps more importantly, uncover his relevant knowledge. How does the program verify the claim that the Arab blockade led to the war? Access to historical events in memory depends upon their organization into temporally ordered chains of related events (see Schank (1979)). These chains are searched by a process called causal-temporal (CT) search. CT search contains the system's knowledge about relations between causality and temporal ordering. Such knowledge includes, for example, the rule that any event up to and including the first event of an episodic sequence can be a cause of that sequence, but no subsequent event can. The program checks the plausibility of 121 above by employing CT search backwards in time from the start of the war, looking for the Arab blockade. However, in the course of this search it naturally discovers the initial event of the war episode, the Israeli attack, which is then noted as . a possible cause of the war. Search continues until the Arab blockade is verified. Then, when the time comes to respond to input [21, the Israeli attack has already been found as a possible rebuttal, so that no explicit search need be initiated. Thus a prime method for finding possible responses is simply to make use of facts that have been noticed by prior inferential memory processing. This mechanism also enables the system to deny false presuppositions without the necessity for a special rule. This is because in the course of relating input which contains a false presupposition to long-term memory, the program would discover the falsehood -- and often the reason why it's false. Having this in hand, it can then be used in a rebuttal. Conclusion Engaging in an argument requires several distinct kinds of knowledge: knowledge of the domain, knowledge of how to reason, and knowledge of how to argue. The proper coordination of these disparate knowledge sources can have a mutually beneficial effect in reducing undirected processing. This kind of result would be impossible in a simplistic model based on strictly separate processing stages. References Birnbaum, L., and Selfridge, M. 1980. Conceptual Analysis of Natural Language, in R. Schank and C. Riesbeck, eds., Inside Computer Understanding, Lawrence Erlbaum Associates, Hillsdale, N.J. de Kleer, J., Doyle, J., Steele, G., and Sussman, G. 1977. AMORD: Explicit Control of Reasoning, in Proc. on Artificial Intellipence and ACM SYUIP. Programming Languages,Rochester, N.Y. Doyle, J. 1979. A Truth Maintenance System, Artificial Intelligence, vol. 12, no. 3. McGuire, R. 1980. Political Primaries and Words of Pain, unpublished ms., Yale University, Dept. of Computer Science, New Haven, CT. Schank, R. 1979. Reminding and Memory Organization: An Introduction to MOPS, Research Report no. 170, Yale University, Dept. of. Computer Science, New Haven, CT. 315	1980	59
55	A PROGRAM MODEL AND KNOWLEDGE BASE FOR COMPUTER AIDED PROGRAM SYNTHESIS' Richard J. Wood Department of Computer Science University of Maryland College Park, Maryland 20742 1 Utroductaon Program synthesis is a complex task comprising many interacting subactivities and requiring access to a variety of knowledge sourcesr Recent investigations have discovered the inadequaciest;ef current synthesis techniques to keep pace with increasing difficulties of managing large intricate problem solutions* An alternative approach to software methodologies is the development of intelligent computer systems that mana e the vast amount of information assimilated an % accessed during this process= The system's ttintelli ence" is characterized not by an innate ability to P nvent solutions, but by the incorporation of an internal model of the problem domain and corresponding program solution. This project ex called the uent- domain expert (the programmer (the cons %?%%onsu tant Bf casting this investigation in the paradigm, the. techniques and knowleci$e general to pro rammln can Isolated 5 2 1n y$~~ipy& exam nedo coo erative frameworkanof program s nthesis the fol!?owing four major categories of ac ivities have I been identified: domain from the client, Formulation: transformation, completion, % refinement of the problem requirements into terms recognized by the system= . tructu : the selection of known techniques for solving a task's subproblems and the combination of these solution fragments to form an overall task solution. . !&& P-ludaan : language the instantiationthztf construct schemata corres ond R to ste s of the solution and w ose execu e ion will achieve the overall program behavior, A system capable of supporting computer aided synthesis must have corn onents each of the above active ies, -& corresponding to Additionally, an automated consultant must access a rich knowledge base of programming information in constructing a model of the evolving the-e F This model must %%trnGf the activities syn;t$sls and must be ;;;;;;ed and consultant the he final program, The Interactive Pro ram Synthesizer (IPS) is a system designed to fulfl 1 - f the role of a consultant The research described in this the Office N00014-76C-0471;: Naval re ortu;;e;unded Researc R b gran 8 Their support and encouragement are gratefully acknowledged+ oomplete process descriptions, but- at a cost of additional processing for referent identificationD This renort focuses on the nature of the program model and required for a Specificall successfu!? sy%h%?ed&~~~~ the pro rammi e' the architecture of the Interacti;; Program Syn hesizer, under current development, described, iS zk~ central datamE;gycture of the IPS system developin program which represents the 2 program during synthesis. The organizat on of the program model must accommodate operations that include the introduction of new terms from the user's problem description, the refinement and further definition of existing the detection &E%i! tion of inconsistencies in the and the efficient retraction of the incons !i stent! assertions* These activities occur in client-consultant programming and correspond to the initial problem description by the client, the explanations and clarifications required by the consultant and the rejection of unfruitful partial solutions* The program model is a record of all the assertions, inferences, and deductions made during the synthesis and the justifications for each assertion* The IPS program model is encoded as a semantic network., a data structure which facilitates the processing of synthesis activitiesi The nodes of describe the cE'f'esi;;jE;ce between an ofg;ratiioz and the set requisite achievement. When the sentence "The screen is thought of as a 40 b 86 byte arrayi" is processed by the system two 0 jects are 5 introduced into the model, 1) an object whose name is "screen" and 2) an object 77 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. which is an instantiation of the two dimensional frame ZZXnce (e,g with the information presented in the the a3 the extents ofeach dimension and These two objects are t p~~gdf tSeiaarray entries) 1. DEF link* that reflects the client's decision ?o consider the abstract object "screen', as a two dimensional array* Similarly, the processing of the sentence "To clear the screen store a blank in ever e position of the screenF,, introduces two objezn; corre;;;;;;; operation ,,clear,' a f to the d;;n;E -array operation and links them via a REF link, The REP and RED links are used in a similar manner represent s z stem generated decisions (i.nferenfc!$ as opposed 0 user specificationss The division of link types parallels the distinction between the two domains of expertise of the client and the consultant. Clarifvina reauests to the user are identified by expressed SDEF and REF yi%E " tZrmirio1;~~ ob.iects (e.g "screenI, and "clear,' in the above e&&nnle) ;E;',,e tk;E;ysi;; automatically infers information array store), RED ObJeCtS (e@g., 2-D array and two state 9 When an inimica~t~~;~;action between in the model is unravels Rieger&London, 1977 B the system the cu&ent solution and selects alternative strategies for achievin the RED state before causing a retraction of tha f state* If the offender is a must appeal to the user for a REF state the system restatement of the goal The system cannot jud e the soundness of a user-supplied decom osition e an % must turn to the user for an alterna e decompositionr Other link types exist in the program model $Z$%dency) ~ZerZnce, feature-description, report. beyond the scope of this The reader is directed to other projects that investigate the foundation networks (eg,, 1979115 [Bra&man, 19779 semantic a:df [Fahlman, The program model is instantiating constructed by schematic programming knowledge with problem specific data presented by the user. The programming knowledge base consists of facts and program construction techniques considered primi;;;; to programming and employed during 9r'; descriptions This collection includes: combination to of data types and rules for their form new (2) criteria abstract required t pes; (3) techniques by .a. type descrip ion; I problem for dz;;mposltlon. solving recognition of st$?s ,;zy interactions: and (4) methods for construction of expressions, statements, conditionals, A fundamental input and out ut characteristic of e knowledge base is that the facts are applicable ",E many programming domains, _ _ The IPS kn;;;;zge base is organized in a hierarchical system, efficient or anization ac ivities: 2 of knowledge for iE0 s nthesis recogn;kcn and inference. % eatures presented during user's behavioral ~~crlpt=on task sugg;;; potential programming objects represent abstract domain Identification of a particular programmin ob Jet ts. supplies information normally associated object w th !3 the object., but not stated in the user's discourse, These inferences provide a basis for queries to the user requesting additional selection information, candidatess of a particular object from a set % The programming frames contain information describing defining characteristics and potential roles of an object in a program, the sentence While processing "The screen is thought of as a 40 by 86 byte array", for example, the prototypical t..~ dimensional array frame is retrieved The IPS is designed to communicate with the user in English. translated Curre;;;xr the sentences into corresponding are rn:daZ! manipulation functions. This transformation will ultimately utilize a kevword narser built around a dictionary of programming terminology, instantiated with the data presented the sentence. ",;g$ult Addit;onall , ci in;;;;;kon .a out the array frame imovides some characteristics common?;' used f indicies of the dimensions an d erminology for components of referencing the array (ergs, ,'rowl and "column for the dimensions), It also suggests queries to the user about requisite information (e&g,, is the size fixed? is this a specific individual?) type description or a kz%ted at?) ., are the values cdnt%ied selectional queries in the uses of the'airay Features describing the ~%%ti% (e-.g., the ability to define subregions of an array> are included in the frame but not processed immediately. If later assertions refer to these roles they can be retrieved from the protytpical frame and instantiated, The knowledge base contains frames for both programming objects and operations, Object frames contain defining characteristics and potential roles of an ob'ect, while o by the set of s ates requisi e i t erations are described for their correct execution and the post conditions and side-effects of the action+ This note describes the nature of two corn onents ? of a computer aided program synthesis the internal program %?w?fdge base model and the structures are of programming information. 19801 part of a larger project C%Z" directed towards the development of I theoretical modekf of program synthesis and an implementation incor orates e this Eode P. rogra%Y:ng system ""2 inves igating project simple assembly microprocessor language programming oth,a activities particular and knowledge ?sed by a co&ultant during the construction of a software package that manages a video dis lay program synthesis in t e concrete and R buffer, By examinin uncluttere % realm of-assembl P language prog ams (as contrasted to abstract high- eve1 languages f progress towards a successful corn uter aided programming system can advance in much t e same manner that advances to R general Investigat!~%??%o problem solvi resulted from the blocks-wor d domain. T Trig Thanks to Chuck Rieger, Steve Small, and Randy for reading drafts of this paper and making help ul B commentsI [Br=$mhm;a'n197~lJ Represent!in A Structural Paradigm for Inc,, Repor & 'KAdwledge, Bolt Beranek and Newman, 3605, Many 1978. [Fahlman, 19791 . Fahlman,*S.Ei ,NETL,: A n fer B ?%!bri!!&'!%a%?$@'%! -, M1T YPress' Results in Knowledge Based J?roc& a-I-79, Tokoyo, [Rieger&London, 19771 Rleger, C, & London, Pa, Subgoal Protection and Unravellin . m, f ambridge, Mass., during Plan S~;;~~;~7). pr0c.a [Wood, 19801 ;,oyf 9 R,J, t Computer Aided Pro ram Synthesis, . of Maryland, TR-861, Jan, $980. 78	1980	6
56	Knowledge Representation for Syntactic/Semantic Processing Robert J. Bobrow Bolt Beranek and Newman Inc. 50 Moulton St. Cambridge, Mass. 02238 Bonnie L. Webber Department of Computer Science The Moore School of Elelectrical Engineering D2 University of Pennsylvania Philadelphia, Pa. 19104 1. Introduction This paper discusses some theoretical implications of recasting parsing and semantic interpretation as a type of inference process which we call incrementalhescription refinement.* It draws upon our recent experience with RUS. a framework for natural language processing developed at BBN and in use in several different natural language systems across the country (for details see [II, C21 and C71>. RUS is a very practical system that is as efficient as a semantic grammar like the SOPHIE parser 161 and as flexible and extensible processor lik? LUiAR modular syntactic/semantic CIOI. It achieves this ccmbination of efficiency and flexibility by cascading Cl21 syntactic and semantic processing- producing the semantic interpretation of an input utterance incremental1 y during the parsing process, and using it to guide the operation of the parser. The research reported in this paw was supported in part by the .Advanced Research Proj ects Agency, and was monitored by ONR under Contract No. NOOO14-77-C-0378. Because RUS provides a very clean interface between syntactic and semantic processing, it has been possible to experiment with a variety of knowledge representations in the different implementations noted above. The most recent such implementation uses the KL-ONE formalism c31, [41, to represent the knowledge needed for incremental processing. (This implementation has been dubbed PSI-KLONE, for "Earsing and Semantic InterpretatioGsmL-ONE1l.) KL-ONE is a uniform Ebject-centered representational scheme based on the idea of structured inheritance in a lattice-structured taxonomy of generic knowledge. As we shall discuss iater, - PSI-KLONE takes advantage of KL-ONE's taxonomic lattice [ill which ccmbines the properties of an inheritance network with those of-a discrimination net. The next section of this paper describes the syntactic/semantic cascade in general terms, and then gives a short example of its operation in the PSI-KLONE implementation. We then define the concent of an incremental describtion refinement (IDR)' process to use as a paradigm for usrstanding the operation of the semantic component of-the cascade. section of the paw, requirements for a general This introduces the last which discusses the frame-like knowledge representation if it is to be capable of supporting such an IDR process. 2. The Syntactic/Semantic Cascade Within the RUS framework, the interaction between the. parser and the semantic interpreter (the interpreter) takes place incrementally as the parser scans the input string from left to right, one word at a time. The semantic interpretation of each syntactic constituent is produced in parallel with the determination of its syntactic structure. Knowledge developed in the course of producing the interpretation is used to control further action by the parser. Parsing supports the processes of semantic interpretation and discourse inference (not discussed in this paper) by finding the constituents of each phrase, determining their syntactic stS;ucture, and labelling their functional relationship to the phrase as a whole (the * We use an extended notion of functional relation here that includes surface syntactic relations, logical syntactic (or shallow case structure) relations, and relations useful for determining discourse structures such as primary focus. 316 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. matrix). These labels are proposed purely on the basisof syntactimrmation, but are intended to reflect a constituent's functional role in the matrix, apd not simply its internal syntactic structure. We will refer to these labels as functional or syntactic labels for constituents. The parser and interpreter engage in a dialogue consisting of transmissions from the parser and responses from the interpreter. A transmission is a proposal by syntax that some snecific functional relation holds between a previously parsed and interpreted constituent and the matrix phrase whose parsing and interpretation is in progress. The proposal takes the form of a matrix/label/constituent triple. The interpreter either rejects the proposal or accepts it and returns a pointer to a KL-ONE data-re which represents-its knowledge of the resulting phrase. (This pointer is not analyzed by the parser, but is rather used in the description of the matrix that syntax includes in its next proposal (transmission) to extend the matrix.) The parser is implemented as an ATN 191, and transmissions occur as actions on the arcs of the ATN grammar. The failure of an arc because of a semantic rejection of a transmission is treated exactly like the failure of an arc because of a syntactic mismatch; alternative arcs on the source state are attempted, and if none are successful, a back-up occurs. 2.1. The role of the semantic interpreter in a cascaded system The PSI-KLONE interpreter must perform two related tasks: 1. provide. feedback to the parser by checking the semantic plausibility of syntactic labels for proposed constituents of a phrase, and 2. build semantic individual phrases interpretations for The mechanism for performing both these tasks is based on the idea of mapping between the (syntactic) functional labels provided by the parser and a set of extended case-frame or semantic relations (defined by the inters) that can hold between a constituent and its matrix phrase. The mapping of functional labels to semantic relations is clearly one to many. For example, the logical subject of a clause whose main verb is "hiV1 might be the agent of the act (e.g. "'Ihe boy hit . ..I'> or the instrument (e.g. "The brick hit 11 . . . 1. A semantic relation (or semantic role), on the other hand, must completely specify the role played by the interpretation of- the constituent the interpretation of the matrix phrase. in example, a noun phrase (NP) can serve various functions in a clause, including logical subject (LSUBJ), logical object (LOBJ), surface subject (SSUBJ), and first NP (FIRSTNP). The task of the interpreter is to determine which, if any, semantic relation could hold between a matrix phrase and a parsed and interpreted constituent, given a functional label proposed by the parser. This task is accomplished with the aid of a- set of pattern-action relation mapping rules (RMRULES) that specify how a given fun- label cmapped into a semantic relation. An RMRULE has a pattern (a matrix/label/constituent triple) that specifies thext=in which it applies, in terms of: 0 the syntactic shape of the matrix (e.g. "It is a transitive clause whose main verb is 'run'."), and the interpretation and semantic role assigned to other constituents (e.8. 'IThe logical subject must be a person and be the Agent of the clause"), o the proposed functional label, and o the interpretation of the be added. constituent to The action of the RMRULE is to map the given functionallabel onto a semantic relation. A proposed syntactic label is semantically plausible if its proposal triple matches the pattern triple(s) of some RMRULE(s). KL-ONE is a good language for describing structured objects such as phrases built up out of constituents, and for representing classes of objects such as the matrix/label/constituent triples that satisfy the constraints given by RMRULE patterns. In PSI-KLONE, each RMRULE pattern is represented as a KL-ONE structure called a Generic Concept (see section 2.2). These Concep55??5- am in a taxonomy that is used as a discrimination net to determine the set of patterns which match each triple. We refer to this as the taxonomy of syntactic/semantic shapes; note that it is generally a lattice and not simply a tree structure. Associated with each semantic relation is a rule (an IRULE) that specifies how the interpretation of the constituent is to be used in building the interpretation of the matrix phrase. When all the constituents of a matrix have been assigned appropriate semantic relations, the interpretation of a phrase is built up by executing all of the IRULEs that apply to the phrase. The separation of RMRULEs from IRULEs allows PSI-KLONE to take full advantage of the properties of the syntactic/semantic cascade. As each new constituent is proposed by the parser, the interpreter uses the RMRULEs to determine which IRULEs apply to that constituent; but it does not actually apply them until the parser indicates that all constituents have been found. This buys That is, the constituent interpretable as a person. labelled LSUBJ must be 317 efficiency by rejecting constituent label assigrments which have no hope of semantic of the PHRASE, and may have 0 or more other SYNTACTIC-CONSTITUENTS which are Modifiers. The interpretation, while deferring the construction of double arrow or SuperC Cable between PHRASE and an interpretation until the syntactic well-formedness of the entire phrase is verified. SYNTACTIC-CONSTITUi?~indicates that every instance of PHRASE is thereby a SYNTACTIC-CONSTITUENT. I 2.2. An example of the cascade The simnlified taxonomv for our example is As a simplified example of the parser-interpreter interaction, and the use of the KL-ONE taxonomy of syntactic/semantic shapes in this interaction, we will briefly describe the process of parsing the clause "John ran the drill press." The simplified ATN grammar we use for this example is shown in Fig. 2-1. given in Fig: 2-3. This indicates that any CLAUSE whose Head is the verb "run" Figure 2-l: A simplified ATN For readers unfamiliar with KL-ONE, we will explain three of its major constructs as we note the information represented in the simple taxonomy shown in Fig. 2-2. In E-ONE, Generic Concepts (ovals in the diagram, boldfaceinhm represent Figure 2-2: A simple EL-ONE network description templates, from which individual descriptions or Individual Concepts (shaded ovals, also boldface in text) are formed. In Fig. 2-2, the most general description SYNTACTIC-CONSTITUENT, which is specialized by tiz two descriptions, PHRASE and WORD. All KL-ONE descriptions are structured objects. The only structuring device of concern here is the Role. A Role (drawn as a small square, and underEd in text) represents a type of relationship between two objects, such as the relation between a whole and one of its parts. Every Role on a Generic Concept indicates what type of object can fill the Role, and how many distinct instances of the relation represented by the Role can occur. The restriction on fillers of a Role is given by a pointer to a Generic Concept, and the ntanber of possible instances of the Role is shown by a number facet (indicated in the form "M < # < NH in thegs In our diagram we indicate that every PHRASE has a WORD associated with it which fills the Head Role Figure 2-3: A simple EL-ONE Syntactic Taxonomy (independent of tense and person/nunber agreement) is an example of a RunCLAUSE. There are two classes of RunCLAUSEs represented in the taxonmy - those whose LSUBJ is a person (the Per sonRunCI.AUSEs), and those whose LSUBJ is a machine (the MachineRunCLAUSEs). The class of PersonRunCLAUSEs is again sub-divided, and its subclasses are RunMachineCLAUSE (in which the LOBJ must be a machine), RunRaceCLAUSE (in which the LOBJ is a race), and SimpleRunCLAUSE (which has no LOBJ). If we get an active sentence like "John ran the drill press", the first stage in the parsing is to PUSH for an NP from the CLAUSE network. For simplicity we assume that the result of this is to parse the noun phrase "John" and produce a pointer to NPl, an Individual Concept which is an instance of the Generic pattern PersonNP. This is the result of interaction of the parser and interpreter To reduce clutter, several superC cables to the Concept NP have been left out. 318 at a lower level of the ATN. Since it is not yet clear what role NPl plays in the clause (i.e. because the clause may be active or passive), the parser must hold onto NPl until it has analyzed the verb. Thus the first transmission from the parser to the interpreter at this level is the proposal that rrrunV (the root of "ran") is the Head of a CLAUSE. The interpreter accepts this and returns a pointer to a new Individual Concept Ckl which it places as an instance of RunCLAUSE. Since the parser has by now determined that the clause is a simple active clause, it can now transmit the proposal that NPl is the LSUBJ of CLl. Because NPl is an instance of a PersonNP, the interpreter can tell that it satisfies the restrictions on the LSUBJ of one of the specializations of RunCLAUSE, and thus it is a semantically plausible assignment. The interpreter fills in the LSUBJ Role of CL1 with NPl, and connects CL1 to PersonRunCLAUSE, since that is the only subConcept of RunCLAUSE which can have a PersonNP as its LSUBJ. Finally, the parser PUSHes for an NP, resulting in a pointer to NP2, an instance of MachineNP. This is transmitted to the interpreter as the LOBJ of CLI. Since CL1 is a PersonRunCLAUSE, the taxonomy indicates that it can be either an instance of a RunRaceCLAUSE or a RunMBCLAUSE, or a SimpleRunCLAUSE. Since-P2 has been classifieras an instance of MachineNP, it is not compatible with being the LOBJ of a RunRaceCLAUSE (whose LOBJ must be interpretable as a race). On the other'hand NP2 is compatible with the restriction on the filler of the LOBJ Role of RunMachineCLAUSE. We assume that the taxonomy indicates all the acceptable subcategories of PersonRunCLAUSE~Thus it is only semantically plausible for NP2 to fill the LOBJ Role of CL1 if CL1 is an instance of RunMachineCLAUSE. This being the case, the interpreter can join CL1 to RunMachineCLAUSE and fill its LOBJ Role with NP2, creating a new version of CL1 whmit returns to the parser. At this point, since there are no more words in the string, the parser transmits a special message to the interpreter, indicating that there are no more constituents to be added to CLI. The interpreter responds by finding the IRULEs inherited by CL1 from RunMachineCLAUSE, PersonRunCLAUSE, etc. and using the actions on those IRULEs to create the interpretation of CLl. It associates that interpretation with CL1 and returns a pointer to CLl, now a fully parsed and interpreted clause, to the parser. Actually, the interpreter creates a Generic subConcept of RunCLAUSE, in order to facilitate sharing of information between alternative paths in the parse, but we will ignore this detail in the remainder of the example. 3. Incremental Description We view the cascaded Refinement analysis and semantic interaction of svntactic interpretation as implementing a recognition paradigm we refer to as intiremental description refinement. In this paradigm we assume we are initially given a domain of structured objects, a space of descriptions, and rules that determine whic$ descriptions apply to each object in the domain. As an example, consider the domain to be strings of words, the structured descriptions to be the parse trees of some grammar, and say that a parse tree applies to a string of words if the leaves in the tree correspond to the sequence of words in the string. In general we assume each description is structured, not only describing the object as a whole, but having ccmponents that describe the parts of the and their relationship to the whole as well. object We consider a situation that corresponds to left-to-right parsing. A machine is presented with facts about an object or its parts in some specified order, such as learning the words in a string one by one in a left-to-right order. As it learns more properties of the object the machine must determine which descriptions are compatible with its current knowledge of the properties object and its parts. the process of: Incremental description refinement o determining the set of the (IDR) is of descriptions ccmpatible with an object known to have a given set of properties, and o refining the set of descriptions as more properties are learned. More precisely, for every set of properties P = {pl ,...,pn) of some object 0 or its parts, there is an associated set of descriptions C(P), the descriptive cover of P. The descriptive cover of P consists of- those descriptions which might possibly be applicable to 0, given that 0 has the properties PI ,-*I b; that is, the set of descriptions which apply to at least one object which has all the properties in P. As one learns more about some object, the set of descriptions consistent with that knowledge shrinks. Hence, the basic step of any IDR process is to take (1) a set of properties P, and (2) its cover C(P), and (3) some extension of P into a set P', and to produce C(P') by removing inapplicable elements from C(P). The difficulty is that it is usually impractical, if not impossible, to represent C(P) extensionally: in many cases C(P) will be infinite. (For example, until the number of words in a string is learned, the nLPnber of parse trees in C(P) remains infinite no matter how many words in the string are known.) Thus, the We assume that at least one to each object in the domain. description applies 319 covering set must be represented intensionally, with the consequence that "removing elements" becomes an inference process which determines the intensional representation of C(P1) given the intensional representation of C(P). Note that just as any element of C(P), represented extensionally, may be structured, so may the intensional representation of C(P) be structured as well. The trick in designing an efficient and effective IDR process is to choose a synergistic inference process/intensional representation pair. One example is the use of a discrimination tree. In such a tree each terminal node represents an individual description, and each non-terminal node represents the set of descriptions corresponding to the terminals below it. Every branch indicates a test or discrimination based on some property (or properties) of the object to be described. Each newly learned property of an object allows the IDR process to take a single step down the tree, as long as the properties are learned in an order ccmpatible with the tree's structure. Each step thus reduces the set of descriptions subsumed. Another IDR process is the operation of a constraint propagation system. In such a system an object is described by a set of nodes, each of which bears a label chosen from some fixed set. The nodes are linkedinto a network, and there is a constraint relation that specifies which pairs of labels can occur on adjoining (i.e. linked) nodes. The facts learned about an object are either links between previously known nodes, or label sets which specify the possible labels at a single node. A descriptive cover is simply the cross-product of some collection of node label sets. The refinement operation consists of (I) extending the analysis to a new node, (2) removing all incompatible labels from adjacent nodes and (3) propagating the effects. Unlike the use of a discrimination net, constraint propagation does not require that information about nodes be considered in some a priori fixed order. - As mentioned earlier, in the RUS framework we are attempting to refine the semantic description of an utterance in parallel with determining its syntactic structure. The relevant properties for this IDR process include the descriptions of various constituents and their functional relations to their matrix (cf. Section 2). Unfortunately, surface variations such as passive forms and dative movement make it difficult to assume any particular order of discovery of properties as the parser considers words in a left to right order. However, the taxonomic lattice of KL-ONE can be used as a generalization of a discrimination tree which is order independent. The actual operation used in PSI-KLONE involves an extended notion of constraint propagation operating on nodes in the taxonomic lattice, and thus the resulting system has interesting analogies to both simpler forms of IDR processes. The complete algorithm for the IDR process in PSI-KLONE is too ccmplex to cover in this paper, and will be described in more detail in a forthcoming report. However, the reader is urged to return to the example in Sec. 2.2 and reconsider 320 it as an IDR process as we have described above. Briefly, we can view the KL-ONE taxonomy of syntactic/semantic shapes as a set of discrimination trees, each with a root labelled by some syntactic phrase type. At each level in a tree the branches make discriminations based on the properties of some single labelled constituent (El, such as the LSUBJ of a CLAUSE. The parser first proposes a phrase type such as CLAUSE, and the IDR process determines which tree has a root with that label. That root beccmes the current active node in the IDR process. All further refim isone within the subtree dominated by an active node. As the parser proposes and transmits new LC's to the IDR, the IDR may respond in one of two ways: 1. it may reject the LC because it is not compatiblewith any branch below the currently active node(s), or 2. it may accept the LC, and replace the current -58376 node(s) with the (set of) node(s) which can be reached by branches whose discriminations are compatible with the LC. 4. The IDR Process and Knowledge Representation We-have- identified four critical characteristics of any general representation scheme that can support an IDR process in which descriptions are structured and covering descriptions are represented intensionally. In such a scheme it must be possible to efficiently infer from the representation: 1. what properties of a structured object provide sufficient information to guarantee the applicability of a description to (some portion of) that object - i.e., criteriality conditions, 2. what mappings are possible between classes of relations - e.g. how functional relationships between syntactic constituents map onto semantic relationships 3. which pairs of descriptions are mutually incompatible - i.e., cannot both apply to a single individual 4. which sub-categorizations of descriptions are exhaustive - i.e., at least one of the sub-categories applies to anything to which the more general description applies. Wr analysis of the assumptions implicit in the current implementation of PSI-KLONE has led us to an understanding of the importance of these four points in a IDR. By making these four points explicit in the next implementation we expect to be able to deal with a larger class of phenomena than the current system handles. In the following sections we illustrate these four points in terms of the behavior of the current version of PSI-KLONE and the improvements we expect to be able to make with more information explicit of types of 4.1. Criteriality Conditions The point here is an obvious one, but bears repeating. If a taxonomy is to be used for recognition, then there must be some way, based on partial evidence, to get into it at the right place for the recognition (IDR) process to begin. That is, for any ultimately recognizable phrase there must be at least one criteria1 condition, i.e. a collection of facts which is sufficient to ensure the abnlicabilitv of some particular descris In the' syntactic"/semantic taxonomy, the criteria1 condition is often, for a phrase, the properties of belonging to a particular syntactic category (e.g., noun phrase, clause, etc.) and having a particular lexical item as head. Recalling the example given in Section 2.2, the evidence that the input had the shape of a CLAUSE and had the verb llrunV as its head constituted sufficient conditions to enter the taxonomy at the node RunCLAUSE - i.e., a specialization of CLAUSE whose head is filled by the verb "run". Without the notion of criteria1 properties, we cannot ensure the applicability of any description and therefore have no way of continuing the recognition process. 4.2. Mapping Syntactic to Semantic Relations In RB, the parser intermittently sends messages to the interpreter asking whether it is semantically plausible for a constituent to fill a specified functional role. The interpreter's ability to answer this question ccmes from its RMRULEs and their organization. This is based on the assumption that a potential constituent can fill some functional role in the matrix phrase if and only if it also fills a semantic role compatible with: o that functional role o the interpretation of that constituent o the head of that matrix phrase o the roles filled of that phrase by the other constituents o other syntactic/semantic propert that phrase and its constituents. ies of With respect to the first of these points, one effective way of representing the compatibility restrictions between syntactic and semantic relations derives from the fact that each purely syntactic relation can be viewed as an abstraction of the syntactic properties shared by some class of semantic relations (i.e., that they have syntactically identical argunents). If 1. a general frame-like system is used to represent the system's syntactic/semantic knowledge, 2. possible syntactic and semantic relations are represented therein as "slots" in a frame, and 3. there is an abstraction hierarchy among slots (the Role hierarchy in KL-ONE), as well as the more common IS-A hierarchy among frames (the SUPERC link between concepts in KL-ONE), then the interpreter can make use of this abstraction hierarchy in answering questions from the parser. As an example, consider a question from the parser, loosely translatable as "Can the PP 'on Sunday' be a PP-modifier of the NP being built whose head is 'party'?". Fig. 4-1jo Figure 4-l: A Simple NP Taxonomy We assume the NP headed by "party" has already been classified by the interpreter as a CelebrationNP. As indicated in Fig. 4-l this concept inherits from EventNP two specializations of the general PP-modifier relation applicable to NP - location-PP-modifier and time-PP-modifier. Thus "on SundaT can be one of% p-modifiers iff it can be either its location-PFmodifier or its time-PP-modifier. The next sEtion will discuss howt& decision can be made. The point here is that there must be some indication of which syntactic relations can map onto which semantic ones, and under what circumstances. An abstraction hierarchy among Roles provides one method of doing so. 4.3. Explicit compatibility/incompatibility annotations As noted above, the semantic interpreter must be able to decide if the interpretation assigned to the already parsed constituent is compatible with the type restrictions on the argLPnents of a semantic relation. For example, the PP "on Sunday" can be a PP-modifier of an NP whose Head is "party" if it isccmpatible with being either a time-PP, and hence capable of instantiating the relation time-PP-modifier, or a location-PP and hencs instaziating the relation location-PP-modifier. - izing There are two plausible strateg the somewhat informal notion of ies for formal compatibility: In this, as in Fig. 4-1, we assume for simplicity that only these two semantic relations are consistent with the syntactic relation PP-modifier for an NP whose head is "party". - 321 1. a constituent is judged ccmpatible with a restriction if its syntactic/semantic shape (and hence interpretation) guarantees consistency with the type restrictions, or 2. it is 3-&W compatible if its interpretation does not guarantee inconsistency. - Consider the problem of rejecting "on Sunday" as a location-PP-modifier. Conceivably one could reject it on thegrounds that "Sunday" doesn't have a syntactic/semantic shape that guarantees that it is a location-NP. This is essentially the strategy followed by the current version of PSI-KLONE. More specifically, the PSI-KLONE system searches along the superC cables of a constituent to find just those semantic relations which are guaranteed to be canpatible with the interpretation of the constituent and matrix. However, "birthday that strategy would have to reject present" as being compatible with apparel-NP (thereby rejecting "Mary wore her birthday present to New York"), vehicle-NP (thereby rejecting "Mary drove her Boston to birthday present from Philadelphia"), rejecting animate-NP (thereby "Mary fed her birthday present some Little Friskies"), etc. Thus, we believe that future systems should incorporate the second strategy, at least as a fall-back when no interpretation is found using only the first strategy. This strategy also makes it easier for the system to handle pronouns and other semantically empty NPs (e.g. etc.) whose "thing" "stuff" syntactic/semantic shapes' guarantei almost nothing, but which are compatible with many semantic interpretations. The imp1 ication here for both processing language and knowledge repr mesentation is that: 1. incompatibility must be marked in the representation, and explicitly 2. the most useful strategy for determining compatibility involves not being able to show explicit inccmpatibility. One caveat and one further observ ation: this strategy is not by itself effective in certain cases of metonymy, which Webster's defines as "the use of the name of one thing for that of another associated with or suggested-by it." semantics would reject For example, "the hamburger" as the subject of,a clause like "the hamburger is getting impatient" which might occur in a conversation between a waiter and a short-order cook. However, the taxonomy would be able to provide information If we assume something like "hamburger" being an instance of definite-food-NP, which is marked as incompatible with animate-NP, the restriction on the subject of "impatient". needed to resolve the metonymy, since it would indicate that "the hamburger" is possibly being used metonymously to refer to some discourse entity which is both describable by an animate-NP and associated with some (unique) hamburger. The observation concerns the way in which semantic interpretation was done in LUNAR [lOI, which was to judge semantic compatibility solely on the basis of positive syntactic/semantic evidence. A semantic interpretation rule could only be applied to a fragment of a parse tree if the rule's left-hand side - a syntactic/semantic template - could be matched against the fragment. The only kinds of semantic constraints actually used in the LUNAR templates were predicates on the head of some tree constituent -- e.g. that the head of the NP object of a PP constituent were of class element, etc. rock, Given this restriction, LUNAR w-t be able to handle an utterance like "give me analyses of alLaninun in NASA's gift to the Royal Academy", where clearly "gift to the Royal Academy" is not incompatible with rock. 4.4. Explicit marking of exhaustive sub-categorization in the taxonomy The algorithm we have developed for incremental description refinement requires that the IDR process be able to distinguish exhaustive from non-exhaustive sub-categorization in the taxonomy of syntactic/semantic shapes. Exhaustiveness marking plays a role similar to that played by inclusive or in a logical framework. That is, it justifies the application of case-analysis techniques to the problem of determining if a proposed constituent is ccmpatible with a given syntactic role. The interpreter is justified in rejecting a proposed label for a constituent only if it has considered all possible ways in which it can correspond to a semantic relation. Exhaustiveness marking also make it possible to infer positive information from negative information as was done in the example in section 2.2. There, the interpreter inferred that the clause was a RunMachineCLAUSE, because it was known to be a PersonRunCLAUSE and the proposed LOBJ was incompatible with it being a RunRaceCLAUSE. Such reasoning is justified only if the subcategories RunMachineCLAUSE, SimpleRunCLAUSE and RunRaceCLAUSE exhaust the possibilities under PersonRunCLAUSE. These types of inference do not always ccme up in systems that are primarily used to reason about inherited properties and defaults. For example, as long as one knows that DOG and CAT are both specializations of PET, one knows what properties they inherit from PET. It is irrelevant to an inheritance process whether there are any other kinds of PET such as MONKEY, BOA-CONSTRICTOR or TARANTULA. Many formalisms, including KL-ONE, do not require the sub-categorization of a node to be exhaustive. So there are two options vis-a-vis the way exhaustiveness can be indicated. A recognition algorithm can act as if every node were exhaustively sub-categorized - a type of Closed World Assumption L-81 - this is essentially the way 322 the current PSI-KLONE system operates. Unfortunately, there are other uses of KL-ONE in the natural language system in which concepts are subcategorized but it is clear that an exhaustive subcategorization has not been made. If the meaning of the links in the representation scheme is to be well-defined, it must be possible to distinguish exhaustive from non-exhaustive sub-categorization. The implication for both knowledge representation and inference is that some clear stand must be taken vis-a-vis the representation of exhaustive sub-categorizations. 5. Conclusion The approach we have taken in RUS is midway between completely decoupled syntactic and semantic processing and the totally merged processing that is characteristic of semantic grammars. RUS has already proven the robustness of this approach in several different systems, each using different knowledge representation techniques for the semantic ccmponent. The RUS grammar is a substantial and general grammar for English, more extensive than the grammar in the LUNAR system ClOl. Although the grammar is represented as an ATN, we have been able to greatly reduce the backtracking that normally occurs in the operation of an ATN parser, allowing RUS to approach the performance of a "deterministic" parser [21. With the aid of a "grammar compiler" [51 this makes it possible to achieve parsing times on the order of .X CPU seconds, on a DEC KLIO, for twenty word sentences. In this paper we have- focused on the latest embodiment of the RUS framework in the PSI-KLONE system -- in particular on the nature of its cascaded syntactic/semantic interactions and the incremental description refinement process they support. We believe that what we are learning about such cascaded structures and IDR processes in building PSI-KLONE is of value for the design of both natural language systems and knowledge representation systems. ACKNOWLEDGEMENTS Cur work on the PSI-KLONE system has not been done in vacua - our colleagues in this research effort includeEd Barton, Madeleine Bates, Ron Bra&-man, Phil Cohen, David Israel, Hector Levesque, Candy Sidner and Bill Woods. We hope this paper reflects well on our joint effort. The authors wish to thank Madeleine Bates, Danny Bobrow, Ron Bra&man, David Israel, Candy Sidner, Brian Smith, and Dave Waltz for their helpful comments on earlier versions of this paper. Our special thanks go to Susan Chase, whose gourmet feasts and general support made the preparation of this paper much more enjoyable than it might otherwise have been. [II PI L-31 [41 [51 Ii61 c71 [81 191 Cl01 Cl11 r121 Woods, W. A. Cascaded ATN Grammars. Amer. J. Computational Linguistics 6(l), Jan:-Mar., 1980. REFERENCES Bobrow, R. J. The RUS System. -- BBN Repom8, Bolt Beranek and Newnan Inc,, 1978. . Bobrow, R. J. & Webber, B. L. PSI-KLONE - Parsing and Semantic Interpretation in the BBN Natural Language Understanding System. In CSCSI/CSEIO Annual Conference. 7z?xrim,*1980. Brachnan, R. J. On the Epistemological Status of Semantic Networks. In Findler, Nicholas V. (editor), Associative Networks - The Representation and Use of Knowledge inomputers, . AcaGi'Gs, New York, 379. Bract-man, R. J. An Introduction to KL-ONE. In Bra&man, R.J., et al. (editors), Research in Natural Language Understanding, Annual Rep- 31 Aug. 79), pages -- -- 13;116, Bolt Berzek and Nekanan Inc, Cambridge, MA, 1980. Burton, R. & Woods, W. A. A Compiling System for Augmented Transition Networks. In COLING 76. Sixth International Conference ?%7Ziputational Linguistics, Ottawa, Canada, June, 1976. Burton, R.. Seelv Br0wn.J. Semantic Grammar: A Technique for ConstrwNa?%ral LanguazInterfaces to Instructional Svstems. BBNTeport 3587, E!ol-?&%%k And Newnan Inc., May, 1977. Mark, W. S. & Barton, G. E. The RUSGrammar Parsing System. m 3243, Genermrmarch Laboratories, 1980. Reiter, R. Closed World Data Bases. In Gallaire, H. & Minker, J. (editor), Logic and Data Bases, . -- Plenum Press, 197T Woods, W. A. Transition Network Grammars for Natural Language Analysis. CACM 13(10), October, 1970. Woods, W. A., Kaplan, R. M. & Nash-Webber. B. The Lunar Sciences Natural Language ' -- Information Systmnal Report. BBN Report 2378, Bolt Beranek andwnan Inc., June, 1972. Woods, W. A. Taxonomic Lattice Structures for Situation Recognition. In Theoretical Issues in Natural Language Processing -r- 1978. - ACL amSIGART, July, 323	1980	60
57	LANGUAGE AND MEMORY: GENERALIZATION AS A PART OF UNDERSTANDING Michael Lebowitz Department of Computer Science Yale University, P.O. Box 2158 New Haven, Connecticut 06520 ABSTRACT process. Here I will present the program's ability to remember Integrated Partial events and make generalizations about This paper presents the the stories it reads. Parser (IPP), a computer model that combines text understanding and memory of events. An extended example of the program's ability to understand MEMORY IN UNDERSTANDING newspaper stories and make generalizations that are useful for memory organization is presented. The addition of information to long-term memory is an integral part of IPP's operation. actively INTRODUCTION Furthermore, the memory update process changes the structure of memory by noticing similarities new memory Memory of specific events has not been a among events, creating structures serious part of previous Artificial Intelligence based upon generalizations about these language understanding systems. Most understanding events, and using these new structures in storing events. systems have been designed to simply parse natural language into an internal representation, use the I will representation for output tasks, and then discard In order to illustrate IPP's memory, present three stories taken This has been true even for systems primarily directly from it. concerned with creating high-level and show how IPP incorporates them into story newspapers, The stories all describe kidnappings that representations in terms of structures such as memory. scripts 1121 [31 [6l, plans and goals [6l [71, took place in Italy. frames [II and other such structures, as well as more syntactically oriented systems. (Sl) Boston Globe, 5 February 79, Italy Three gunmen kidnapped a 67 year-old retired The Integrated Partial Parser (IPP) is an industrialist yesterday outside his house near understanding program that addresses the problems this north Italian town, police said. of integrating parsing and memory updating. By making use of memory, IPP is able to achieve a high (S2) New York Times, 15 April 79, Italy A building contractor kidnapped here on Jan. level of performance when understanding texts that 17 was released last night after payment of an it has not been specially prepared for. IPP has been designed to read and remember news stories undisclosed ransom, the police said. about international terrorism taken directly from newspapers and the UPI news wire. (S3) New York Times, 25 June 79, Italy Kidnappers released an Italian shoe manufacturer here today after payment of an IPP performs six different understanding undisclosed ransom, the police said. tasks, in a single, integrated process. These tasks include the addition of new events to After a person has read these three stories he long-term memory, being reminded of previous would have undoubtedly drawn some conclusions about stories, noticing the interesting aspects of kidnapping in Italy. The similar nature of the stories, making generalizations to improve the victims - all businessmen of one sort or another - quality of its world knowledge, predicting likely In what follows I will future events, and, of course, parsing stories from is immediately apparent. show the natural language into an internal representation. actual generalizations IPP made upon reading this sequence of stories. The output that follows In this paper, I will mention IPP's parsing is from IPP's operation on the three stories above. abilities only in passing. The interested reader is refered to [41 for more details of the parsing Story: Sl (2 5 79) ITALY -------------------- (THREE GUNMEN KIDNAPPED A 67 YEAR-OLD RETIRED INDUSTRIALIST YESTERDAY OUTSIDE HIS HOUSE NEAR THIS This work was supported in part by the Advanced NORTH ITALIAN TOWN POLICE SAID) Research Projects Agency of the Department of Defense and monitored under the Office of Naval * Parsing and incremental memory access * Research under contract N00014-75-C-1111. 324 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. >>> Beginning final memory incorporation . . . Story analysis: EVl (S-MOP = S-EXTORT) HOSTAGES AGE OLD OCCUPATION-TYPE RETIRED ROLE BUSINESSMAN POL-POS ESTAB GENDER MALE ACTOR NUMBER 3 METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Indexing EVl as variant of S-EXTORT >>> Memory incorporation complete Stories are represented in IPP by MOPS (Memory Organization Packets) [51. MOPS were designed to represent events in memory, and are thus particularly well suited for maintaining a permanent the purpose of episodic memory. Sl is represented by S-EXTORT, a simple MOP (S-MOP) that is intended to capture the common aspects of extortion and SKIDNAP, a script that represents the low-level events of a kidnapping. IPP's long term memory is organized around S-MOPS, structures that represent generalities among scripts, and the ways in which different stories vary from the S-MOPS. The first step in adding a new story to memory is to extract the various features of each instantiated S-MOP. These features consist of the scripts that are instantiated subordinate to the S-MOP (methods and results, typically), and features from each of the role fillers (such as the actor and victim). The output above shows the features that IPP accumulated about the S-EXTORT S-MOP as it read Sl. Since there were no similar events in memory when IPP read this story, its addition to memory consisted of indexing it under S-EXTORT using simply each of the features as an index. Now consider they way IPP reads S2. The collection of the features of the S-EXTORT occurs during parsing of the story. Story: S2 (4 15 79) ITALY (A BUILDING CONTRACTOR KIDNAPPED HERE ON JAN 17 WAS RELEASED LAST NIGHT AFTER PAYMENT OF AN UNDISCLOSED RANSOM THE POLICE SAID) * Parsing and incremental memory access * >>> Beginning final memory incorporation . . . Story analysis: EV3 (S-MOP = S-EXTORT) RESULTS SCRIPT SS-GET-RANSOM SCRIPT SS-RELEASE-HOSTAGES HOSTAGES ROLE BUSINESSMAN POL-POS ESTAB GENDER MALE METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Creating more specific S-EXTORT (SpMl) from events EV3 EVl with features: HOSTAGES ROLE BUSINESSMAN POL-POS ESTAB GENDER MALE METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Reminded of: EVl (during spec-MOP [EVl is from Sll >>> Memory incorporation complete creation) While reading S2, IPP adds it to long-term memory. However, as it uses the indexing procedure described above, it naturally finds Sl, which is already indexed under S-EXTORT in many of the same ways as S2 would be. Thus IPP is reminded of Sl, and what is more, since there are such a large number of features in common, IPP makes a tentative generalization that these features normally occur together. Informally speaking, IPP has concluded that the targets of kidnappings in Italy are frequently businessmen. The new generalization causes the creation of a new node in memory, known as a spec-MOP, to be used to remember events that are examples of the generalization. A spec-MOP is simply a new MOP, equivalent in kind to the S-MOPS that IPP starts with, that is used to embody the generalizations that have been made. The events that went into the making of the generalization are then indexed off of the spec-MOP by any additional features they may have. Story: S3 (6 25 79) ITALY (KIDNAPPERS RELEASED AN ITALIAN SHOE MANUFACTURER HERE TODAY AFTER PAYMENT OF AN UNDISCLOSED RANSOM THE POLICE SAID) Processing: KIDNAPPERS : Interesting token - KIDNAPPERS Instantiated SKIDNAP -- S-EXTORT 325 >>> Beginning memory update . . . New features: EV5 (S-EXTORT) METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Best existing S-MOP(s) -- SpMl [the spec-MOP just created] Predicted features (from SpMl) HOSTAGES GENDER MALE POL-POS ESTAB ROLE BUSINESSMAN >>> Memory update complete . . [rest of the parsing process] >>> Beginning final memory incorporation . . . Story analysis: EV5 (S-MOP = S-EXTORT) RESULTS SCRIPT SS-GET-RANSOM HOSTAGES NATIONALITY ITALY ROLE BUSINESSMAN POL-POS ESTAB GENDER MALE RESULTS SCRIPT SS-RELEASE-HOSTAGES METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Creating more specific S-EXTORT (SpM2) than SpMl from events EV5 EV3 with features: RESULTS SCRIPT SS-GET-RANSOM SCRIPT SS-RELEASE-HOSTAGES HOSTAGES ROLE BUSINESSMAN POL-POS ESTAB GENDER MALE METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Reminded of: EV3 (during spec-MOP creation) [EV3 is from S21 >>> Memory incorporation complete As it finishes reading S3, IPP completes its addition of the new event to memory. The processing is basically the same as that we saw for s2, but instead of beginning the updating process with the basic S-EXTORT S-MOP, IPP has already decided that the new story should be considered as a variant of the first spec-MOP created. IPP is able to create another new spec-MOP from S3, this one including all the features of the first spec-MOP, plus the RELEASE-HOSTAGES and GET-RANSOM results of S-EXTORT. IPP has noticed that these are frequent results of kidnappings of businessmen in Italy. So after having read these three stories, IPP has begun to create a model of kidnapping in Italy. This is the sort of behavior displayed by interested human readers. It is also the kind of behavior that cannot be captured by an understanding system that does not involve long-term memory. CONCLUSION The inclusion of long-term memory as a part of IPP has been a key factor in allowing the program to be a powerful, robust understanding system. (To date it has a vocabulary of about 3200 words and has processed over 500 stories from newspapers and the UP1 news wire, producing an accurate representation for 60 - 70%.) Furthermore, the addition of memory makes possible the inclusion of many familiar language-related phenomena, such as reminding and generalization, into a model of the understanding. As shown in the examples in this paper, memory updating can be implemented as a natural part of language processing. In fact without it, the ability of a program to fully understand a story is doubtful, since it cannot improve its knowledge of the world, and adapt its processing to fit the knowledge it has obtained - important measures of the human understanding ability. A program without episodic memory can never conclude that Italian kidnappings are often against businessmen, or any of a myriad of other generalizations that people make all the time. And that is a crucial part of understanding. [II [21 [31 [41 I51 t61 [71 REFERENCES Charniak, E. On the use of framed knowledge in language comprehension. Research Report #137, Department of Computer Science, Yale University, 1978. Cullingford, R. Script application: computer understanding of newspaper stories. Research Report #116, Department of Computer Science, Yale University, 1978 DeJong, G. F. (1979) Skimming stories in real time: An experiment in integrated understanding. Research Report #158, Department of Computer Science, Yale University. Lebowitz, M. Reading with a purpose. In Proceedings of 17th Annual Meeting; of the Association of Computational Linguistics, g Diego, CA, 1979. Schank, R. C. Reminding and memory organization: An introduction to MOPS. Research Report #l70, Department of Computer Science, Yale University, 1979. Schank, R. C and Abelson, R. P. Scripts, Plans, Goals and Understanding. Lawrence Erlbaum Associates, Hillsdale, New Jersey, 1977. Wilensky, R. Understanding Goal-Based Stories. Research Report #l40, Department of Computer Science, Yale University, 1978. 326	1980	61
58	FAILURES IN NATURAL LANGUACE SYSTEIG : APPLICATIONS ?o DATA BASE WERY SYSTEMS Eric Mays Department of Canputer and Information Science University of Pennsylvania Philadelphia, PA 19104 ABSTRACT A significant class of failures in interactions with data base query systems are attributable to misconceptions or incunplete knowledge regarding the danain of discourse on the part of the user. This paper describes several types of user failures, namely, intensional failures of presumptions. These failures are distinguished fran extensional failures of presumptions since they are dependent on the structure rather than the contents of the data base e A knowledge representation has been developed for the recognition of intensional failures that are due to the assumption of non-existent relationships between entities. Several other intensional failures which depend on more sophisticated knowledge representations are also discussed. Appropriate forms of corrective behavior are outlined which would enable the user to formulate queries directed to the solution of his/her particular task and compatible with the knowledge organization. I INTRCDUcrICN An important aspect of natural language interaction with intelligent systems is the ability to deal constructively with failure. Failures can be viewed as being of two types. One can be ascribed to a lack of syntactic, semantic, or pragmatic coverage by the system. This will be termed system failure , and manifests itself in the inability of the system to assign an interpretation to the user's input. Recent work has been done in responding to these types of failures, see for example, Weischedel and Black [aI, and Kwasny and Sondheimer E31 - A second class of failures may be termed user failures. A user failure arises when his/hermiefs about the domain of system. ** discourse diverge from those of the This work is partially supported by a grant frcm the National Science Foundation, NSF-W 79-08401. Sane user beliefs regarding the domain of discourse are implicitly encoded in questions posedtothe system. The beliefs held by the system are explicit in its knowledge representation, either procedurally or declaratively. To avoid confusion, a clear distinction should be made between failures and errors. An error occurs when the system's response to an input is incorrect. Errors generally manifest themselves as incorrect resolution of ambiguities in word sense or modifier placement. These errors would usually be detected by the user when presented with a paraphrase that differs in a meaningful way frcan the original input [61. More serious errors result fran incorrect coding of domain knowledge, and are often undetectable by the user. This paper concerns itself with the recognition and correction of user failures in natural language data base query systems -- in particular, failures that arise fran the user@s beliefs about the structure, rather than the cOntent, of the data base. The data base model that has been implemented for the recognition and correction of simple user failures about the data base structure is presented. Several other failures which depend on more sophisticated knowledge representation are also discussed. II PRESUPFDSITICNANDPRESUMPTICN The linguistic notion of presupposition provides a formal basis for the inference of a significant class of user beliefs. There is a less restrictive notion, presumption, which allows the inference of larger class of user beliefs, namely, that knowledge which the user must assume when posing a question. A presupposition is a proposition that is entailed by all the direct answers of a question. A presumption is either a presupposition or it is a proposition that is entailed by all but one of the direct answers of a question [21. Hence, presupposition is a stronger version of presumption , and a presupposition is a preswtption by definition. For example, question (la) has several direct answers such as, "John", "Sue" , etc., and, of course, "no one". Proposition (lb) is entailed by all the direct answers to (la) except the last one, i.e., "no -------m-e The complete definition of includes the condition that the question, direct answer pair presupposition. presupposition negation of a entails the 327 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. one". Therefore, (lb) is a presumption of (la). Proposition (Id) is a presupposition of (lc), since it is entailed by all of the question's direct answers. (la) Which faculty members teach CSEllO? (lb) Faculty members teach CSEllO. (lc) When does John take CSEllO? (ld) John takes CSEllO. Presumptions can be classified on the basis of what is asserted -- i.e., an "intensional" statement about the structure of the data base or an "extensional" statement about its contents. Thus an extensional failure of a presumption occurs based on the current contents of the data base, while an intensional failure occurs based on the structure or organization. For example, question (2a) presumes propositions (2b), (2c), and ( a) l Presumption (2b) is subject to intensional failure if the data base does not allow for the relation "teach" to hold between "faculty" and "course" entities. An extensional failure of presumption (2b) would occur if the data base did not contain any "faculty member" that "teaches" a "course". Also note that the truth of (2b) is a pre-condition for the truth of (2c) l (2a) Which faculty members teach CSEllO? (2b) Faculty members teach courses. (2~) Faculty members teach CSEllO. (Xi) CSEllO is a course. Although a presumption which fails intensionally will of necessity fail extensionally, it is important to differentiate between them, since an intensional failure that occurs will occur consistently for a given data base structure, whereas extensional failure is a transitory function of the current contents of the data base. This is not meant to imply that a data base structure is not subject to change. However, such a change usually represents a fundamental modification of the organization of the enterprise that is modelled. One can observe that structural modifications occur over long periods of time (many months to years, for example), while the data base contents are subject to change over relatively shorter periods of time (hourly, daily, or monthly, for example). Kaplan [2] has investigated the -tation and correction of extensional failures of preslrmptions. The approach taken there involves accessing the contents of the data base to determine if a presumption has a non-empty extension. The remainder of this paper discusses several ways a presumption might be subject to intensional failure. These inferences are made from the structural information of the data base. III DATA F3ASE M3DEL --- In order to recognize failures of presumptions concerning the structure of the data base, it is necessary to use a robust data model. The discussion here will assume a data base model similar to that proposed by Lee and Gerritsen [4], which incorporates the generalization dimension developed by Smith and Smith [7] into the entity-relationship model [U Basically, entities participate in relationships along two orthogonal dimensions, aggregation bxmg dissimilar entities) and generalization (among similar entities), as well as having attributes that assume values. As an example of this type of structure consider the data base model fragment for a typical university in figure 1. Entity sets are designated by ovals, aggregation relationships by diamonds, and generalization relationships by edges from the super-entity set to the sub-entity set. The parallel arcs denote mutual exclusion. 328 Mutual exclusion is used to infer the difference between "men that are also faculty" (a possibly non-empty set) and "men that are also women" (an empty set by definition), for exmple given figure 1. This distinction can be made by prohibiting the traversal of a path in the data model that includes two entity sets which are mutually exclusive. Furthermore, the path in the generalization dimension is restricted to "upward" traversals followed by "downward" traversals. An upward (downward) traversal is from a sub-entity (super-entity) set to a super-entity (sub-entity) set. This restriction is made to prevent over-specialization of an entity set when traversing downward edges. The set of inferences that can be made in the presence of this restriction is not overly constrained, since any two entity sets that have a ccmmon intersection (sub-entity set) will also have a common union (super-entity set).* IV INTENSICNAL FAILURES A. Non-existent Relationships The most basic intensional failure that can occur is the presumption of a non-existent relationship between entity sets. In the university data base model fragment given above, such a failure occurs in the question "Which faculty take courses?". This question presumes that a "take" relationship could exist between "faculty" and "courses" entities. Since no such relationship can be established, that presumption has failed intensionally. Recognizing the failure is only part of the problem -- it is also useful to provide the user with related intensional knowledge. Given a relation R, entities X and Y, and a failed presumption (R X Y), salient intensional knowledge can be found by abstracting on R, X, or Y to create a new relation. For example, consider the following exchange: Q: Which faculty take courses?" A: "I don't believe that faculty can take courses. Faculty teach courses. Students take courses." A similar failure occurs in the presumption of a non-existent attribute of an entity set. For example, What is the cost of all courses taught by teaching assistants?", incorrectly presumes that in this data base, "cost" is an attribute of "courses". B. Inapplicable Functions Intensional failures may also occur when attempting to apply a function on a domain. The question, What is the average grade in CSEllO?", will cause no processing problems provided grades are assigned over the real numbers. But if grades ranged fron A to F, then the system should inform the user that averages can not be performed on character data. (Note that the clever system designer might trap this case and make numerical assignments to the letter grades.) A more significant aspect is the notion of a function to be meaningful over a particular dcmain. That is, certain operations, even though they might be applicable, may not be meaningful. An example would be "average social security number". The user who requested such a amputation does not really understand what the data is supposed to represent. In such cases a short explanation regarding the function of the data would be appropriate. To achieve this type of behavior, of course, the data base model must be augmented to include type and functional information. C. Higher Order Failures The mutual exclusion operator allows the detection of a failure when the question specifies a restriction of an entity set by any two of its mutually exclusive sub-entity sets. For examplep Which teachers that advise students take courses?" presumes that there could be same "teachers" that are both "faculty" and "students". Since this situation could never arise, given the structure in figure 1 , it should be cammunicated to the user as an intensional failure. If an exhaustiveness operator is incorporated as well, unnecessary restrictions of an entity set by disjunction of all of its exhaustive sub-entity sets can be detected. Although this would not constitute a failure, it does indicate that there is scane misconception regarding the structure of the data base on the part of the user. If the sub-entity sets were known to be exhaustive by the user, there would be no reason to make the restriction. As an example, the addition of the fact that "grads" and "undergrads" were exhaustive sub-entity sets of "students" would cause this misconception to arise in the question "Which students are either grads or undergrads?". The following behavior would be desired in these cases: Q: "Which teachers that advise students take courses?" A: "Faculty advise students. Students take courses. I don't believe that a teacher can be both a faculty member and a student." *See [5] for a more detailed description. 329 D. Data Currency Same failures depend on the currency of the data. One such example occurs in a naval data base about ships, subs, and aircraft. The question "What is the position of the Kitty Hawk?" presumes that timely data is maintained. Actually, positions of friendly vessels are current, while those of enemy ships might be hopelessly out of date. In this case, the failures would be extensional since the last update of the attribute must be checked for currency. It may be the case that sane data is current while other data is not. However, the update processing time lag from actual event occurence to capture in the data base might be sufficiently long that such presumptions might be subject to intensional failure. Thus the user could be made aware that current data was never available. v coNcLus1oN In this paper we have discussed several.kinds of failures of presumptions that depend on knowledge about the structure or organization of the data base. It is important to distinguish between structure and content, since there is a significant difference in the rate at which they change. When responding to intensional failures of presumptions, simply pointing out the failure is in most cases inadequate. The user must also be informed with regard to related knowledge about the structure of the data base in order to formulate queries directed at solving his/her particular problem. The technique described for recognizing intensional failures that are due to the presumption of non-existent relationships between entities and attributes of entities has been implemented. Further work is aimed at developing knowledge representations for temporal and functional information. We hope to eventually develop a general account of user failures in natural language query systems. I would like to thank Peter Buneman, Aravind Joshi, Kathy McKeown, and Bonnie Webber for their valuable camments on an earlier draft of this paper l [2] Kaplan, S.J., Cooperative Responses Fran a Portable Natural Language Data Base Query -em; Ph.D. Dision, Ccmpz and Information Science Department, University of Pennsylvania, Philadelphia, PA., 1979, [31 Kwasny, S.C.r and Sondheimer, N.K., "Ungrammaticality and Extra-Grannnaticality in Natural Language Understanding sys terns" , Proceedings of the Conference of the Association for Cqutational Linguistics, La Jolla, CA., August 1979. [4] Lee, R.M. and Gerritsen, R., "A Hybrid Representation for Database Semantics", Working Paper 78-01-01, Decision Sciences Department, University of Pennsylvania, 1978. [51 Maysr E., "Correcting Misconceptions About Data Base Structure", Proceedings of the Conference of the Canadian Society for Computational Studies of Intelligence, Victoria, British Columbia, Canada, May 1980. [63 McKeown, K., "Paraphrasing Using Given and New Information in a Question-Answer system" , Proceedings of the Conference of the Association for Canputational Linguistics, La Jolla, CA., August 1979. [7] Smith, J.M. and Smith, D.C.P., "Database Abstractions: Aggregation and Generalization", ACM Transactions on Database Systems, Vol. 2, No. 2, June 1977. 183 Weischedel, R.M., and Black, J., "Responding Intelligently to Unparseable Sentences", American Journal of Canputational Linguistics, Vol. 6, No. 2, April-June 1980. [l] Chen, P.P.S., "The Entity-Relationship Model -- Tawards a Unified View of Data", ACM Transactions on Database Systems, Vol. 1, No. 1, March 1976. 330	1980	62
59	WHEN EXPECTATION FAILS: -- Towards a self-correcting inference system Richard H. Granger, Jr. Artificial Intelligence Project Department of Information and Computer Science University of California Irvine, California 92717 ABSTRACT Contextual understanding depends on a reader's ability to correctly infer a context within which to interpret the events in a story. This "context-selection problem" has traditionally been expressed in terms of heuristics for making the correct initial selection of a story context. This paper presents a view of context selection as an ongoing process spread throuqhout the understanding process. This view requires that the understander be capable of recognizing and correcting erroneous initial context inferences. A computer program called ARTHUR is described, which selects the correct context for a story by dynamically re-evaluating its own initial inferences in light of subsequent information in a story. INTRODUCTION Consider the following simple story: [l] Geoffrey Huggins walked into the Roger Sherman movie theater. He went up to the balcony, where Willy North was waiting with a gram of cocaine. Geoff paid Willy in large bills and left quickly. Why did Geoff go into the movie theater? Most people infer that he did so in order to buy some coke, since that was the outcome of the story. The alternative possibility, that Geoff went to the theater to see a movie and then coincidentally ran into Willy and decided to buy some aoke from him, seems to go virtually unnoticed by most readers in informal experiments. On the basis of pure logic, either of these inferences is equally plausible. However, people overwhelmingly choose the first inference to explain this story, maintaining that Geoff did not go into the theater to see a movie. The problem is that the most plausible initial inference from the story's first sentence is that Geoff did go inside to see a movie. Hence, selection of the correct inference about Geoff's goal in this story requires rejection of this initial inference. This paper describes a program called ARTHUR (A Reader THat understands Reflectively) which understands stories like [l] by generating tentative initial context inferences and then re-evaluating its own inferences in light of subsequent information in the story. By this process ARTHUR understands misleading and surprising stories, and expresses its surprise in English. For example, from the above story, ARTHUR answers the following question about Geoff's intentions: Q) Why did Geoff go into the movie theater? A) AT FIRST I THOUGHT IT WAS BECAUSE HE WANTED TO WATCH A MOVIE, BUT ACTUALLY IT'S BECAUSE HE l.&NTED To BUY COCAINE. (For a much more complete description of ARTHUR, see Granger [19801). We call the problem of finding the correct inference in a story the "context-selection problem" (after the "script-selection problem" in Cullingford El9781 and Dejong [1979], which is a special case (see Section 4.2)). All the "contexts" (or "context inferences") referred to in this paper are goals, plans or scripts, as presented by Schank and Abelson [1977]. Other theories of contextual understanding (Charniak C19721, Schank [19731, Wilks [1975], Schank and Abelson [1977], Cullingford [19781, [1978]) involve Wilensky the selection of a context which is then used to interpret subsequent events in the story, but these theories fail to understand stories such as [l], in which the initially selected context turns out to be incorrect. ARThllR operates by maintaining an "inference-fate g rayh", containing the tentative inferences generated during story processing, alony with information about the current status of each inference. BACKGROUND: SCRIPTS, PLANS, GOALS & UNDERSTANDING 2.1 Contextual understanding ARTHUR's representational scheme is adopted from Schank and Abelson's [1977] framework for representing human intentions (goals) and methods of achieving those goals (plans and scripts). The problem ARTHUR addresses is the process by which a given story representation is generated from a story. It will be seen that this process of mapping a story onto a representation is not straightforward, and may involve the generation of a number of intermediate representations which are discarded by the time the final story representation is complete. Recall the first sentence of story r11: "Geoffrey Huggins walked into the Roger Sherman movie theater." ARlHUR's attempt to infer a context for this event is based on knowledge of typical functions associated with objects and locations. In this instance, a movie theater is a known location with an associated activity: "scripty" viewing a movie. Hence, whenever a story character goes to such a location, one of the plausible inferences from this action is that the character may intend to perform this activity. Seeing a movie also has a default goal associated with it: being entertained. Thus, ARTHUR infers that Geoff plans to see a movie to entertain himself. 301 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. When the next sentence is read, "He went up to the balcony, where Willy North was waiting with a gram of cocaine", ARTHUR again performs this bottom-up inference process, resulting in the inference that Geoff may have been planning to take part in a drug sale. Now ARTHUR attempts to connect this inference with the previously inferred goal of watching a movie for entertainment. Now, however, ARTHUR fails to find a connection between the goal of wanting to see a movie and the action of meeting a uocaine dealer. Understanding the story requires ARTHUR to resolve this connection failure. 2.2 Correcting an erroneous inference - Having failed to specify a connecting inferential path between the initial goal and the new action, ARTHUR generates an alternative goal inference from the action. In this case, the new inference is that Geoff wanted to entertain himself by intoxicating himself with cocaine. (Note that this inference too is only a tentative inference, and could be supplanted if it failed to account for the other events in the story.) ARTHUR now has a disconnected representation for the story so far: it has generated two separate goal inferences to explain Geoff's two actions. ARTHUR thinks that Geoff went to the theater in order to see a movie, but that he then met up with Willy in order to buy some coke. This is not an adequate representation for the story at this point. The correct representation would indicate that Geoff performed both of his actions in service of a single goal of getting coke, and that he never intended to see a movie there at all; the theater was just a meeting place. Hence, ARTHUR instead infers that Geoff's action of going into the theater was in service of the newly inferred goal, and discards the initial inference (wanting to see a movie) which previously explained this action. We call this process supplanting an inference: ARTHUR supplants its initial "see-movie" inference by the new "yet-coke" inference, as the explanation for Geoff's two actions. ARTHUR's representation of the story now consists of a single inference about Geoff's intentions (he wanted to acquire some coke) and two plans performed in service of that goal (getting to the movie theater and getting to Willy), each of which was carried out by a physical action (PTRANSing to the theater and PTRANSing to Willy). At this point, the initial goal inference (that Geoff wanted to see a movie) has been supplanted: it is no longer considered to be a valid inference about Geoff's intentions in light of the events in the story. 3.1 OPERATICN OF THE ARTHUR PROGRAM Annotated run-time output The following represents actual annotated run-time output of the ARTHUR program. The input to the program is the following deceptively simple story: [2] Mary picked up a magazine. She swatted a fly. The first sentence causes ARTHUR to generate the plausible inference that Mary plans to read the magazine for entertainment, since that is stored in ARTHUR's memory as the default use for a magazine. ARTHUR's internal representation of this situation consists of an "explanation triple": a goal (being entertained), an event (picking up the magazine), and an inferential path connecting the event and goal (reading the magazine). The following ARTHUR output is generated from the processing of the second sentence. (ARTHUR'S output has been shortened and simplified here for pedagogical and financial reasons.) :CURRENT EXPLANATION-GRAPH: GOAL: (E-ENTERTAIN (PLANNER MARY) (OBJECT MAG)) EW: (GRASP (ACTOR MARY) (OE+JECT M&)) PATHO: (READ (PLANNER MARY) (OBJECT MAG)) ARTHUR's explanation of the first sentence'has a goal (being ENTERTAINed), an act (GRASPiq a magazine) and an inferential path connecting the action and goal (READing the magazine). :NEXTSENTENCE CD: (PROPEL (ACTOR MARY) (OBJECT NIL) (TO FLY)) The Conceptual Dependency for Mary's acti she struck a fly with an unknown object. On: :FAILURE TO CONNECT TO EXISTING GOAL CONTEXT: ARTHUR's initial goal inference (Mary planned to entertain herself by reading the magazine) fails to explain her action of swatting a fly. :SUPPLANTING WITH NEW PLAUSIBLE GOAL CONTEXT: (PHYS-STATE (PLANNER MARY) (OEUECT MAG) (VAL -10)) ARTHUR now generates an alternative goal on the basis of Mary's new action: she may want to destroy the fly, i.e., want its physical state to be -10. This new goal also serves to explain her previous action (getting a magazine) as a precondition to the action of swatting the fly, once AKI'HUR infers that the magazine was the INSTRument in Mary's plan to damage the fly. :FINAL EXPLANATION-TRIPLE: GOALl: (PHYS-STATE (PLANNER MARY) (OBJECT FLY) (VAL -10)) EVl: (GRASP (ACtOR MARY) (OBJECT MAG)) PATHl: (DELTA-CONTROL (PLANNER MARY) (OBJECT Mpc,) EV2: (PROPEL (ACTOR MARY) (OBJECT MAG) (TO FLY)) PAm2: (CHANGE-~Ys-STATE (PLANNER MARY) (ow~cr FLY) (DIRECTION NM;) (INSTR MAG)) 302 This representation says that Mary wanted to destroy a fly (GOALl), so she planned to damage it (PATH2). Her first step in doing so was to get an instrumental object (PATHl). These two plans were realized (Events 1,2) by her picking up a magazine and hitting the fly with it. :READY FOR QUESTIONS: >Why did Mary pick up a magazine? AT FIRST I THOUGHT IT c91S BECAUSE SHE WANTED To READ IT, BUT ACTUALLY IT'S BECAUSE SHE WTED TO USE IT TO GET RID OF A FLY. The question asks for the inferred goal under- lying Mary's action of GRASPing the magazine. This answer is generated according to ARTHUR's suw$aed inference about the action (READ) and the active inference about the action (CHANGE-PhYS-STATE) . The English generation mechanism used is described in Granger [1980]. 3.2 The parsimony principle AKIWUR's answer as given above is not the only possible interpretation of the story; it is only one of the following three alternatives, all of which are valid on the basis of what the story says: (2a) Mary picked up a magazine to read it. She then was annoyed by a fly, and she swatted it with the magazine she was holding. (2b) Mary picked up a magazine to read it. she then was annoyed by a fly, and she swatted it with a flyswatter that was handy. (2~) Mary picked up a magazine to swat a fly with it. The last interpretation (2~) reflects a story representation which consists of a single goal, getting rid of a fly, which both of Mary's actions were performed in service of. The other interpretations both consist of two separate goals, each of which explains one of Mary's actions. In [21, as in [U, the interpretation generated by the reader is the most parsimonious of the possible interpretations. That is, the preferred interpretation is the one in which the fewest number of inferred goals of a story character account for the maximum number of his actions. We ammarize this observation in the, following principle: The Parsimony Principle The best context inference is the one which accents for the most actions of a story character. - 303 In other words, the decisicn to replace a previous inference by a new one is not based on the explicit contradiction of that inference by subsequent information in the story. Example [2], for instance, has three possible interpretations, none of which can be ruled out on strictly logical grounds. Rather, the reader prefers the most parsimnious story representation over less parsimonious ones, that is, the representation which includes the fewest goal inferences to account for the actions in the story. This is true even when it requires the reader to do the extra work of replacing one of its own previous inferences, as in example [2]. CATEGORIES OF ERRONEOUS INEERENCES 4.1 Goals ARTHUR is capable of recognizing and correcting erroneous context inferences in order to maintain a parsimonious explanation of a story. The examples given so far have dealt only with erroneous goal inferences, but other conceptual categories of inferences can be generated erroneously as well. In this section, examples of other classes of erroneous inferences will be given, and it will be shown why each different class presents its own unique difficulties to ARTHUR's correction processes. 4.2 Plans and scripts -- Consider the following simple story: [3] Carl was bored. He picked up the newspaper. He reached under it to get the tennis racket that the newspaper had been covering. This is an example in which ARTHUR correctly infers the goal of the story character, but erroneously infers the plan that he is going to perform in service of his goal. ARTHUR first infers that Carl will read the newspaper to alleviate his boredom, but this inference fails to explain why Carl then gets his tennis racket. ARTHUR at this point attempts to supplant the initial goal inference, but in this case ARTHUR knows that that goal was correctly inferred, because it was implicitly stated in the first sentence of the story (that Carl was bored). Hence ARTHUR infers instead that it erroneously inferred the plan by which Carl intended to satisfy his goal (reading the newspaper). Rather, Carl planned to alleviate his boredom by playing tennis. The problem now is to connect Carl's action of picking up the newspaper with his plan of playing tennis. Instead of using the newspaper as a functional object (in this case, reading material), Carl has treated it as an instrumental object that must be moved as a precondition to the implementation of his intended plan. (Preconditions are discussed in Schank and Abelson [1977]) e ARTHUR recognizes that an object can be used either functionally or instrumentally. Furthermore, when an action is performed as a precondition to a plan, typically the objects used in the action are used instrumentally, as in [31 . ARTHUR's initial inference about Carl's plan was based on the functionality of a newspaper. It is able to supplant this inference by an inference that Carl instead used the newspaper instrumentally, as a precondition to getting to his tennis racket, which in turn was a presumed precondition to using the racket to play tennis with. Hence, correcting this erroneous plan inference required ARTHUR to re-evaluate its inference about the intended use of a functional object. 4.3 Causal state changes -- Consider the following example: [4] Kathy and Chris were playing golf. Kathy hit a shot deep into the rough. We assume that Kathy did not intend to hit her ball into the rough, since she's playing golf, which implies that she probably has a goal of winning or at least playing well. Her action, therefore, is probably not goal-oriented behavior, but is accidental: that is, it is an action which causally results in some state which may have an effect on her goal. This situation differs from stories like [l], 121 and [31, in that ARTHUR does not change its mind about its inference of Kathy's goal. Rather than assuming that .the goal inference was erroneous, ARTHUR infers that the causal result state hinders the achievement of Kathy’s goal. Any causal state which affects a character's goal, either positively or negatively, appears in ARTHUR's story representation in one of the following four relationships to an existing goal: l- the state helps the achievement of the goal; 2 - the state hinders achievement of the goal; 3 - the state aZG%Z the goal entirely; or 4- the state thwarts the goal entirely. If ARTHUR did assume that Kathy's shot was intentional, then the concomitant inference is that she didn't really want to win the game at all; or, in other words, that the initial inference was erroneous. This is the case in the following example: [S] Kathy and Chris were playing golf. Kathy hit a shot deep into the rough. She wanted to let her good friend Chris win the game. Understanding this story requires ARTHUR first to infer that Kathy intends to win the game; then to notice that her action has hindered her goal, and finally to recognize that the initial goal inference was erroneous, and to supplant it by the inference that Kathy actually intended to lose the game, not win it. 4.4 Travelling down the garden path . . . -- If the correct context inference for a story remains unknown until some significant fraction of the story has been read, the story can be thought of as a "garden path" story. This term is borrowed from so-called garden path sentences, in which the correct representation of the sentence is not resolved until relatively late in the sentence. We will call a garden path story any story which causes the reader to generte an initial inference which turns out to be erroneous on the basis of subsequent story events. Obvious examples of garden path stories are those in which we experience a surprise ending, e.g., mystery stories, jokes, fables. Since ARTHUR operates by generating tentative initial inferences and then re-evaluating those inferences in light of subsequent information, ARTHUR understands simple garden path stories. Not all garden path stories cause us to experience surprise. For example , many readers of story [2] do not notice that Mary might have been planning to read the magazine, unless that intermediate inference is pinted out to them. Hence we hypothesize that the processes involved in understanding stories with surprise endings must differ from the processes of understanding other garden path stories. Hence, ARTHUR's understanding mechanism is not entirely psychologically plausible in that it does not differentiate between stories with surprise endings and other garden path stories. A more sophisticated version of ARTHUR (call it “Macro-MTHUR”) might differentiate between "strong" default inferences and "weak" tentative inferences when generating an initial context inference. If a strong initial inference is generated, then MacARTHUR would consciously "notice" this inference being supplanted, thereby experiencing surprise that the inference was incorrect. Conversely, if the initial inference is weak, MacAKTHUR may not commit itself to that inference, but rather may choose to keep around other possible alternatives. In this case MacARTHUR would onlv exmrience further specification of the initial- tentative set of inferences, rather than supplanting a sinqle strong inference. The question of-when readers processes consciously versus unconsciously is still an open question in psychology. Future psychological studies of the cognitive phenomena underlying human story understanding (such as in Thorndyke [19761, [1977], Norman and Bobrow [19751 , and Hayes-Roth [1977], to name a few) may be able to provide data which will shed further light on this issue. 304 a)NCLUSIONS: WHERE WE'VE BEEN/ WHERE WE'RE HEADING REFERENCES 5.1 Process and representation in understanding This paper has presented a process for [ll Cullingford, R. (1978). Application: Script Computer Understanding of Newspaper Stories. Ph.D. Thesis, Yale University, New Haven, building story representations which contain inferences not explicitly stated in the story. The representations themselves are not new; they are based on those presented by Schank and Abelson [19771. What is new here is the process of corm m [2l Granger, R. (19813). Adaptive Understanding: Correcting Erroneous Inferences. Ph.D. Thesis, Yale University, New Haven, Conn. arriving at a story representation. Most contextual understanders (e.g. Charniak E19751, Cullingford [1978], Wilensky [1978]) would fail to arrive at the correct story representations for any of the examples in this paper, because initial statements in the examples trigger inferences which prove to be erroneous in light of subsequent story statements. ARTHUR's processing of these examples shows that arriving at a given story representation may require the reader to generate a nunber of intermediate inferences which get discarded along the way, and which therefore play no role in the final representation of the story. [3] Hayes-Roth, B. (1977). Implications of human pattern processing for the desiyn of artificial knowledge systems. In Pattern-directed inference systems (Waterman and Hayes-Roth, eds.). Academic Press, N.Y. [4] Kintsch, h. (1977) e Memory @ Cognition. Wiley, New York. [S] Norman, D. and Bobrow, D (1975). On the role of active memory processes in perception and cognition. In The structure of human memory -- (Coffer, C., ed.). Freeman, San Francisco. Thus a final story representation may not completely wecify the process by which it was generated, since there may have been intermediate inferences which are not contained in the final [6] Schank, R. C. and Abelson, R. P. (1977). Scripts, Plans, Goals and Understanding. Erlbaum Press, Hill=, E. representation. Yet we know that when people have understood one of these examples, they can express these intermediate inferences with phrases like "At first I thought X, but actually it's Y." ARTHUR keeps track of its intermediate inferences while understanding a story, and maintains an "inference-fate graph" containing all inferences generated during story processing, whether they end up in the final story representation or not. [71 Thorndyke, P. (1977). Pattern-directed processing of knowledge from texts. In Pattern-directed inference systems (Waterman and Hayes-Roth, eds.). Academic Press, N.Y. [8] Wilensky, R. (1978). Understanding Goal-based Stories. Ph.D. Thesis, Yale University, New Haven, Conn. The point here is that the relationship between a given story representation on the one hand, and the process of arriving at that representation on the other, may be far from straightforward. The path to a final story representation may involve sidetracks and spurious inferences which must be recognized and corrected. Therefore, specifying the representations corresponding to natural language inputs is not enough for a theory of natural language processing; such a theory must also include descriptions of the processes by which a final representation is constructed. ARTHUR has demonstrated one area in which specification of process and representation diverge: the area of correcting erroneous inferences during understanding. Further work will be directed towards specifying other conditions under which process and representation are not straightforwardly related in natural language tasks. [9] Wilks, Y. (1975). Seven Theses on Artificial Intelligence and Natural Language, Research Report No. 17, Instituto per gli Studi Semantici e Cognitivi, Castagnola, Switzerland. 305	1980	63
60	ORGANIZING MEMORY AND KEEPING IT ORGANIZED Janet L. Kolodner Dept. of Computer Science Yale University, P. 0. BOX 2158 New Haven, CT 06520 ABSTRACT Maintaining good memory organization is important in large memory systems. This paper presents a scheme for automatically reorganizing event information in memory. The processes are implemented in a computer program called CYRUS. INTRODUCTION People are quite good at retrieving episodes from their long term memories. In fact, they are much better at information retrieval than any current computer system. Psychologists have described human memory as a reconstructive process (e.g., 131). When people attempt to remember events and episodes from their lives, they often must go through a complicated reasoning and search process ([6l and [51). These processes are dependent on good memory Drganization. In order to keep memory well organized as new data is added, memory organization must support the creation of new memory categories and the building up of generalized knowledge. If a memory held only 10 events that could be described as meetings, a "meetings" category would be useful. But, unless new meeting sub-categories were created as additional meetings were added to the memory, retrieval of meetings would become very inefficient. Thus, a memory system needs the ability to create new categories automatically from old ones. CYRUS is a computer program which implements a theory of human memory organization and retrieval. The program is designed to store and retrieve episodic information about important people, and is based on a theory of the way people organize and remember information about themselves. Right now, it,holds information about Secretaries of State Cyrus Vance and Edmund Muskie. CYRUS answers questions posed to it in English, using search strategies [51 to search its memory and a set of constructive strategies 121 to construct search keys. CYRUS is connected to the FRUMP program [ll, which produces summaries of stories off the UP1 This work was supported in part by the Advanced Research Projects Agency of the Department of Defense and monitored under the Office of Naval Research under contract NOOOl4-75-C-1111. wire. When FRUMP sends CYRUS new information about Vance or Muskie, CYRUS automatically updates its memory. As it does that updating, it reorganizes its memory and builds up generalized knowledge about the information it stores. CYRUS' memory is organized around episodes using Memory Organization Packets (MOPS) [41. Because episodes include references to persons who participated in them, their locations, other episodes they are related to, etc., they are good global organizers of otherwise disjoint information. For example, Camp David and Menachim Begin have in common that Camp David was the location of an important summit conference that Begin attended. There are a number of problems that must be addressed in maintaining a self-updating memory. 1. what constitutes a good category in memory, i.e., what attributes does a good category have? 2. what kind of knowledge must be stored about each category to enable retrieval and new category creation? 3. how do categories relate to each other? 4. when is it appropriate to reorganize a category into smaller pieces? 5. how can generalized knowledge be added to new categories? The remainder of this paper will address some of these problems. RECOGNIZING SIMILARITIES BETWEEN EPISODES People notice common aspects of episodes and make generalizations in the normal course of understanding. Reorganization of memory requires noticing similarities and making generalizations based on those similarities. Generalized knowledge is needed to predict future occurrences, to elaborate on a context being understood, to help direct memory update, and as an aid in directing search during retrieval. Like people, CYRUS notices similarities between episodes and makes generalizations from them. Similar episodes in CYRUS are stored in the same MOP, along with the generalized knowledge built UP from them. MOPS act as event categories in memory holding episodes and knowledge about those episodes. The generalized information a MOP 331 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. holds resides in its "content frame" 141 and includes such things as typical preconditions and enablement conditions for its episodes, their typical sequence of events, larger episodes they are usually part of, their usual results, typical location, duration, participants, etc. The structure of individual episodes provides a framework for deciding whether two episodes are similar to each other. If, on two different trips, Vance is welcomed at the airport and then has a meeting with the president of the country he is in, then the episodic structure of the two trips will look alike, and we can say that the two episodes are similar. While on the second trip, he might be reminded [41 of the first one because of their similarities. In the same way, a person hearing about the two trips might be reminded of the first when hearing about the second. If the result of the first trip had been a signed accord, then he may predict, or at least hope, that an accord would be signed at the end of this trip also. If an accord is signed, he will generalize that when the first event of a diplomatic trip is a meeting with the head of state, then an accord will be reached. Later, he will be able to use that knowledge in understanding and retrieval. REORGANIZING EVENT CATEGORIES In order for such reminding and subsequent generalization to occur in CYRUS, its MOPS must be highly structured internally. Episodes are indexed in MOPS according to their important aspects. Indexing in a MOP is by content-frame components and includes sequence of events, location, participants, duration, etc. When an event is indexed similarly to an event already in memory, reminding occurs and generalizations are made based on their similarities. As a result, a sub-MOP of the larger MOP is formed to hold those episodes and their generalizations. With the addition of more episodes, new sub-MOPS are further divided. In creating new MOPS and building up generalized information, new knowledge, which can be used for later understanding and retrieval, is added to the data base. The actual processing when an event is added to a MOP depends on its relationship to events already in the MOP. One of the following four things is true about each component of an event description: 1. It is unique to this event in the MOP 2. It is semi-unique to this event in the MOP (it has happened once or a small number of times before) 3. It is often true of events in the MOP 4. It is typical of events in the MOP In case 1, when the descriptive property is unique, the event is indexed under that aspect in the MOP. For instance, one of the discriminations CYRUS makes on meetings is the topic of the contract being discussed. A meeting about the Camp David Accords is indexed under "contract topic = peace", and a meeting about military aid- to Pakistan is indexed under "contract topic = military aid". The first time CYRUS hears about a meeting in which Vance discusses mili tary aid, it will index that meeting uniquely in the "diplomatic meetings" MOP under the property "contract topic = military aid". If it were also the first meeting he had with a defense minister, then it would also be indexed uniquely under the occupation of its participants (because meetings are occupational). If a property being indexed has occured once before in a MOP (case 21, then reminding occurs, the two events are compared to see which other common aspects they have, and generalizations are made. When Vance meets again about military aid, CYRUS is reminded of the prior meeting because both have the same topic. It checks the descriptions of both to see what other similarities they have. If both, for example, are with defense ministers, it will conclude that meetings about military aid are usually with defense ministers. It also begins indexing within the new MOP: Jr********* Adding SMEET actor (Vance) others (defense minister of Israel) topic (Military aid to Israel) place (Jerusalem) to memory . . . Reminded of SMEET actor (Vance) others (defense minister of Pakistan) topic (Military aid to Pakistan) place (Washington) because both are "diplomatic meetings' both have contract topic "military aid" creating new MOP: meetings about military aid generalizing that when Vance meets about military aid, often he meets with a defense minister *Jr***** Later, if CYRUS hears about a third meeting whose topic is military aid, it will assume that the meeting is with the defense minister of the country requesting aid (unless it is given contrary information). If asked for the participants of that event' i&w.'.11 be able to answer "probably the defense minister". If, on the other hand, a number of meetings about military aid with participants something other than defense ministers are added to memory, CYRUS will remove that generalization and attempt a better one instead. On entering the next meetings about military aid to memory, CYRUS will index them along with other events already indexed there. A new meeting about military aid will be entered into the "meetings about military aid" sub-MOP of "diplomatic meetings" and will be indexed within that MOP (case 3). In this way, reminding, generalization, and new MOP creation will occur within newly created MOPS. If a new meeting about military aid to Pakistan is added to memory, CYRUS will be reminded of the first because both will be indexed under "contract sides = Pakistan" in the "meetings about military aid" MOP. 332 No discrimination is done on properties that are typical (case 4) of events in a MOP (i.e., almost all events in the MOP fit that description). In that way, generalization can control the expansion of MOPS in memory. If memory has generalized that meetings are called to discuss contracts, then the fact that the topic of a later meeting is a contract will never be indexed. Appropriate aspects of the contract, however, will be indexed. Thus, if a new event with no unique aspects is added to memory, reminding of individual events does not occur, but generalizations already made are confirmed or disconfirmed. Disconfinned generalizations are removed. When CYRUS hears about yet another meeting in the Mid East about the Camp David Accords, it will not be reminded of any specific episodes, but will put the new meeting into the MOPS it fits into. Jr******** Adding SMEET actor (Vance) others (Begin) topic (Camp David Accords) place (Jerusalem) to memory . . . Putting it into MOP: meetings with Begin in Israel confirming generalizations Putting it into MOP: meetings about the Camp David Accords with Israeli participants confirming generalizations Putting it into MOP: meetings in Israel confirming generalizations a.. Jr******** IMPLICATIONS IN RETRIEVAL What are the implications of this indexing scheme in retrieval? The retrievability of an event depends on how distinct its description is, or how many of its features turn out to be significant. As events with similar properties are added to memory, their common aspects lose significance as good retrieval cues and category specifiers (case 4). An event with no unique or semi-unique descriptors will become lost in memory or "forgotten". Since events are indexed by their differences, they can be retrieved whenever an appropriate set of those differences is specified, but specification of only common aspects of events will not allow distinct episodes to be retrieved. Generalized knowledge can be used during retrieval in a number of ways. One important use is in guiding search strategy application (see 151). Generalized knowledge can also be used for search key elaboration. There is not always enough information given in a question to direct search to a relevant MOP or to a unique episode within a MOP. Generalizations and a MOP's indexing scheme can be used to direct the filling in of missing details. Only those aspects of a MOP that are indexed need be elaborated. Generalized information can be used to answer questions when more specific information can't be found. If CYRUS has made a generalization that Gromyko is usually a participant in SALT negotiations, it will be able to give the answer "probably Gromyko" to "Last time Vance negotiated SALT, who did he meet with?", even if it could not retrieve the last negotiating episode. In the same way, generalizations can be used for making predictions during understanding. CONCLUSIONS Good memory organization is crucial in a large information system. Some important processes memory organization must support include dynamic reorganization of memory, creation of new memory categories, and generalization. In this paper, we've tried to show how a long-term memory for episodic information can keep itself usefully organized. This requires a good initial organization plus rules for reorganization. Reorganization of memory happens through reminding, or noticing similarities between episodes, and generalization. The generalizations produced are useful both for controlling later memory reorganization and for retrieval. Some related problems which have not been addressed in this paper, but which are important, are updating generalized knowledge (particularly recovery from bad generalizations), judging the usefulness of new categories, and guiding indexing properly so that new MOPS and generalizations are useful and relevant. Those topics are currently being addressed. REFERENCES [ll DeJong, G. F. (1979). Skimming stories in real time: An experiment in integrated understanding. Research Report #158. Department of Computer Science, Yale University. [21 Kolodner, J. L. (1980). Organization and Retrieval from Long Term Episodic Memory. Ph.D. Thesis (forthcoming). Department of Computer Science, Yale University. [31 Norman, D. A. and Bobrow, D. G. (1977). Descriptions: a basis for memory acquisition and retrieval. Report #7703. Center for Human Information Processing, La Jolla, California. [41 Schank, R. C. (1979). Reminding and memory organization: An introduction to MOPS. Research Report #170. Department of Computer Science, Yale University. 151 Schank, R. C. and Kolodner, J. L. (1979). Retrieving Information from an Episodic Memory Research Report #159. Department of Computer Science, Yale University. also in IJCAI-6. [61 Williams, M. D. (1978). The process of retrieval from very long term memory. Center for Human Information Processing Memo CHIP-75, La Jolla, California. 333	1980	64
61	Meta-planning Robert Wilensky Computer Science Division De artment of EECS University o F California, Berkeley Berkeley, California 94720 1.0 INTRODUCTION This paper is concerned with the problems of planning related and understanding. because These problems are natural understander must apply knEwledge about people s language goals and necessary f lans in order to make the inferences o in a sto explain the behavior of a character story u~er~~~~;;;ky, 19,;3Fa). a planner, it mus? Thus while embody a theory of plilning knowledge. I have construction story dt?;e+;get rl * understanding PIZYhApp?ieZh$Z anliZm> F": concerned not with t R rogram. This paper is itself, but e understanding mechanism which is that part of its inde lanning knowledge is used to 21 * endent of whe her that knowledge if someone's behavior or to generate a p% "f,'", one's own use. One part of this theory of knowledge is essentially world knowled includes a plan;;i;fg e. classification of in entional f them, plans are used to achieve goals, etc.) and an actual body of knowledge about particula; elements (e getting some hing from someone 'f' 2 asking for som 7 thing is a way of . When one attempts to use this world knowledge to understand the intentions of a story's characters, a number of problems soon become apparent. particular, what ' difficult in understan%ng a person s behavii: iz iE':perating under, but the fact so much understanding the goal and that t#$% are usually numerous a situation. It is f oals and plans present in he interactions between these intentional elements that cause much of complexity in both understanding and planning. For example, consider the stories: following (1) John was in a hurr Vegas, I* to get to Las but he no iced that there were a lot of cops around so he stuck to the speed limit. (2) John was noticed eating dinner when he to break in to his house. that a thief was ti$;;f he finished his dessert, John called the police. Likewise, (2) strikes most strange since John should have rea%", 'Fo t:z z intruder more strongly. The unusualness of this ;;; ry is due not to kno;r;dge ;,";ut the plans goals involved, scheduling of these plans. ap arent unproductive 1 more intelligent planner would have dealt with t? e threat immediately, and then perhaps returned to ;+s meal when that situation had been disposed . Thus to understand the behavior of character, or to enerate an intelligent plana it is necessary f0 take into account th;! interacti ns between goals. programs. e.g., Sussman 1975, 9 y;;;py;+f;g deal interactions bv providi&tkpec??-!E pr~&a~~%emechanisms to deal with particular situations. Sussman s HACKER has a celebrated FZL tZ""~iEt knows about goals clobbering "brother oalslt, and detects this bug in plans suggested $Y the plan synthesizer. The difficulty with this ty e is that burying this knowledge a t of solution out how to plan in a rocedure assures that such knowled not %e shared by a program that wishe 2 et;o;zi this knowledge to understand someone else's behavior in a complicated situation. addition, there is a lot i"f structure as I hope to show to this knowle6ge that is missed in this fashion, to the tasks and which is extreme~l.l~seful both of planning as as plan understanding. 2.0 META-PLANNING One solution to this problem is to create a second body of planning knowledge that is called meta-planning. By this I mean that knowledge about now to plan should itself be expressed in P ;;ys,;f a set of g alsaf,;r thesg;an;pg '5 E recess meta-goals , achieve th meta-plans e~re~m~~Z1~n~~.atheMe~~~~oa~~~~~~~~ mechanism (or plan understa der) that is u produce a plan of action if or explanation 7 ed to from ordinary plans. For consider the following situation, example either Prom the point of view of plan understanding or plan generation: (7) John's wife called him and told him they were all out of milk. He decided to pick some up on his way home from work. Most intelligent planners would come up with John's -plan, assumi by a grocery store on T they knew that they pass he ;;ute home. In order to produce this plan, is necessary to go through the following processes: 1. 334 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. 2. A&j;sting one's plans according1 case, the plan is modifie 3' so Ia: to Produce a route that takes the planner near the grocery store. 2. The "go home" plan is suspended at the point at which the grocery store is reached. 3. 4. The "get milk" plan is executed. The "go home" plan is resumed. 2.1 Kinds Of Meta-goals The following is a brief description of the more important meta-goals along with the situations in so far encountered, which the and some standard plans applicable I arise o them. This list is not meant to be corn lete. merely reflects the current sta e ? of 0:: analysis. META-GOALS, SITUATIONS, AND META-PLANS 1 . Don't Waste Resources Situations to Detect 1 . Goal Overlap Associated meta-plans: 1 . Schedule Common Subgoals First 2. Plan Integration In terms.of meta-planning, this situation has the ;f;;o;;; structure: There ' an imnortant - d "Don't Waste ResouEes". Plan Piggybacking (find a new 'a g~~nflalls k",;k goa~;~ultaneously meta-pian fulfills the 'Don't Waste Resources" meta-goal. The advantage of the meta- is that the problem of how to Ei lanning approach eal with complex goal interactions can be stated as a problem to be solved ,,by applies to the same planning mechanism one ordinary" goals. For example, one may first try out a number of canned sol;;;ons, then some standard planning procedures, lf all else fails, try to construct a novel solution. Note that there are at least three important differences between meta-plannin and planning using constraints + Constraints and plan generators Ze asGike:ZE in that constraints reject plans, but they don't themselves propose new ones. In constrast meta-goals not violations, but su gest only pick up f new plans to fix the problem. Me a-goals are declarative structures, and thus may be used in the ex lanation process as well as in planning. In 3 are domain a dition, independent, meta-goals encoding about planning in general. only knowledge McDermott's notion of a AotP,'n 0; secondary task comes closest to meta-planning I propose here. A policy is essentially constraint. ex licitly Theli;;rnary x* ifferences r;epeEezz;ted polic 3 and a - oal are inclu e goals tha 'i that meta-goal: are not constraints necessarily facts about p ,"~ni~~'as E meta-goals refer CXI;T,;~ policies their domain, include domain information; may specific of policies often entail the creation psuedo-tasks, whereas meta-goals have meta-plans that deviate less from the structure of normal plans. Hayes-Roth and Hayes-Roth (1978) uses the term meta-nlannina to refer to decisions about the planning process. While my use of the term similar to theirs the cyanning decisions under t K include all types of is meta-planning is name, and their not formulated in terms of explicit meta- 5 oals term to and meta-plans. I use the re er of this knowledge t and only to a subset convenien ly only when that knowledge is expressable in meta-goals and meta-plans. terms of explicit 2. Multiple Planning Options (more l%tin g:l) plan is applicable to a Associated meta-plans: 1 . Select Less Costly Plan 3. Plan Non-integrability situations in which the execution o two plahs will adversely affect one another e.5, one undoes B subgoal established by the other Associated meta-plans: 1. Schedule Sequentially 4. Recurring Goals (A goal repeatedly) arises Associated meta-plans: 1. 9 ubsume Recurring Goal Establish state that fulfills a r re:ondition for a plan for ndures he goal and which 7 over a P%~ of time see Wilensky, -5. Recursive subgoals (a subgoal is identical to a higher level goa , causing a potential infinite loop $ Associated meta-plans: 1. Select Alternate Plan 2. Achieve As Many Goals As Possible Situations to detect 1. Goal Conflict Associated meta-plans: various conflict resolution plans (see Wilensky 1978a) 2. Ass mme;;$;otGoal 3[ Conflict (Both goa s be accomplished‘if A kgr ormed F erformed beff;;; B, but B being difficulty) poses no 335 Associated meta-plans: Schedule Innocuous Plan First 2. Plan Splicing (If one plan has already been started, suspend it 6 divert to the other plan, ie"w p%?%s been executed) original plan when 3. Goal Competition (Goal interference with the goal of another planner) Associated meta-plans: 1. various anti-plans (plans to deal specificly with opposition) 2. various plans for resolving the competition 3. Maximize the Value of the Achieved Goals Situations to detect 1. Unresolvable Goal Conflict Associated meta-plans: 1. Abandon Less Important Goal 4. Don't Violate Desireable States Situations to detect 1. Danger Associated meta-plans: 1. Create a preservation goal 2. Maintenance Time Associated meta-plans: 1. Perform Maintenance 3. Anticipated Preservation Goal performance of another plan $?? cause the plann r e to preservation goal have a Associated meta-plans: 1. Select Alternate Plan 2. Protective Modification (Modify original plan so as n t to provoke preservation goal P Ef a stored lan for this % rotective clot ing, it R- wou d f oal is to wear be scheduled efore the initial plan. If not, then we could establish a subgoal of getting a raincoat. This mia:; spawn a plan that involves going would violate outside, the Recursive Subgoals condition. another x: The ~;ta--~la;ai~e;; * lan. t?nd to choose "Achieve one, the s Many Subgoals As Possible" meta-goal is activated, as a goal conflict is now seen to exist. resol- ution The meta-plans for goal conflict are attempted. If they fail, then an unresolvable goal conflict situation exists, and Maximize the Goals Achieved is activated. The meta-plan here is to abandon the less goal. The nlanner selects whichever important goal he values more and then abandons the other. 3 .O APPLICATIONS We are meta-planning understanding interactions :Ytple goal length el currently attemptin F use in two programs. AM,? sto ?$StE:derstand uses kno;;EtfEsabout Y w involving S. As PAM has been discussed sewhere, we will forego a discuss- ion of rts use of meta-planning here. Meta- P lanning is also being used in the developmen of a planning program called PANDORA (Plan ANalyzer with D namic Revision and Application . ai Organization, PANDORA is given a d,,esrcrit ;io;ozfsa $tuation and. creates. a plan R PANDORA . may have in that situation. developmen$z, dynamically told about new and changes it plans accordingly. References 11 Hayes-Roth, B. and Hayes-Roth, F. (1978). Cognitive Processes in Planning. RAND Report R-2366-ONR. 21 McDermott, Drew (1?78>. PLanning and $cting. In Cognitive Science vol. 2, no. . 31 Sacerdoti, E. D. (1977). A Structure for Plans and Behavior. ElZevier North-H?iIlanu, Amsterdam. 41 Schank, R. C. and Abelson, R. P. (1977). %%%%in~?@~'L%%&e%$lbaum Press, Billsdale, N.J. 51 Sussman G. J. (1975). A Corn uter Model of $%:l~o~;quisition. Aiiie?&&ETs~,- 61 Wilensky R. (1978). Understanding goal-based stories. Yale University Research Report #140. 336	1980	65
62	NARRATIVE TEXT SUMMARIZATION Wendy Lehnert Yale University Department of Computer Science New Haven, Connecticut ABSTRACT In order to summarize a story it is necessary to access a high level analysis that highlights the story's central concepts. A technique of memory representation based on affect units appears to provide the necessary foundation for such an analysis. Affect units are conceptual structures that overlap with each other when a narrative is cohesive. When overlapping intersections are interpreted as arcs in a graph of affect units, the resulting graph encodes the plot of the story. Structural features of the graph then reveal which concepts are central to the story. Affect unit analysis is currently being investigated as a processing strategy for narrative summarization. When a reader summarizes a story, vast amounts of information in memory are selectively ignored in order to produce a distilled version of the original narrative. This process of simplification relies on a global structuring of memory that allows search procedures to concentrate on central elements of the story while ignoring peripheral details. It is apparent that some hierarchical structure is holding memory together, but the precise formulation of this structure is much more elusive. How is the hierarchical ordering of a memory representation constructed at the time of understanding? Exactly what elements of the memory representation are critical in building this structure? What search processes examine memory during summarization? How are summaries produced after memory has been accessed? In this paper we will outline a strategy for narrative summarization that addresses each of these issues. This proposed representation for high level narrative analysis relies on affect units. An affect unit is an abstract structure composed of three affect states and four affect links. AFFECT STATES AFFECT LINES Positive Events (+> Motivation cm> Negative Events (-1 Actualization (a> Mental States (M) Termination (t) Equivalence (e> ------- This work was supported in part by ARPA contract N00014-75-C-1111 and in part by NSF contract IST7918463. For example, if John wants to buy a house, his desire is a mental state (Ml. If John subsequently buys the house, his desire is actualized by a positive event (+>. But if someone else buys it instead, John will experience that transaction as a negative event (-1 signalling actualization failure. These particular affect states are derived by recognizing an initiated goal (Ml, an achieved goal (+>, and a thwarted goal (-1. The status of a goal is just one way that an affect state can be recognized. A more complete account of affect state recognition is presented in [31. All affect states are relative to a particular character. If another buyer (Mary) takes the house, we have a negative event for John, and a positive event for Mary. We use a diagonal cross-character link to identify their two affect states as reactions to the same event: John Mary wants to buy (, M MJ" wants to buy a house is sold -/+ buys house The above configuration of four states and three links is the affect unit for "competition." Two actors have a goal, and success for one means failure for the other. "Success" and "failure" are primitive affect units contained within the competition unit. Success is recognized whenever a mental state is actualized by a positive event. Failure is the non-actualization of a mental state through a negative event. Now suppose John decides to get even by setting the house on fire. And suppose further that it takes two tries to get it going. John Mary wants to buy L M U ,:.' a wants to buy buys house house is sold desires fire 4- can't set fire e e desires fire "GM 5 gets fire going +-+-k house burns down The sale of the house to Mary motivates John to set the house on fire (Ml. This mental state fails to be actualized (-1 the first time he tries to commit 337 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. arson. But his desire persists in an equivalent mental state (M) and is then successfully actualized (+I by John setting the fire. This fire is a positive event (+I for John, but a negative event (-1 for Mary who suffers a loss. We first derive a baseline summary from the pivotal unit by accessing a "generational frame" associated with the pivotal unit. For example, a generational frame for retaliation is: "When Y caused a negative event for X, X caused a negative event for Y.” "Loss" is an affect unit that occurs whenever a negative event terminates a positive event in the sense of removing whatever satisfaction was derived from that positive event. When a loss wipes out a previous success, we get the affect unit for "fleeting success." When a smaller unit is embedded in a larger unit (e.g. "loss" is embedded in "fleeting success"), we recognize the structure of the larger unit as a "top level" affect unit. Using this convention, our story about John and Mary contains 4 top level affect units. (11 (21 (31 This is a conceptually abstract description of retaliation. To produce a reasonable summary, we must (1) instantiate the generational frame, and (2) augment it with information from units adjacent to the pivotal unit. We will try to convey what's involved by showing how a baseline summary evolves into a reasonable summary with the addition of information from adjacent units. (This sequence is not intended to reflect actual processing stages). Sl = Retaliation (the baseline summary) "When Mary prevented John from gett wanted, John set her house on fire. ing 11 something he s2 = Sl + Competition (11 represents "competition", (21 "fleeting success", and (3) "perseverance after failure." A fourth affect unit is recognized by merging the two equivalent mental states of John: "When Mary bought something that John set her house on fire." wanted, John s3 = S2 + Fleeting Success "retaliation" 'When Mary bought a hous set the house on fire." e that John wanted, John ,(x> mcA- aG; A, s4 = S3 + Perseverance After Failure "When Mary bought a house that John wanted, set the house on fire after two tries." John The unspecified (X) in the retaliation unit can be any affect state. In our story, John's negative event happened to be a positive event for Mary. If the information from the perseverance unit seems less important than the other contributions, there is a good reason for this. "Perseverence after failure" resides between two equivalent mental states that are merged within the retaliation unit. It is often desirable to ignore units that are lost when equivalent mental states are merged. Top level affect units for a narrative can be used as the basis for a graph structure that describes narrative cohesion. The nodes of the graph represent top level affect units, and an arc exists between two nodes whenever the corresponding affect units share at least one common affect state. The affect unit graph structure for our simple story looks like: Suppose for comparison, that John gave up on his intended arson after the first unsuccessful attempt. Then our affect analysis for the story would be a truncated version of the original: wants to buy M MiP wants to buy 4 /------+ buys house house is sold desires fire mS;; can't set fire a 4, Where C = "competition", F = "fleeting success", R = "retaliation", and P = "perseverance after failure." In general, the affect unit graph for a cohesive narrative will be connected. And in many cases, the graph will have a unique node whose degree (number of incident arcs) is maximal over all nodes in the graph. In our example, the retaliation unit has a uniquely maximal degree of 3. We will call any node of maximal degree a "pivotal unit." If a story has a unique pivotal unit, then that unit encodes the "gist" of the story. A good summary for the story will be based on the pivotal unit and its adjacent units. We st ill have a competition unit, but the other level units are now "motivation" and "fail ure": top "motivation" "failure" The a ffect unit graph now contains three connected units , with motivation acting as the pivotal unit: 338 The baseline summary is therefore built from a generational frame associated with motivation: “When a negative event happened to X, X wanted Z.” Augmenting this baseline summary with information from the competition and failure units, we derive a reasonable summary: Sl = Motivation (the baseline summary) “When Mary prevented John from getting something he wanted, John wanted to set her house on fire.” s2 = Sl + Competition “When Mary bought a house that wanted to set it on fire.” s3 = S2 + Failure “When Mary bought a house that John wanted, unsuccessfully tried to set it on fire.” John wanted, John These two examples illustrate how pivotal units and their adjacent units can be used to drive processes of narrative summarization. While many simple stories will succumb to an algorithm that uses a pivotal unit for the baseline summary, other stories yield affect unit graphs that do not have unique pivotal units. For example, consider “The Gift of the Magi” by 0. Henry. In this story a young couple want to buy each other Christmas presents. They are very poor but Della has long beautiful hair, and Jim has a prized pocket watch. To get money for the presents, Della sells her hair and Jim sells his pocket watch. Then she buys him a gold watch chain, and he buys her expensive ornaments for her hair. When they realize what they’ve done, they feel consoled by the love behind each other’s sacrifices. The affect unit analysis is perfectly symmetrical across the two characters. Both characters have affect units for nested subgoals, a regrettable mistake, two distinct losses, and a hidden blessing. The affect unit graph for this story is connected, but there is no unique pivotal unit: Both “HM” and “WM” are pivotal units. These units correspond to their regrettable mistakes. Let the family of a node N be the set of nodes adjacent to N. Then this graph can be partitioned into the families of “HM” and “WM.” “HN”, “WN”, “HLl”, and “WLl” are boundary units in the sense that each of their families cross this partition. It is not easy to come up with a one sentence summary of “The Gift of the Magi ,I’ but it can be done by concentrating on the boundary units of maximal degree (“HN” and “WN”). These are the units for their nested subgoals: “Della sold her long locks of hair to buy her husband a watch chain, and he sold his watch to buy her ornaments for her hair.” This example shows how the summarization algorithm must be sensitive to structural features of affect unit graphs. In this case the connected graph can be partitioned into two families of two pivotal units, and the simplest summary originates from the boundary units of maximal degree. The process of narrative text summarization relies on (1) a high level of conceptual representation that readily encodes coherence within the narrative, and (2) a process of language generation that can easily be driven by that high level memory representation. In this paper we have attempted to show how affect units and their resulting graph structures are well-suited to these requirements. We have necessarily omitted important explanations concerning techniques of recognition for affect units and the processes of generation that express target summaries in English. The representational system itself requires further explication concerning which affect unit configurations are legitimate (there are 15 legal configurations of the form “state” - “link” _ “state” rather than 36). the combinatorially possible Using these 15 primitive configurations, we can represent speech acts, voluntary compliance, coerced compliance, the notion of a double-cross, and similar abstractions of complexity [31 e equivalent conceptual The use of affect units in narrative summarization is currently being explored by psychological experiments on text comprehension and within a computer implementation for the BORIS system 121. While related work on text summarization has been conducted using story grammars , there are serious flaws in that approach due to the top-down nature of story grammars Ill. These difficulties will not arise with affect unit approach because affect units are constructed by bottom-up processing at the time of understanding. The resulting affect unit graphs are consequently far more flexible in their content and structure than the rigid hierarchies of fixed story grammars. This flexibility is the key to recognizing a diverse range of plot structures without recourse to an a priori taxonomy of all possible plot types. REFERENCES 111 Black, J. B., and Wilensky, R. (1979) “An Evaluation of Story Grammars .” Cognitive Science. Vol. 3, No. 3, pp. 213-230. [21 Dyer, M. and Lehnert, W, (1980) “Memory Organization and Search Processes for Narratives . ” Department of Computer Science TR i/175. Yale University. New Haven, Conn. [31 Lehnert, W. (1980) “Affect Units and Narrative Summarization” Department of Computer Science TR #179. Yale University. New Haven, Conn. 339	1980	66
63	HEARSAY-III: A Domain-Independent Framework for Ex Robert Balzer Lee Ermzrn Philip London Chuck Williams USC/Information Sciences Institute* Marina del Rey, CA 90291 Abstract Hearsay-Ill is a conceptually simple extenslon of the basic ideas in the Hearsay-II speech-understanding system [a]. That domain- dependent expert system was, in turn, a product of a tradition of increasingly sophisticated production-rule-based expert systems. The use of production systems to encapsulate expert knowledge in manageable and relatively independent chunks has been a strong recurrent theme in Al. These systems have steadily grown more sophisticated in their pattern-match and action languages, and in their conflict-resolution mechanisms [la]. In this paper, we describe the Hearsay-Ill framework, concentrating on its departures from Hearsay-II. 1. The Heritage From Hearsay-II Hearsay-II provided two major advances -- the structuring of the workspace, called the bhckboard in Hearsay, and the structuring of the search, via scheduling mechanisms. The blackboard provided a two-dimensional structure for incrementally building hierarchical interpretations of the utterance: - levels which contained different representations (and levels of abstraction) of the domain (phones, syllables, words, phrases, etc.). - a location dimension (the time within the spoken utterance) which positioned each partial interpretation within its level. Knowledge sources (KSs), relatively large production rules, were agents which reacted to blackboard changes produced by other KSs and in turn produced new changes. The expertise was thus organized around the activity of building higher-level, more encompassing partial interpretations from several nearby lower- level partial interpretations (e.g., aggregating three contiguous syllables into a word) and producing lower-level ones from higher-level ones (e.g., predicting an edjacent word on the basis of an exisiting phrase interpretation). Within this aggregation-based interpretation-building paradigm, Hearsay-II also provided a method for exploring alternative interpretations, i.e., handling search. Interpretations conflicted if they occupied the same or overlapping locations of a level; conflicting interpretations competed as alternatives. Thus, in addition to organizing activity around the interpretation-building process, Hearsay-II also had to allocate resources among competing interpretations. This required expertise in the form of critics and evaluators, and necessitated a more complex ‘This research wss supported by Defense Advrnced Research Projects Agency contrect DAK15 72 C 0308 View.s and conclusions contained in this document ere those of the authors and rho&l not be interpreted em representing the official opinion or policy of DARPA, the U.S. Government, or eny other person or egency connected with thorn. scheduler, which at each point one previously-matched KS. chose for execution the action of 2. The Directions for Hearsay-III To this heritage, we bring two notions that motivate most of our changes: - Through simple generalization, the can be made domain independent. Hearsay approach - Scheduling is itself so complex a task that the Hearsay blackboard-oriented knowledge-based approach is needed to build adequate schedulers.* Our generalizations activities: consist of systematizing the main blackboard - aggregating several interpretations at one a composite interpretation at a higher level, level into - manipulating alternative interpretations (by creating a placeholder for an unmade decision, indicating the alternatives of that decision, and ultimately replacing the placeholder by a selected alternative), and - criticizing proposed interpretations. The complexity of scheduling is handled by introducing a separate, scheduling blackboard whose base data is the dynamically created activation records of KSs. These include both the domain-dependent KSs, which react to the regular, domain blackboard, and scheduling KSs, which react to changes on the scheduling blackboard as well. The organization of these activations (with agendas, priorities, etc.) is left to the application writer; Hearsay-Ill provides only the basic mechanisms for building expert systems. Thus domain KSs can be viewed as the legal move generators (competence knowledge) with the scheduling KSs controlling search (performance knowledge). 3. Blackboard Structure In Hearsay-II, nodes on the blackboard, which represented partial interpretations, were called hypotheses. In Hearsay-Ill, we adopt the more neutral term unit. Hearsay-Ill provides primitives for creating units and aggregating them, i.e., associating them hierarchically. The blackboard is implemented in a general- purpose, typed, relational database system (built on top of INTERLJSP), called A83. AP3 has a pattern-matching language; this “A good discussion of scheduhng WI Hearsay-II csn be found in (51 “‘This notlon, in one form owemple, [6], (21. and [III another, is common to 8 number of others, for 108 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. is used for retrieval from the blackboard. AP3 also has demons; the triggering pattern which the application writer supplies as part of the definition of a KS is turned into an AP3 demon. The blackboard levels of Hearsay-II have been generalized somewhat into a tree-structure of classes. Each unit is created permanently as an instance of some class. The unit is, by inheritance, also an instance of all superclasses of that class. The apex of the class tree is the general class Unit. The immediate subclasses of Unit are DomainUnit and SchedulingUnit; these classes serve to define the domain and scheduling blackboards. All other subclasses are declared by the application writer, appropriate to his domain. For example, in the SAFE application, which is a system for building formal specifications of programs from informal specifications [l], one of the subclasses of DomainUnit is ControlFragment, and it has subclasses Sequence, Parallel, Loop, Conditional, Demon, etc. The semantics of the unit classes other than Unit, DomainUnit, and SchedulingUnit are left to the application writer. Any unit may serve to denote competing alternative interpretations. Such a unit, called a Choice Set, represents a choice point in the problem-solving. The Choice Set is a place- holder for the interpretation it represents; it can be dealt with as any other unit, including its incorporation as a component into higher-level units. Associated with a Choice Set unit are the alternatives of the choice. These may be explicit existing units or they may be implicit in a generator function associated with the Choice Set. When appropriate, a KS may execute a Select operation on a Choice Set, replacing it with the selected alternative. The Selection can be done in a destructive, irrevocable manner, or it can be done in a new contezt, retaining the ability to Select another alternative. Contexts are described more in Section 5. Hearsay-II’s location dimension (e.g., time-within-utterance in the speech-understanding domain) is not imposed on the Hearsay-III blackboard. The application writer may create such a dimension, either inherently in the interconnection structure of units or explicitly as values associated with the units. The flexibility of the underlying relational database system allows such constructs to have first-class status, for example, to be used in KS triggering patterns. 4. Scheduling Hearsay-Ill retains Hearsay-II’s basic sequencing of KS execution: When the triggering pattern of a KS is matched by a configuration of data on the blackboard, an activation record is created containing the information needed to execute the KS in the environment of the match. At some later time, the activation record may be selected and saecuted, i.e., the KS’s action, which is arbitrary code, is run. The executing KS has available to it the blackboard data that triggered it, which usually serves as the initial focus for the activity of the execution. Each KS execution is indivisible; it runs to completion and is not interrupted for the execution of any other KS activation. The effect of a KS execution is an updated blackboard. lndpendent activations of the same KS can pursue the same exploration by retrieving (potentially private) state information from the blackboard. The scheduling problem is: given the current state of the system, select the appropriate activation record to execute next. The separation of KS execution from triggering allows for complex schedultng schemes (i.e., a large collection of activations may be available from which to select). To allow the application writer to use the Hearsay problem-solving features for building such schemes, several mechanisms were added in Hearsay-Ill: - Each activation record is a unit on the scheduling blackboard. The application writer supplies, as part of the definition of each KS, code to be executed when the triggering pattern is matched; this code computes a scheduling-blackboard class (level) in which the activation record will be created. - When executed, scheduling KSs are expected to make changes to the scheduling blackboard to facilitate organizing the selection of activation records. In addition to triggering on changes to the domain blackboard, scheduling KSs can trigger on changes to the scheduling blackboard, including the creation of activation records. The actions a scheduling KS may take include associating information with activation records (e.g., assigning priorities) and creating new units that represent meta-information about the domain blackboard (e.g., pointers to the current highest-rated units on the domain blackboard). The scheduling blackboard is the database for the scheduling problem. - The application writer provides a base scheduler procedure that actually calls the primitive Eaecute operation for executing KS activations. We intend the base scheduler to be very simple; most of the knowledge about scheduling should be in the scheduling KSs. For example, if the scheduling KSs organize the activation records into a queue, the base scheduler need consist simply of a loop that removes the first element from the queue and calls for its execution. If the queue is ever empty, the base scheduler simply terminates, marking the end of system execution. 5. Context Mechanism While Choice Sets provide a means for representing an unmade decision about alternative interpretations, we still need a method of investigating those alternatives independently. For that, Hearsay-Ill supports a context mechanism similar to those found in Al programming languages such as QA4 [lo] and CONNIVER [9]. The method by which KS triggering interacts with the context mechanism allows controlled pursuit of alternative lines of reasoning. A KS triggers in the most general context (highest in the tree) in which its pattern matches. ixecution of that KS occurs in the same context and, unless it explicitly switches contexts, its changes are made in that context and are inherited down toward the leaves. Contexts are sometimes denoted as unsuitable for executing KSs -- a condition called poisoned. Poisoned contexts arise from the violation of a Hearsay constraint (e.g., attempting to aggregate conflicting units). In addition, a KS can explicitly poison a context if, for example, the KS discovers a violated domain constraint. A KS activation whose execution context is poisoned is placed in a wait state until the context is unpoisoned. Special KSs, called poison handlers, are allowed to run in poisoned contexts, and specifically serve to diagnose and correct the problems that gave rise to the poisoning. A common application for the context mechanism arises when alternative interpretations lack good “locality”. First consider the 109 exampfe of SAFE% Pfanning Phase, which uses Choice Sets to represent alternative interpretations for control fragments. In the case of the input sentence “Send an acknowledgment to the imp and pass the message on to the host.” a Choice Set served well. The possible interpretations for this sentence include being put in parallel or in sequence with an existing structure; since all alternatives would be positioned identically in the existing aggregate structure, the Choice Set unit can be placed where the chosen interpretation eventually will go. In some cases, however, locality is lacking. An example is the input sentence, “After receiving the message, the imp passes it to the host.” The possible interpretations for this include a demon (“The occurrance of r triggers y”) and a sequence to be embedded in an existing procedure (“After r do y”). Since the demon interpretation resides at the same structural level as the procedure into which the sequence would be embedded, there is no convenient place to put the Choice Set representing these alternatives. Instead, the KSs producing these alternative interpretations put them in brother contexts, so that each can be pursued independently. 6. Relational Database As mentioned earlier, the blackboard and all publicly accessible Hearsay-Ill data structures are represented in the AP3 relational database. In addition, any domain information which is to cause KS firing must also be represented in the database. This is because KSs are AP3 demons, and their triggering is controlled by activity in the database. The AP3 database is similar to those available in languages such as PLANNER [7], but also includes strong typing for each of the relational arguments in both assertion and retrieval. These typed relational capabilities are available for modeling directly the application domain. 7. Implementation and Current Status The Hearsay-Ill system is implemented in AP3, which in turn is implemented in INTERLISP [12]. AP3 was chosen as an implementation language because it already contained the mechanisms needed to support Hearsay-Ill (e.g., contexts, demons and constraints, and strong typing). In fact, the design of Hearsay-Ill’s initial implementation was almost trivial, being largely a set of AP3 usage conventions. However, efficiency considerations have forced a substantial implementation effort. Hearsay-Ill has been tested on two small applications: a cryptarithmatic problem and a cryptogram decoding problem. Three major implementation efforts are currently underway. The first of these, as described above, is the reimplementation of the SAFE system [l]. Second, Hearsay is being used as the basis for a system for producing natural language descriptions of expert= system data structures [8]. Finally, the system is being used as the basis for a “jitterer” which automatically transforms a program so tHat a transformation chosen by a user is applicable 143. The Hearsay-Ill architecture seems to be a helpful one. The separation of competence knowledge from p.erformance knowledge helps in rapidly formulating the expert knowledge required for a solution. Pretiminary experience with the larger applications now under development seem to bear this out, and seem to indicate that performance (scheduling) is a difficult issue. The flexibility that the Hearsay-III architecture gives toward developing scheduling algorithms will undoubtably go a long way toward simplifying this aspect of the overall problem-solving process. Acknowledgments We wish to thank Jeff Barnett, Mark Fox, and Bill Mann for their contributions to the Hearsay-Ill design. Neil Goldman has provided excellent and responsive support of the AP3 relational database system. Steve Fickas, Neil Goldman, Bill Mann, Jim Moore, and Dave Wile have served as helpful and patient initial users of the Hearsay-Ill system. References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. Balzer, R., N. Goldman, and D. Wile, “Informality in Program Specifications,” IEEE Trans. Software Enp. SE-4, (21, March 1978. Davis, R., Meta-Rules: Reasoning About Control, MIT Al Laboratory, Al Memo 576, March 1980. Erman, L. D., F. Hayes-Roth, V. R. Lesser, and D. R. Reddy, “The Hearsay-II Speech-Understanding System: Integrating Knowledge to Resolve Uncertainty,” Computing Surveys 12, (2), June 1980. (To appear) Fickas, S., “Automatic Goal-Directed Program Transformation,” in 1st NationaL Artificial Intelligence Conf., Palo Alto, CA, August 1980. (submitted) Hayes-Roth, F., and V. R. Lesser, “Focus of Attention in the Hearsay-II System,” in Proc. 5th International Joint Conference on Artificial Intefligence, pp. 27-35, Cambridge, MA, 1977. Hayes-Roth, B., and F. Hayes-Roth, Cognitive Prooesses in Phning, The Rand Corporation, Technical Report R-2366-ONR, 1979. Hewitt, C. E., Description and Theoretical Analysis (Using Schemata) of PLANNER: A Language for Proving Theorems and Manipulating Models in a Robot, MIT Al Laboratory, Technical Report TR-258, 1972. Mann, W. C., and J. A. Moore, Computer as Author -- Results and Prospects, USC/Information Sciences Institute, Technical Report RR-79-82, 1979. McDermott, D., and G. J. Sussman, The CONNIVER Reference Manual, MIT Al Laboratory, Memo 259a, 1974. Rulifson, J. F., R. J. Waldinger, and J. A, Derksen, “A Language for Writing Problem-Solving Programs,” in ZFIP 71, pp. 201-205, North-Holland, Amsterdam, 1972. Stefik, M., Planning with Constraints, Ph.D. thesis, Stanford University, Computer Science Department, January 1980. Teitelman, W., lnterlisp Reference ManuaL, Xerox Palo Alto Research Center, 1978. Waterman, D. A., and F. Hayes-Roth, Pattern-Directed fnference Systems, Academic Press, New York, 1978. 110	1980	67
64	QUANTIFYING AND SIMULATING THE BEHAVIOR OF KNOWLEDGE-BASED INTERPRETATION SYSTEMS* V.R. Lesser, S. Reed and J. Pavlin Computer and Information Science Department University of Massachusetts Amherst, Mass. 01003 ABSTRACT The beginnings of a methodology for quantifying the performance of knowledge-sources (KSs) and schedulers in a knowledge-based interpretation sys tern are presented. As part of this methodology, measures for the “reliability” of an intermediate state of system processing and the effectiveness of KSs and schedulers are developed. Based on the measures, techniques for simulating KSs and schedulers of arbitrary effectiveness are described. I INTRODUCTION The development and performance-tuning of a knowledge-based interpretation sys tern like the Hearsay-II speech understanding system [I] is still an art. There currently does not exist sufficient formal methodology for relating the performance characteristics of such a system to the performance characteristics of its components; i.e., knowledge-sources (KSs) and schedulers*. For that matter, there does not even exist an adequate framework for quantifying the performance of the system and its components in an uniform and integrated way. Thus, when the initial operational configuration, Cl, of the Hearsay-II speech understanding system had poor performance, there existed no methodology for detailing in a quantifiable way what types of performance improvements in specific components would be needed National -0412 and Projects of Naval The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or imp1 ied, of the National Science Foundation, the Defense Advanced Research Projects Agency, or the US Government. made some efforts in this we feel that his model of these abstract to capture adequately important issues. (A detailed be found in [41.) to improve significantly the overal 1 system performance . Therefore, the development of the completely reorganized C2 configuration, which turned out to have much superior performance, was based on “seat of the pants” intuitions. These intuit ions were testable only when the new set of KSs were integrated together into a working system. In the following sections we present the beginnings of a methodology for quantifying the performance of KSs and schedulers. We then show how this methodology can be used to simulate the performance of an upgraded component in a working system, so that more accurate estimates can be made of the overall performance improvement that would be realized if the component were actually upgraded. II A MODEL FOR A HEARSAY-LIKE ---v-p KNOWLEDGE-BASED SYSTEM -- 111 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. into more abstract and encompassing higher-level hypotheses (partial interpretations). The lower-level hypotheses are said to support the higher-level hypotheses. This aggregation process, accomplished by synthesis KSs , involves the detection of local consistency (or inconsistency) relationships among hypotheses. In KS processing, the belief-values of supporting hypotheses are not changed as a result of detection of local consistency, as would be the case in a relaxation process. The construction of a higher-level hypothesis and its associated belief-value is an explicit encoding of the nature and degree of consistency found among its supporting hypotheses. The hope is that this incremental aggregation process resolves the uncertainty among competing interpretations and, simultaneously, distinguishes between correct and incorrect interpretations. We have not discussed KSs as solely reducing the uncertainty in the system, because uncertainty is a measure of the distribution of belief-values and does not reflect the accuracy of hypotheses. We feel that KS processing causes changes in both certainty and accuracy in a system’s database and we have developed a measure, called ‘lreliability”, that combines the two. A good KS will produce higher-level hypotheses which are more reliable than the lower-level hypotheses supporting them. III MEASURING THE SYSTEM STATE Basic to our view of processing in knowledge-based systems is the concept of system state. The system state at any point in time is -- the current set of hypotheses and their relationships to the input data. It is through measures of the system state that we can talk in a uniform manner about the performance of KSs and schedulers. For purposes of measuring reliability, we associate a hidden attribute with each hypothesis which we call its truth-value. This attribute measures the closeness of the hypothesized event to the correct event. For task domains in which a solution is either totally correct or incorrect, we quantify truth-value as either 1 (true) or 0 (false) , while in domains in which there are solutions of varying degrees of acceptability, truth-values range between 1 and 0. One way of evaluating an intermediate state of processing is by measuring the reliability of the set of all possible complete interpretations (final answers> that are supported (at least in part) by hypotheses in the current state. We feel that a direct measure of this sort is not feasible because it is very difficult in general to relate the set of partial interpretations to the very large set of complete interpretations they can support. We take an alternative approach, called + To simplify this presentation, we focus here on synthesis KSs only, though prediction, verification, and extrapolation KSs also have a place in our model. reflecting-back, which is based on two premises. First, the creation of a high-level hypothesis is a result of detecting the constitency among its supporting hypotheses. This creation process iS an alternative to actually changing the belief-values of the supporting hypotheses, as occurs in the relaxation paradigm. Thus, the creation of a high-level hypothesis implicitly changes the reliability of its supporting hypo the se s . This change can be traced down to the input hypotheses whose reliability is implicitly improved to the extent that they are aggregated into reliable high-level hypo the se s . Second, we assume that processing which implicitly improves the reliability of the input hypotheses al so improves the reliability of the complete interpretations supported by these hypotheses. In the reflecting-back approach, we associate with each input hypothesis the highest belief hypothesis it supports. The truth- and belief-values of this highest belief hypothesis are reflected-back to the input hypothesis . The process is illustrated in Figure 1. It should be stressed that the hypotheses’ truth-values and reflected-back values are used only for measuring the system state; they are not available to KSs during processing. Our measure for the reliability of an intermediate system state is based on a measure of the reliability of input competitor sets computed from the reflected-back belief- and truth-values. Intuitively, a measure of reliability for a competitor set should have the following properties, based on both the be1 ief- and truth-values of its hypotheses: 1. With respect to accuracy, reliability should be high if a true hypothesis has high belief-value or if a false hypothesis has a low belief-value, while reliability should be low if a true hypothesis has low belief-value or a false hypothesis has high belief-value; 2. With respect to uncertainty, reliability should be high if one hypothesis in a competitor set has a high belief-value and the rest have low belief-values, while reliability should be low if all the hypotheses have similar belief-values. A measure for the reliability, RC(S), of a competitor set, S, that captures and adequately combines both of these properties is: RC(S) = 1 - avg iTV(h)-BV(h) I h in S which is equivalent, in the case of binary truth-values, to the correllation of truth- and belief-values: RC( s) = avg CBV(h)TV(h) + (l-BV(h) )(I-TV(h))1 h in S where BV(h) is the belief-value of hypothesis h and 112 TV(h) is the truth-value of hypothesis h. Other measures may also be applicable, but of those we considered this one best captures our intuitive notion of reliability. Based on this measure of competitor set reliability, we can construe t, for instance, a measure of processing effectiveness associated with the current intermediate sys tern state. This measure is the average over the input competitor sets of the difference between the initial reliability and the reliability of the intermediate state. The initial reliability is the average reliability of the competitor sets formed from input hypotheses, using the initial belief- and truth-values. The reliability of the intermediate -- state is the average reliability of the competitor sets formed from the input hypotheses, where the reflected-back truth- and belief-values of hypotheses are used in place of the original ones. In Figure 1, if the three competitor sets were the only ones on the input level and each one contained just the two hypotheses shown, then the initial rel iab il i ty , RI, is .517, the reliability of the intermediate state, RS, is .575, and processing effectiveness, PE, is .058: RI = l/3 (.7+.6)/2 + (.5+.4)/2 + (.7+.2)/21 = .517 RS q 1/3[(.6+.5)/2 + (.7+.5)/2 + (.45+.7)/21 = .575 PE q RS - RI = .058 A positive PE value, as in this example, indicates that some uncertainty and error has been resolved. The larger the value, the more resolution, the more effective is the processing. IV MEASURING THE RESOLVING POWER OF A KS ___-- - We define the instantaneous resolving power of a KS as a change in reliability due to a single KS - execution. This change is measured on competitor sets construe ted from the KS input hypotheses. Thus, instead of calculating the reflected-back reliability of the entire system state, the procedure is localized only to the subset of the state directly affected by the KS execution. We Figure 1: an example of the reflecting back process. INPUT LEVEL: COMPETITOR SETS The system state contains three competitor sets (represented by boxes) on the input level, and a number of higher level hypotheses. The hypotheses are represented as circles with belief-values on the left and truth-values CT=1 and F=O) on the right. The initial values are in the bottom of the circle, the reflected-back values are in the top half. The thick 1 ines show the support 1 ink to the highest belief-value hypothesis supported by each input hypothesis and indicate the source of the reflected-back values. Hypotheses which intervene between the inputs and their highest-belief supported hypotheses are not shown. 113 measure the change in reliablilty as a result of KS execution by measuring before KS processing the reliability of the KS input competitor sets, then measuring after KS processing the reliability of these sets based on values reflected-back from the KS output hypotheses, and finally taking the difference between the results obtained. The resolving power of a KS can now be defined ---- as its average instantaneous resolving power over a series of system executions. This view of KS resolving power does not take into account the global impact of KS execution on the entire system state and on future processing. Rather, it is a local, instantaneous measure of the effects of KS execution. The global effects occur through KS interactions, which we believe should be separated from our measure of the resolving power of a single KS. V SIMULATING A KS -- Given a formal measure of KS resolving power, we can simulate KSs of any desired power. This is accomplished by introducing an lloracle" which knows how to judge the closeness of a hypothesized interpretation to the correct interpretation (this is the source of the truth-values). Our reliability measures can thus be calculated during processing rather than in retrospect, after the system has completed processing. Therefore, a system does not hav% to complete execution in order to be evaluated. A KS is simulated in stages: candidate We believe that our approach to simulating KSs of different resolving power, which makes heavy use of an oracle, will prove useful in designing and debugging knowledge-based systems. However, there are some limitations: ------------------ I In many cases, it is relatively easy to design a KS which provides moderate accuracy. Most of the effort in knowledge engineering is spent in increasing this accuracy to gain superior performance. Our simulation of KS resolving power is based on a combination of simple knowledge about local consistency and reference to an oracle, while real KSs infer truth from local consistency alone (and falsehood from local inconsistency). The behavior of different simulated KSs sharing similar errors in knowledge will not be correlated due to our statistical approach to KS simulation. Given these limitations, we do not expect a simulated KS to behave exactly the same as a real KS. We hope, however, the essential behavior of a KS has been captured so that system phenomena are adequately modelled. In order to validate our models of KS power, we plan to analyze the behavior of KSs in some existing knowledge-based systems. A measures of KS power will be taken for an existing KS and then the KS will be replaced by a simulated KS of the same power, and the overall system behavior compared in the two cases. The results of these experiments should give us some understanding of the extent to which data derived from our simulation studies can be used to predict the behavior of real systems. VI SIMULATION OF - ACCURACY IN THE -- SCHEDULER Reliability measures can also be used in the simulation of a scheduler of a specific accuracy. The task of a scheduler is choosing a KS instantiation for execution. A KS instantiation is a KS-stimulus pair, where the stimulus is the set of hypotheses which caused the KS to be considered for scheduling. The scheduler evaluates alternative instantiations according to its knowledge of the characteristics of the KSs, the stimuli, and the current state of processing. The effects of future processing are not factored into this model of scheduling; we take an instantaneous view of scheduling decisions. Because of this, we are unable to model scheduling algorithms such as the "shortfall density scoring method" 171 which use information about future processing. We hope to develop a formulation that includes this type of information. A good scheduler chooses for execution the KS instantiation that will most improve the reliability of the current system state. The accuracy of a single scheduling decision is defined --- relative to the performance of an optimum scheduler, which uses accurate information about the resolving power of the KSs and the reliability of the KS stimuli and system state. The accuracy of a scheduler is the average of the accuracy of -- many scheduling decisions. We steps: view the optimum scheduling process in ------------------ The work of Paxton C5l comes the closest our approach, but was much more limited. two to 114 1. For each KS instantiation on the scheduling queue, make accurate predictions concerning its instantaneous resolving power. These predictions involve determining the truth-value of the stimulus hypotheses (using the oracle) and knowledge of the resolving power of the KS. 2. Make accurate predictions as to the global system state which would result from scheduling each instantiation given the predictions of step 1. These predictions will determine optimum ratings for the instantiations and result in an optimum schedule. Our approach to modelling the scheduler is to obtain statistically accurate ratings for the instantiations, based on the optimum schedule, and then choose for execution an instantiation from within the ordering which results. The position in the ordering of the chosen instantiation depends on the desired accuracy of the scheduler being modelled; the closer to the top of the order, the more accurate the scheduler. We feel it would be an error to model scheduling only as a function of the truth-value of stimulus hypotheses. Real schedulers do not have access to the truth-values of hypotheses, but only infer truth from belief-values and processing history. The point is that two instantiations of the same KS, whose stimulus hypotheses have equivalent characteristics (same belief-value, level of abstraction, database region, processing history, etc.) except for their truth-values would be rated the same by even the best scheduler. Additionally, in order to determine the rating of a KS instantiation, real schedulers [ 31 consider other factors, besides the characteristics of the stimulus hypotheses. For example, schedulers take into account such factors as the balance between depth-first vs. breadth-first processing or between executing KSs that work in areas with rich processing history vs. executing KSs that work where little processing has been done. These additional considerations are, in fact, heuristics which attempt to capture the concept of improvement in the reliability of the system state. Thus, in our view, a scheduler should be characterized in terms of its ability to estimate the improvement in system state reliability, rather than its ability to detect the truthfulness of the instantiation's stimulus hypotheses. We could have modelled the scheduler just as we modelled KSs, with a candidate evaluator and a scheduling resolver. The candidate evaluator would take the generated KS instantiations and give them ratings based on simple scheduling knowledge. The scheduling resolver would minimally alter these ratings (with statistical perturbation) to produce an ordering for the instantiations which corresponds to a desired scheduler accuracy. For several reasons, too complicated to discuss in this short paper, we have not used such an approach for modelling schedulers. Further details of this issue and a more detailed formulation of scheduling measures ar e disc this paper c41. in an extended version of VII SUMMARY This work represents the beginnings of a methodology for understanding in quantitative terms the relationship between performance of a knowledge-based system and the characteristics of its ccmponents. This quantification may also allow us to develop simulations of these systems which can accurately predict the performance of alternative designs. ACKNOWLEDGMENTS We would like to recognize the helpful comments on various drafts of this paper given by Daniel Corkill and Lee Erman. REFERENCES II11 c21 c31 r41 [51 [61 c71 Erman, L. D., F. Hayes-Roth, V. R. Lesser and R. Reddy (1980). "The Hearsay-II Speech Understanding System: Integrating Knowledge to Resolve Uncertainty," Computing Surveys, 12:2, June 1980. Fox, M. S. (1979). "Organizational Structuring: Designing Large Complex Software," Technical Report CMU-CS-79-155, Department of Computer Science, Carnegie-Mellon University, Pittsburgh, Pennsylvania. Hayes-Roth, F. and V. R. Lesser (1977>, "FOCUS of Attention in the Hearsay-II Speech Understanding System," In Proceedings of the Fifth International Joint Conference on Artificial Intelligence-1977, P. 27-35, Cambridge, Massachusetts, 1977. Lesser, V. R., J. Pavlin and S. Reed (19801, "First Steps Towards Quantifying the Behavior of Knowledge-Based Systems," Technical Report, Department of Computer and Information Science, University of Massachusetts, Amherst, Massachusetts. Paxton, W. H. (19781, I1 The Executive System," in D. E. Walker, (editor), Understanding Spoken Language, Elsevier, North-Holland, N. Y., 1978. Rosenfeld, A. R. , R. A. Hummel , and s. w. Zucker (19761, "Scene Lab eling by Rel axation Operators," IEEE Transactions on Systems, Man and Cybernetics, SMC-6, pg. 420-433, 1976. Woods, W.A. (1977)) "Shortfall and Density Scoring Strategies for Speech Understanding Control,11 In Proceedings of the Fifth -- International Joint Conference on Artificial Intelligence-1977, p. 18-26, - Cambridge, Massachusetts, 1977. 115	1980	68
65	Representation of Task-Specific Knowledge in a Gracefully Interacting User Interface Eugene Ball and Phil Hayes Computer Science Department, Carnegie-Mellon University Pittsburgh, PA 15213, USA Abstract Command interfaces to current interactive systems often appear inflexible and unfriendly to casual and expert users alike.’ We are constructing an interface that will behave more cooperatively (by correcting spelling and grammatical errors, asking the user to resolve ambiguities in subparts of commands, etc.). Given that present-day interfaces often absorb a major portion of implementation effort, such a gracefully interacting interface can only be practical if it is independent of the specific tool or functional subsystem with which it is used. Our interface is tool-independent in the sense that all its information about a particular tool is expressed in a declarative tool description. This tool description contains schemas for each operation that the tool can perform, and for each kind of object known to the system. The operation schemas describe tne relevant parameters, their types and defaults, and the object schemas give corresponding structural descriptions in terms of defining and derived subcomponents. The schemas also include input syntax, display formats, and explanatory text. We discuss how these schemas can be used by the tool-independent interface to provide a graceful interface to the tool they describe, 1. Introduction Command interfaces to most current interactive computer systems tend to be inflexible and unfriendly. If the user of such a system issues a command with a trivial (to a human) syntactic error, he is likely to receive an uninformative error message, and must re-enter the entire command. The system is incapable of correcting the error in the “obvious” way, or of asking him to retype only the erroneous segment, or of providing an explanation of what the correct syntax really is. Anyone who has used an interactive computing system is only too familiar with such situations, and knows well how frustrating and time-consuming they are, for expert as well as novice users. We are involved in a project to build an interface which will behave in a more flexible and friendly way, one that will inferact graceful/y. As we have described in earlier work [3, 41, graceful interaction involves a number of relatively independent skills including: m the parsing of ungrammatical input, either to correct it or to recognize any grammatical substrings; 1 This research was sponsored by the Defense Advanced Research Projects Agency (DOD), ARPA Order No. 3597, monitored by the Air Force Avionics Laboratory Under Contract F33615-78-C-1551. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed OI Implied, of the Defense Advanced Research Projects Agency or the US Government. o robust communication techniques to ensure that any assumptions the system makes about the user’s intentions are implicitly or explicitly confirmed by the user; o the ability to give explanations of how to use the system or the system’s current state; e interacting to resolve ambiguities or contradictions in the user’s specification of objects known to the system; o keeping track of the user’s focus of attention; e describing system objects in terms appropriate to the current dialogue context. Providing these facilities is clearly a major programming task requiring extensive use of Artificial Intelligence techniques (see [2] for just the flexible parsing aspect). We believe that it is unrealistic to expect the designers of each interactive sub-system (or tool) to implement a user interface with these capabilities. Therefore, instead of constructing a gracefully interacting interface for a single application, we are attempting to build a tool independent system, which can serve as the user interface for a variety of functional sub-systems. The availability of a tool independent user interface would greatly simplify the construction of new computer sub-systems. Currently, even if the new system is not intended to be gracefully interacting but merely to perform according to minimal standards, a large amount of implementatron effort must be devoted to user interface issues. The system designers must decide on a style of interaction with the user, select the general format and detailed syntax of all commands, and provide for the de.tection of illegal input. The command language must then be thoroughly checked to ensure that it does not contain ambiguities or misleading constructions and that likely error sequences will not be misinterpreted and cause unrecoverable system actions. Often, the design can only be completed after an Initial implementation of the system has produced feedback about the usability of the human interface. This design process represents the minimum effort necessary to produce a system that is even usable by a large number of people; if a superior (but still far from gracefully interacting) interface or one which can be used by non-programmers is required. much more work must be expended. Editing facilities, which are required in most interactive systems (at least for correction of typed input), must be fully integrated into the sub-system; compatibility with other editors in common use on the computer must be considered, even though this may lead to difficult interactions with the sub-system command language. Error detection and reporting must be Improved; generating coherent diagnostics for the inexperienced user can be very difficult indeed. Online documentation must be provided, 116 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. including reasonable facilities which allow a user to quickly find the answer to specific (although vaguely expressed) questions. The complexity of this task often means that most of the implementation effort in adding a new tool to a computer system is absorbed by the user interface. Technological trends are aggravating the problem by raising the level of performance expected of an interface. In particular, as high-resolution graphics displays equipped with pointing devices become available, users expect to use menu-selection and other more sophisticated forms of input, and to see system output displayed in an attractive graphical format. The very recent, but growing, availability of speech input and output will intensify this pressure for sophistication. An additional reason for constructing a tool-independent interface is to make the computer system as a whole appear consistent to the user. If the interfaces for different tools use different conventions, then no matter how sophisticated each of them is individually. the user is likely to be confused as he moves from one to another because the expectations raised by one may not be filled by the other. For all these reasons, we are attempting to make our gracefully interacting interface as tool-independent as possible. In the remainder of this paper we outline the system structure we have developed, and go on to give further details about one component of this structure, the declarative format in which information about the tool is made available to the interface, together with sketches of how the tool-independent part of the system uses the information thus represented. 2. System Structure The basis for our system structure is the requirement that the interface contain no tool-dependent information. All such information must be contained in a declarative data base called the tool description. In an effort further to improve portability and reduce duplication of effort between interfaces implemented on different hardware configurations, we have made a second major separation between the device-dependent and -independent parts of the interface. The resulting structure is illustrated in figure 1. User Agent Figure 1. User Interface System Structure The intelligent functions of the interface, those itemized above, are isolated in a tool and device independent User Agenf, which interacts with the tool through a narrow interface that is completely specified by the declarative tool description. Communication between the Agent and the user is not direct, but goes via a device-dependent Front-End, which allows the Agent to specify its output in a high-level device-Independent manner, and which preprocesses the user’s input into a standard, device-independent, format. Communication between the Agent and Front-End is thus restricted to a well-defined format of input and output requests. Display formats in which to realize the tool’s and Agent’s high-level output requests are specified declaratively in the tool description. The basic function of the Agent is to establish from the user’s input what functional capability of the tool the user wishes to invoke and with what parameters he wishes to invoke it. Once this is established, the Agent issues the appropriate request to the tool and reports to the user relevant portions of the tool’s response. To make this possible, the tool description includes a specification of all the operations provided by the tool in terms of their parameters and their types, defaults, etc., plus specifications of all the abstract objects manipulated by the tool in terms of their defining (and descriptive) sub-components. This representation of operations and objects follows Minsky’s frames paradigm [7] in the spirit of KRL [l] or FRL [13]. The representation allows the Agent to follow the user’s focus of attention down to arbitrarily deeply nested aspects of object or operation descriptions to resolve ambiguities or contradictions. This facility depends on the tool to provide resolution of object descriptions into sets of referents. The tool description also specifies the syntax for the user’s input descriptions of the objects and operations. The Agent applies the grammar thus specified to the user’s input (as pre-processed by the Front-End) in a flexible way, providing the kinds of flexible parsing facilities mentioned above. The user may also request information about the tool or other help, and the Agent will attempt to answer the query with information extracted from the tool description, and displayed according to tool-independent rules. Besides requesting that the Front-End output text strings to the user, the Agent may also specify instances of system objects. The Front-End will then drsplay the objects according to a display format specified in the tool description. For the Front-End we are using, whrch operates through a graphics display equipped with a pointing device. this allows the user to refer directly to system objects by pointing. The Front-End reports such pointing events to the Agent in terms of the system object refered to. Typed input is pre-processed into a stream of lexical items, and other pointing events, such as to menus, can also be reported as lexical items. We are also experimenting with a limited-vocabulary, single word (or phrase) speech recognizer, isolated in the Front-End. It’s output can also be reported as a lexical item. This concludes the overview of the system structure. For the remainder of the paper, we will concentrate on the representation employed in the tool-description, and the way the information, thus represented, is used by the remainder of the system. Our examples will be in terms of the tool being used as a test-bed for the development of the Agent and Front-End: a multi-media message system, capable of transmitting, receiving, filing, and retrieving pieces of electronic mail whose bodies contain mixtures of text, speech, graphics, and fax. 3. Representation of task specific information A functional sub-system, or tool, is characterized for the user interface program by a data base which describes the objects it manipulates and the operations it can perform. This fool description is a static information structure (provided by the sub-system implementor) which specrfies everything that lhe User 117 Agent needs to know about the tool. We’ll first give a brief overview of the structure of the tool description and how the Agent uses it to provide an interface to the sub-system. Then the content of the data base will be explained in more detail, with examples showing how this information is utilized in the processing of commands from the human user. The tool description consists of: o Declarations of the data objects used by the tool. These declarations specify the internal structure of each object type defined within a particular sub-system. The tool may also contain references to object types that are defined in a global data base and are used by many different tools (e.g. files, user names, dates and times). The object declaration provides rules for displaying an instance of the object, syntax for descriptions of it in commands, and documentation that can be used to explain its function to the user. o Descriptions of the operations which the tool can perform. Each operation entry specifies the parameters that the Agent must provide to the tool to invoke that action. It also defines the legal syntax for the command, provides some simple measures of its cost and reversability, and supplies a text explanation of its purpose. As mentioned earlier, the primary goal of the Agent is to help the human user to specify sub-system operations to be executed. To carry out this function, it parses the user’s commands (including text, pointing, and possibly spoken input) according to the syntax specifications in the tool description. It decides which operation has been selected and attempts to fill out the parameter template associated with it. This process may involve interpreting descriptions of sub-system objects, negotiating with the user about errors or ambiguities that are discovered, and explaining the meaning of command options, 3.1. Object Descriptions The tool description contains a declaration of each data type that is defined within that sub-system. Data objects which will be manipulated by both the Agent and the tool are represented as lists and property lists (sets of name-value pairs), using the formalism defined by Postel for the communication of Internet messages [9]. This representation is self-describing in that the structure and type of each data element is represented explicity in the object. Thus, complexly structured objects can be transferred between the User Agent and the tool, and the Agent can interpret them according to the information contained in the tool description. For example, the following is the internal representation of a simple message (primitive elements are integers or text strings, and brackets are used to delimit sets of name-value pairs): [ StructureType: Objectlnstance ObjectName: Message Sender: [ PersonName: [ First: John Middle: Eugene Last: Bali] Host: [ Site: CMU Machine: A ] 1 Recipient: [ PersonName: [ First: Phil Last: Hayes ] Host: [ Site: CMU Machine: A ] 1 Copies: [] Date: [ Year:1980 Month:April Day:10 Weekday:Thursday AMPM: AM Hour: 11 Minutes: 16 Seconds: 371 Subject: “Meeting tomorrow?” Body: “Phil, Could we meet tommorrow at 1 pm? -Gene” 1 The structure of a Message is defined in the tool data base by a message schema. This schema declares each legal field in the object and its type, which may be a primitive type like TEXT or an object type defined by another schema in the tool description. The schema also specifies the number of values that each field may have, and may declare default values for new instances of the object. The following is a simplified schema for a message object and some of its components: StructureType: ObjectSchema ObjectName: Message DescriptionEvaluation: ToolKnows Schema: Sender: [ FillerType: Mailbox ] Recipient: [ FillerType: Mailbox Number: OneOrMore ] Copies: [ FillerType: Mailbox Number: NoneOrMore ] Date: [ FillerType: Date ] Subject: [ FillerType: MultiMediaDocument ] Body: [ FillerType: MultiMediaDocument ] After: [ FillerType: Date UseAs: DescriptionOnly ] Before: [ FillerType: Date UseAs: DescriptionOnly ] StructureType: ObjectSchema ObjectName: Mailbox DescriptionEvaluation: ToolKnows Schema: [ PersonName: [ FillerType: PersonName ] Host: [ FillerType: Host ] 11 StructureType: ObjectSchema ObjectName: PersonName DescriptionEvaluation: OpenEnded Schema: [ First: [ FillerType: TEXT ] Middle: [ FillerType: TEXT Number: NoneOrMore] Last: [ FillerType: TEXT ] II StructureType: ObjectSchema ObjectName: Host DescriptionEvaluation: ToolKnows Schema: [ Site: [ FillerType: TEXT Default: CMU ] Machine: [ FillerType: TEXT Default: A ] 11 In addition to defining the structure of an instance of a message object, the schema includes fields which are used by the Agent to interpret descriptions of messages. The DescriptionEvaluation field tells the Agent how to evaluate a description of an object in this class. For example, ToolKnows indicates that the sub-system is prepared to evaluate a description structure and return a list of instances matching the description. A description structure is an instance of the object with special wild card values for some fields and with possible extra DescriptionOnly entries, such as the After field in the example above. Since a description of one object may reference other objects (“the messages from members of the GI project”), the Agent uses the hierarchy defined by the object declarations to guide its evaluation of the user’s commands. Each level generates a new sub-task in the Agent which processes that portion of the description, and is responsible for resolving ambiguities that it may encounter. This structure also makes it possible to follow the user’s focus of attention, since new input may apply to any of the currently active subgoals (“No, only the ones since October” or “No, only the ones at ISI”). 118 Each object declaration also includes information which is used information network at a node representing the data type, which is by the Front-End module to display instances of that object type. connected to Zog frames for other objects referenced by that type. Several different formats may be defined; the tool (or Agent) The user can find information about a related sub-system object selects an appropriate format by name each time it displays an by choosing a link to follow (with a menu selection); that frame is object. The format declaration specifies which fields of the object quickly displayed. The legal syntax for descriptions of the object to display and their layout; it may also provide parameters to the and links to frames for operations which manipulate it are also Front-End which invoke special capabilities (highlighting, font included in the Zog network. In the following example selection) of the display hardware. For example, the following documentation frame for Message, the italicized entries are section of the description for a Message defines two different buttons which can be activated to request more information about styles of message header display. a specific topic: DisplayFormat: [ ShortHeader: // style: From Hayes on 18-Mar [ Text: “From &Sndr& on &Day&” Sndr: [Field: Sender/PersonName/Last FaceCode: Bold] Day: [ Field: Date Style: ShortDay ] 1 FullHeader: // From Eugene Ball on lo-Apr-80 11:16am about ‘Meeting tommorrow? [ Text: “From &Sndr& on &Day& about ‘&Subj&“’ Sndr: [ Field: SenderIPersonName Style: FullName ] Day: [ Field: Date Style: FullDate ] Subj: [ Field: Subject Style: FullSubj ] 11 Messase: (Multi-Media Message System) Each message contains the author and date of origination and is addressed to (one or more) destination mailboxes; copies may optionally be sent to additional destinations. The body of a message may contain uninterpreted text, images, sketches, and voice recordings. Example syntax: ‘messages [from Person] [before/after Date] [about String]’ Detailed syntax Each object declaration also defines the legal syntax that can be used in commands that refer to objects of that type. In order to understand a phrase like “the messages from Phil since March”, the Agent uses the syntax definition associated with the object type Message, which in turn refers to the syntax for other objects like Date and PersonName. In the example below, question marks indicate optronal syntactic elements, asterisks mark fields that can be repeated, slashes indicate word class identifiers, and ampersands mark places where the syntax for other object types is to be expanded. Some syntactic entries also specify correspondences to particular fields in the object; as a command is parsed, the Agent builds description structures which represent the objects referred to in the command. Thus, the phrase “since March” results in an appropriate After clause in the Message description. The grammar defined by these syntax entries is applied to the user’s input in a flexible way [2], so that grammatical deviations such as misspellings, words run together, fragmentary input, etc. can still be parsed correctly. Related object types: Mailbox Display Multi Media Document Operations: Send Date Edit 3.2. Operation Descriptions Each sub-system operation which can be invoked by the Agent is also described by an entry in the tool data base. An operation entry specifies the parameters that the Agent must provide to the tool to have it perform that action. The object type of each parameter is declared and the tool description can optionally indicate that a parameter position may be filled by a set of such objects. In addition, constraints on the legal values of a parameter are sometimes provided, which can help the Agent to avoid requesting an illegal operation. Syntax: [ Pattern: (?/Determiner /MessageHead */MessageCase) Determiner: (the (all ?of ?the) every) MessageHead: (messages notes letters mail) MessageCase: [ StructureType: Operation OperationName: Forward Reversible: false Cost: moderate Parameters: [ Message: [ FillerType: Message Number: OneOrMore ] Recipient: [ FillerType: Mailbox Number: OneOrMore ] Forwarder: [ FillerType: Mailbox MustBe: CurrentUser ] ( [ Syntax: (/From &Mailbox) StructureToAdd: [ Sender: &Mailbox ]] [ Syntax: (/Before &Date) StructureToAdd: [ Before: &Date]] [Syntax: (/After &Date) 1 Syntax: [ Pattern: (/Forward %Message to %Recipient) Forward: (forward send mail (pass on) deliver redeliver) 1 Explanation: “Message Forwarding StructureToAdd: [ After: &Date ]] 1 From: (from (arriving from) (that came from) (/Mailed by)) Mailed: (mailed sent delivered) Before: (before (dated before) (that arrived before)) After: (after since (dated after)) 1 A copy of a message that was delivered to you can be sent to another person with the Forward command. You must specify the message to forward and the destination mailbox. Sample syntax: ‘Forward the last message from Phil to Adams at ISIE”’ Finally, the object description provides information which is used to automatically construct documentation and provide answers to user requests for help. Each object contains a brief text explanation of its structure and purpose in the sub-system, which can be presented to the user in response to a request for information. The documentation is also placed into a Zog [14] The example entry for the forward operation also mcludes a declaration of the legal syntax for the command, and a text entry which will be included in its documentation frame. It also indicates that this command is not reversible (once executed it cannot be undone), and that it is moderately expensive to execute. This information is used by the Agent to select an appropriate style of interaction with the User; for example, irreversible operations will usually require explicit confirmation before the request is given to the sub-system for execution. 119 4. Conclusion References The design and implementation of a good user interface for a computer sub-system is a difficult and time-consuming task; as new techniques for communication with computers (especially high resolution displays and speech) gain widespread use, we expect this task to become even more expensive. However, we also feel that the typical user interface must be made much more robust, graceful, and intelligent. For this goal to be feasible, substantial portions of the interaction with the user must be independent of the details of the application, so that the development cost of the user interface code can be shared by many sub-systems. Therefore, we are designing a generalized User Agent which can be used to control a variety of different sub-systems. The Agent carries on a dialog with the human user; it can understand a variety of different command styles, recognize and correct minor syntactic or spelling errors, supply default values for command arguments based on context, and provide explanations when requested. All of the information that the Agent needs to know about the application system is explicitly stored in a tool description provided by the sub-system implementor. This paper 1. Bobrow, D. G. and Winograd, T. “An Overview of KRL-0, a Knowl&ge Representation Language.” Cognitive Science 1,1 (1977). 2. Hayes, P. J. and Mouradian, G. V. Flexible Parsing. PrOC. of 18th Annual Meeting of the Assoc. for Comput. Ling., Philadelphia, June, 1980. 3. Hayes, P. J., and Reddy, R. Graceful Interaction in Man-Machine Communication. Proc. Sixth Int. Jt. Conf. on Artificial Intelligence, Tokyo, 1979, pp. 372-374. 4- Hayes, P. J., and Reddy, R. An Anatomy of Graceful Interaction in Man-Machine Communication. Tech. report, Computer Science Department, Carnegie-Mellon University, 1979. 5. Kernighan, Brian W. and Ritchie, Dennis M. The C Programming Language. Prentice-Hall, Inc., 1978. 6. Metcalf, Robert and Boggs, David. “Ethernet: Distributed has concentrated on the content of that data base, detailing the information represented there and demonstrating how the Agent can apply it to provide a sophisticated interface to a specific Packet Switching for Local Computer Networks.” Comm. ACM 79, 7 (July 1976), 395404. application system. 7. Minsky, M. A Framework for Representing Knowledge. In The tool description is represented in a unified formalism, which enables us to maintain a single data base which specifies all of the task-specific attributes of a particular sub-system. Because the information is stored in a single format, it can easily be utilized by multiple portions of the interface system. For example, a single syntax description is used to parse user commands, to generate Winston, P., Ed., The Psychology of Computer-Vision, McGraw Hill, 1975, pp. 211-277. 8. Newell, A., Fahlman, S., and Sproull, R.F. Proposal for a joint effort in personal scientific computing. Tech. Rept. , Computer Science Department, Carnegie-Mellon University, August, 1979. explanations of system actions, and to construct documentation Of the options available in the tool. 9. Postel, J. Internet Message Protocol. Draft Internet Experiment Note, Information Sciences Institute, Univ. of The initial implementation of the Agent will provide the user Southern California, April, 1980. interface for the Multi-Media Message System; an electronic mail facility which manipulates messages containing mixtures of text, lo- Rashid, R. A proposed DARPA standard inter-process recorded speech, graphics, and images. The system is being communication facility for UNIX version seven. Tech. Rept. , implemented as a multiple machine, multiple language distributed Computer Science Department, Carnegie-Mellon University, system: screen management (multiple windows) and graphics February, 1980. support are provided in Bcpl [l I] on a Xerox Alto [15] with a high resolution raster display and pointing device: audio recording and 1 1 - Richards, M. BCPL: A tool for compiler writing and systems playback is controlled by a DEC PDP-11 in C [5]; the Front-End programming. Proceedings of the Spring Joint Computer module and User Agent are implemented in C and LISP Conference, AFIPS, May, 1969, pp. 34557-566. respectively, on a VAX-l l/780 running Unix [12]; and the tool (message system) runs in C on the VAX. The system modules 12. Ritchie, D. M. and Thompson, K. “The UNIX Time-Sharing communicate using a message based Inter-Process System.“ Comm. ACM 77, 7 (July 1974), 365375. Communication facility [lo] within Unix, and a packet broadcast network (Xerox Ethernet [6]) between machines. Most of the 13. Roberts, R. B. and Goldstein, I. P. The FRL Manual. A. I. system components are currently running as individual modules, the first version of a single integrated system should be completed Memo 409, MIT Al Lab, Cambridge, Mass., 1977. by June 1980. Because of our goal of a smoothly working, robust, 14. and graceful system, we expect to continue tuning and improving Robertson, G., Newell, A., and Ramakrishna, K. ZOG: A the implementation for at least another year. The system will Man-Machine Communication Philosophy. Tech. Rept. , eventually be moved to a single powerful personal computer, Carnegie-Mellon University Computer Science Department, where we expect it to make substantial contributions to the CMU August, 1977. Spice (Scientifc Personal Integrated Computing Environment [8]) development effort. 1 6 . Thacker, C.P., McCreight, E.M., Lampson, B.W., Sproull, R.F., and Boggs, D.R. Alto: A personal computer. In Computer Structures: Readings and Examples, McGraw-Hill, 1980. Edited by D. Siewiorek, C.G. Bell, and A. Newell, second edition, in press. 120	1980	69
66	AN EFFICIENT RELEVANCE CRITERION FOR MECHANICAL THEOREM PROVING* David A. Plaisted Department of Computer Science University of Illinois Urbana, Illinois 61801 ABSTRACT To solve problems in the presence of large knowledge bases, it is important to be able to de- cide which knowledge is relevant to the problem at hand. This issue is discussed in [l 1. We pre- sent efficient algorithms for selecting a relevant subset of knowledge. These algorithms are presene ed in terms of resolution theorem proving in the first-order predicate calculus, but the concepts are sufficiently general to apply to other logics and other inference rules as well. These ideas should be particularly important when there are tens or hundreds of thousands of input clauses. We also present a complete theorem proving strate- gy which selects at each step the resolvents that appear most relevant. Thisstrategy is compatible with arbitrary conventional strategies such as P - deduction, locking resolution, et cetera. Also, 1 this strategy uses nontrivial semantic information and "associations" between facts in a to human problem-solving processes. I RELEVANCE FUNCTIONS Definition. A support set for a set S of clauses is a subset Ti of S such that S-Ti is con- way similar sistent. A support class for S is a set {Tl,..., Tk] of support sets for S. Definition. A (resolution) proof of C from S is a sequence C1,C2 ,...,C, of clauses in which Cn is C and each clause C i is either an element of S (an input clause) or a resolvent of two preceding clauses in S. (Possibly both parents of Ci are identical.) The length of such a refutation is n. A refutation from S is a proof of NIL (the empty clause) from S. Definition. A relevance function is a func- tion R which, given a set S of clauses, a support class T for S, and an integer n, maps onto a sub- set Rn(S, T) of S having the following property: If there is a length n refutation from S, then there is a refutation from Rn(S, T) of length n or less. Thus if we are searching for length n refutations from S, we need only search for length n refuta- * This research was partially supported by the Na- tional Science Foundation under grant MCS-79-04897. tions from R,(S, T). In fact, the derivation from R,(S, T) will be a subderivation of the derivation from S, for all relevance functions considered here. Thus if there is a length n P1-deduction from S, there will be a P l-deduction of length n or less from R,(S, T), and similarly for other corn plete strategies. Definition. Suppose S is a set of clauses. The connection graph of S, denoted G(S), is the graph whose nodes are the clauses of S, and which has a directed arc from C to C 1 2 labeled (Ll, L2)' if there are literals L 1 c C 1 and L2 E C2 such that Ll and z2 are unifiable. Such graphs have been introduced and discussed in [2]. Note that there will also be an arc labeled (L 2' Ll> from C 2 to Cl in the above case. Definition. A path from Cl to Cn in G(S) is a sequence Cl,C2,..., Cn of clauses of S such that there is an arc from C i to c i+l in G(S), for lli<n. Also, the length of the path is n. Definition. The distance d(Cl, C2) between Cl and C2 in G(S) is the length of the shortest path from Cl to C2 in G(S), and m if no such path exists. Definition. If S is a set of clauses, T is a support class for S, and n is a nonnegative inte- ger, then Q,(S, T) is EC E S: d(C, Ti)(n in G(S) for all Ti in T}, where d(C, Ti) is min{d(C, D):D E Ti}. Intuitively, if d(Cl, C2) is small, Cl and C 2 are "closely related." Also, Q,(S, T) is the clauses that are closely related to all the sup- port sets. Typically we will know that several clauses are essential to prove a theorem, and each such clause by itself can be made into a support set. Definition. A set S of clauses is fully matched if for all C there exists C2cS an a ES for all literals L LicC2 such that Ll an EC a 2 L1 are unifiable. 79 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Definition. R (S, T) is the largest fully matched subset of Q"(S, T). Thus we obtain R (S, T) from Qn(z, T) bynrepeatedly deleting Claus& containing unmatched" literals. This definition is not ambiguous, since if Q (S, T) contains more than one nonempty fully mate ed subset, then Rn(S, R T) is the union of all such subsets. Theorem 1. The function R is a relevance function. That is, if there is a length n refuta- tion from S, and T is support class, then there is a refutation from R (S, T) of length n or less. In fact, there is s&h a refutation from R Pl 6, T) Proof. Assume without loss of generalgty that NIL appears only once in the refutation and that every clause in the refutation contributes to the derivation of NIL. Let S1 be the set of input clauses appearing in the refutation. Then S1 is connected, intersects all the support sets, and has at most n elements. binary trees we can show that Using properties ;f elements. Note that R (S T) Sl has at most r;;l Vance criterion. Thatnisl; is a "globaL" relg- it depends in a non- trivial way on all the input clauses and on inter- actions between all the support sets in T. II EXAMPLES Let S be the Pl P2 -- Pl P2 E Ql p2 Q2 P3 Q7 following Q6 ?% 44 Q5 set of clauses: Q4 P4 LB 44 P3 P4 EL Q7 P4 Here ?i Ql indicates (??, Ql], i.e., E V Ql et cetera. Let Tlbe {{Pl], {P2]], let T2 be {{E]), and T = {Tl, T2). Then R (S, T) = R,(S, T) = R -- (S, T) = $j but R (S, T) =+Pl],{P2}" {Pl P2 P3? {p3,P4], {x}] which is in fact a miAimal'in:on-' sistent subset of S. For a second example, let S be the following: IN(a,box) IN(x,box) 1 IN(x,room) IN(x,room) > IN(x,house) E(x,house) ON(x,box) 1 m(x,box) ON(x,street) ' %?(x,house) IN(x,house) > IN(x,village) G(house,box) AT(house,street) ON(b,box) ON(c,street) m(d,village) Let Tl be {{IN(a,box))) and let T2 be ([IN(x, house)]]. Also, T = {Tl, T2]. R2(S, T) = R3(S, T) = 0 but R Then Rl(S, T) = (S, T) = {(IN(a, box)], (m&,box), IN(x,room) 4, {IN(x,room) IN(x, house)], {IN(x,house))]. This is a minimal'incon- sistent subset of S. Here "box", "room", "house", "a" et cetera are constants and x is a variable. Note that we cannot always guarantee that this re- levance criterion will yield minimal inconsistent . 80 sets as in these examples. is a set of clauses. Let Suppose S I I 4 1 be the length of S in characters when written in the usual way. Let Lits(S) be the sum over all clauses III ALGORITHMS C in S, of the number of literals in C. If S is a set of propositional clauses and IISII = m9 then G(S) may have O(m2) arcs. However, we can construct a modified version G,(S) of G(S) which has the same distances between klauses as G(S) does but which has only O(m) arcs. The idea is to reduce the number of arcs as follows: sup- pose C ,C. are the clauses containing L and D1, . . . . Dk%;e'tha clauses containing z. Then we add a node NL and arcs as follows in Gl(S): The numbers indicate the lengths of the arcs. Sim- larly, there are arcs of the form Di~-Nr,.f3C i2' Although Gl(S) is not a connection graph, and has arcs of length 0 and 1, it preserves distances be- tween clauses as in G(S). Using this modified con- nection graph, we have linear time algorithms to do the following, if S is a set of propositional clauses ,T=C T1 ,...,Tk] is a support class a suppo rt set, andn is a posi tive in teger: 2. Construct G1(S) from S. Find {C c S. d(C, T;> 5 nl. , Ti is 3. Given Q (S, T) for stpport class T, to find Rnn( S, T). Since step 2 must be performed 1 TI times to obtain Q (S, T), the total algorithm to obtain R (S, T) rzquires O(lT/IIS/ I) time. (Here ) Tl isnthe num- ber of support sets.) The algorithm to find {C E S: d(C,T.) < r$is a simple modification of standard shortestipath al- gorithms. gorithms see [ 31. For a presentation of these standard al- This can be done in linear time because the edge lengths are all 0 and 1. pute R,(S, T) as follows: We com- Definition. If S is a set of clauses, let M(S) be the largest fully matched subset of S. Note that Rn(S, T) = M(Q,(S, T)). The following algorithm M1 computes M(S) for set S of propositional clauses in linear time. This algorithm can therefore be used to compute R,(S, T) if S is a set of propositional clauses. Note that t is a push-down stack. procedure Ml(S); tt- empty stack; for all L such that LEC or ~EC for some C&S do clauses(L) f {C~S:LEC]; - count(L) + I clauses(L)1 od; for all C&S do member(C)T TRUE: for all LEC do ' - if count (L) = 0 then - push C on t; member(C) + FALSE fi od while t not empty do - pop c off t; for all L&C- do count(L) ycount(L) - 1; if count(L) = 0 then - for all C I% E clauses(l) do if mem er(C,) then - - push C on t; member Cl) + FALSE fi t - od fi - - od ;- return((CeS: member(C) = TRUE]); end Ml; If S is a set of first-order clauses, then G(S) can be constructed in O(Lits(S)I IS I I) time using a linear unification method [4]. This bound results because a unification must be attempted between all pairs of literals of S. The number of edges in G(S) is at most Lits(S)2. Given G(S), we can find {C E S: d(C, T.) < n] in time propor- tional to the number of &dges in G(S) (since all edge lengths are one). Also, given Q (S, T), we can find R (S, T) in time proportiona f to the num- ber of edggs in G(S) by a procedure similar to M above. The total time to find R (S, T) is there-l fore O<I T19cLits(S)2 + I IS) ILitsTfS)). If 11 Sl I= m then the time to find R (S, T) is O(m2( TI). By considering only the przdicate symbols of literals, the propositional calculus algorithms can be used as linear time preprocessing step to eliminate some clauses from S. An interesting problem is to compute R (S, T) efficiently for many values of n at the sa#ie time. There are methods for doing this, but we do not discuss them here. IV REFINEMENTS A. Connected Components Proposition 1. If there is a length n refu- tation from S, and T is a support class for S, then there is a length n refutation from one of the connected components of R IF1 (S, T). Also, the connected components can be found in linear time [ 31. B. Iteration Definition. If T = {T1,T2, . . ..Tk} is a sup- port class for S and Sl is a subset of S then Th Sl (T restricted to S1) is {Tl n S may be that T//R (S, ,...,Tkn Sl]. It T) # T or t at distances in ii G(Rn(S, T)) arendifferent than distances in G(S). This motivates the following definitions. Definition. R;(S, T) = S, R;(S, T) = Rn(S,T), and if i>l then Ri(S, T) = Rn(Riql(S, T),TlRi-l(S, n T)). Also, Ry(S, T) is the limit of the sequence R;(S, T), R$, T), R;(S, T), .. l Proposition 2. Rn i+l(S, T) GRi (S, T) for i>l. - Therefore the limit Ri(S, T) exists. Also, Rw(S, T) can be computed in at most I SI iterations oh@ >. thannth&? Can it be computed more efficiently Theorem 2. If there is a length n refutation from S, and T is a support class for S, then there is a length n refutation from one of the connected components of R m (S, T). $1 Proposition 3. For all i>O there exist n, S and T such that R:+'(S, T) = R: (S, T) # Rn i-l(S, T). Thus this computation can take arbi- trarily long to converge. Proof. Let n = 2, S = {P. 1 Pi+l: l<i<k] U CP i+l 3 P. : l<i<k). -- T. = {Pi: P -- Let T = {+l,T2,T3} where J i+l: i s j(mod3)) U {Picl 1 P. : isj (mod 3)). Then R;(S, T) = @ if 2azk but iy(S, T) # 6 if 2a < k. C. Selecting Support Clauses We now give another approach. Theorem 3. Suppose there is a length n refu- tation from S and T = IT ,...,T 1 is a support class for S. Then there'exist tlauses Ci E Ti, l<i<k, such that -- a) {Cl,C2,...,Ck] C R co (S, {CC,], CC,}, El . . . . {Ck}}) and L b) thzre is a length n refutation from R $1 (S, CCC,}, . . . . k.,H). Thus it is possible to select particular clauses from the support sets and use them to define R. There may be many sets R co (S, {{Cl), I"1 . . . , {C,}}) satisfying condition a) $bove, but they may be smaller than the connected components of R ~0 Pl (S, T). It is possible to construct examples z av- ing this property. Therefore it may help to use the above sets rather than R co (S, T) when search- ing for refutations. rg Another advantage is that it is possible to examine the clauses C. E T. in some heuristically determined order. Fu&herm;re, the 81 above approach is useful when the support sets Ti are not known in advance but are obtained one clause at a time. We now give a recursive procedure "re13" for generating all the sets as in Thereom 3. The idea is to order the Ti so as to reduce the branching factor as much as possible near the beginning of the search. Thus we use the principle of "least commitment." This procedure has as input sets Sl and S of clauses, integer n, and support class T Ri(S, {(D1),..., {Dj}, {Cl}, . . . . {Ck})) having ,...,Ck} as a subset, for Ci e Ti, ere is a length n refutation from S, and T is a support class for S, then there will be a length n refutation from some set output by rel3(lb, rtl, S, T). Definition. If S = {Dl, = {{D,) , . . . . {Dj% . . ..Dj) then Single(S) procedure rel3(S,,n,S,T) S2 -+ <(S, T 6 Single(S1)) if s1 C S2 then if T = fl then output (S2) else - T,+- TG I choose 'I$ e Tl minimizing I T21; for all C E T2 do od: re13(S1U {C),n,S2,Tl-T2) fi;-’ fi; end re13; By searching for such sets Sl, we can some- times obtain much better relevance criteria than by previous methods. The use of centers insures that elements of Sl will be closer together than in pre- vious methods. To implement this method, let S2 be Q p+2,@ T) . For each C E S2, let S3 be RrnT2 (S,{{b}}4).' -I 1-p If S3 intersects all support sets, then it is a candidate set of input clauses for a length n refu- tation. Here S, is a set of possible centers. Note that two clauses of S will have distance at most rtl + 1 in G(S3). No?e also that if n=6 then IF1 = 2 and if n = 10 then r?l = 3. Thus we can get somewhat nontrivial refutations with quite small distance bounds. E. Typing Variables For these relevance criteria to be useful, there must exist clauses C and C of S such that dC+ C2) is large. Howe&r, if z he axiom x=y 1 y=x is in S then two clauses of the form tl = t2 d D and t3 # t4 Tib v D2 will have distance 3 or less. is may cause everything to be close to everything else. To reduce this problem, we propose that all variables be typed as integer, Boolean, list, string, et cetera and unifications only succeed if the types match. Thus the above clauses would not necessarily be within distance 3 if tl and t4 or t2 and t3 have different types. The use of types may increase the number of clauses, since more than one copy of some clauses may be needed. However, the overall effect may still be beneficial. F. Logical Consequences D. Center Clauses By using the idea that graphs have "centers," we can reduce the distance needed to search for relevant clauses by another factor of 2. Theorem 4. Suppose there is a length n refu- tation from set S of clauses, and T is a support class for S. Then there exists a clause C E S and a set S 1 c S having the following properties: 1. ;: There is a length n refutation from S1 Sl is fully matched 4. Sl intersects all the support sets in T C E Sl and for all Cl E Sl, d(C, Cl) 2 +1 in G(Sl). Proof. Let Sl be the input clauses actually used insome minimal refutation from S. Then I s1 I Choose a "center" C of Sl, and note that The preceding ideas can also be applied to derivations of clauses other than NIL from S. Definition. A support set for S relative to C is a subset V of S such that C is not a logical consequence of S-V. A support class for S relative to C is a collection of support sets for S relative to c. For example, if I is an interpretation of S in which C is false, and V is the set of clauses of S that are false in I, then V is a support set for S relative to C. Definition. M(S, C) is the largest subset of S in which all literals are matched, except possi- bly those having literals of C as instances. Definition. Rn(s,T,c) is M(Q,(s,T) 23. Theorem 5. If there is a length n derivation of something subsuming C from S, and T is a support class for S relative to C, then there is a length n derivation of something subsuming C from R (S, T,O. $1 As before, we can introduce R O" (S,T,C) and other relevance criteria. 121 82 Plaisted, D. 5 G. Procedures scribed earlier. To incorporate procedural and heuristic in- formation, we may add clauses expressing the as- sertion A(x) 3 (sy)B(x,y) where A and B are input and output assertions for the procedure and x and y are input and output variables. To account for the fact that heuristics may fail, we assign pro- babilities of truth to clauses. The task then is to find a set S of clauses from which the desired consequence can'possibly be derived, subject to the condition that the product of the probabili- ties of the clauses in Sl is as large as possible. One way to do this is to run many trials, gene- rating relevant subsets of S, where the clauses of S are chosen to be present or absent with the appropriate probability. We then select a rele- vant set of clauses from among those clauses that have been found to be relevant in many of the trials. Note that if procedures are encoded as above, then a short proof may correspond to a so- lution using a few procedure calls, but each pro- cedure may require much time to execute. H. Subgoals If procedures are encoded as above, then each procedure may call the whole theorem prover recur- sively. This provides a possible subgoal mechan- ism. By storing the clauses from all subgoals in a common knowledge base, we may get interesting interactions between the subgoals. By noticing when subgoals are simpler than the original pro- blem in some well-founded ordering, we may be able to get mathematical induction in the system. The use of clauses, procedures, subgoals, and relevance criteria as indicated here provides a candidate for a general top-level control structure for an artificial intelligence system. V A COMPLETE STRATEGY The following procedure attempts to construct a refutation from set S of first-order clauses: procedure refute(s); for d = 1 step 1 until (NIL -- for j = 1 step 1 until (j refl(S, j, d) od od; -- is derived) do - > d) do - end refute; procedure refl(S, i, d); let T be a support class for S; RtR?(S, '0; if R 'is empty then return fi; ~Ru llevel 1 resolvents from R); if NIL E V or d = 1 then return fi; for j = 1 step 1 until(NIL is delved) do - refl(V, j, d - 1) &; end refl; This procedure selects at each step the clauses that seem most relevant and attempts to construct a refutation from them. Similar procedures can be given using other of the relevance functions de- A. Generating Support Sets One way to generate support sets for the above procedure is to let each support set be the subset of S in which specified predicate symbols occur with specified signs. This would yield 2n support sets for n predicate symbols. Of course, it is not necessary to use all of these support sets. A more interesting possibility is to have a collection {I,, 12, . . . , Sandtolet T. Ikl of interpretations of be the set of clauses that are false in I.. 'If I, has a finite domain then T. can be comiuted by ehaustive testing. Otherwisk, special methods may be necessary to determine if a clause C is true in I.. If I. has an infinite domain, a possible he?iristic 3s to let T. be the set of clauses that are false on some fiiite sub- set of the domain. If f is an abstraction mapping or a weak abstraction mapping [5] and I is an interpretation, then{CeS: some clause in f(C) is false in 11 is a support set for S. This approach may allow the use of nontrivial support sets which are easy to compute, especially if all elements of f(C) are ground clauses for all C in S. Note that T may include support sets obtained both syntac- tically and semantically. Although it may require much work to test if C is true in I., this kind of effort is of the kind that humans s&em to do when searching for proofs. Also, this provides a meaningful way of incorporating nontrivial seman- tic information into the theorem prover. The arcs in the connection graph resemble "associations" between facts, providing another similarity with human problem solving methods. REFERENCES [l] Gallaire, H. and J. Minker, eds. Logic and Data Bases. New York: Plenum Press, 1978. -- [2] Kowalski, R., "A Proof Procedure using Connection Graphs". J-ACM 22(1975)572-595. -- [3] Reingold, E. M., J. Nievergelt, and N. Deo. Combinatorial Algorithms: Theory and Practice. 7 Englewood Cliffs, New Jersey: Prentice-Hall, 1977. [4] Paterson, M. S. and M. N. Wegman, "Linear Unification", IBM Research Report 5304, IBM, 1976. [5] Plaisted, D., "Theorem Proving with Abstrac- tion, Part I", Departmental Report UIUCDCS-R- 79-961, University of Illinois, February 1979. 83	1980	7
67	This paper presents the results of research done on the representation of control knowledge in rule-based expert systems.’ It discusses the problems of representing co&o1 knowledge implicitly in object-level inference rules and presents specific examples from a MYCIN-like consultation system called PUFF. As an alternative, the explicit representation of conerol knowledge in sloes of a frame-like data structure is demonstrated in the CENTAUR system. Explicit representation of control knowledge has significant advantages both for the acquisition and modification of domain knowledge and for explanations of how knowledge is used in the expert system. REPRESENTATION OF CONTROL KNOWLEDGE IN EXPERT SYSTEMS Janice S. Aikins Computer Science Department Stanford University Stanford, California 94305 ABSTRACT I INTRODUCTION This paper emphasizes the importance of representing domain-specific control knowledge explicitly and separately from other forms of domain knowledge in expert systems. The particular focus of research on this topic has been MYCIN-like consultaeion systems [S] which represent their domain knowledge in the form of condition-action or production rules. Examples in this paper are taken from the PUFF system [43 which performs consultations in the domain of pulmonary (lung) physiology. The CENTAUR system was created in response to several knowledge representation and control structure problems in the rule-based systems, among which were the problems caused by ehe implicit representation of control knowledge. CENTAUR provides a framework for performing tasks using an hypothesize and match approach [5] to problem solving. This approach focuses the search for new information around recognized patterns of knowledge in the domain, a strategy that was not represented in the rule-based systems. Knowledge in CENTAUR is represented in the form of frame-like structures, called prototypes, which represent the expected patterns of knowledge, and in production rules, which serve as a stylized form of procedural attachment and are used to infer values or “fill in” slots in the prototype. This knowledge of prototypical situations is used for control of the consultation, for explanation of system performance, and also as a guide for acquiring additional knowledge and for modifying the existing knowledge base. ’ This work was supported by the Advanced Research Projects Agency under contract MDA 903-77-C-0322. Computer facilities were provided by the SUMEX-AIM facility at Stanford University under National Institutes of Health grant RR-00785-07. The author is sponsored by the Xerox Corporation under the direction of the Xerox Palo Alto Research Center. CENTAUR’s combination of prototypes and rules results in a knowledge representation that is expressive enough to allow the many kinds of domain knowledge necessary for system performance to be explicitly represented. Control know/edge for the consultation is represented in slots associated with each prototype, separately from the inference rules. Rules are associated with prototypes as the explicit contexts in which the rules are applied. The slots in the prototype specify the function of the attached rules, such as to summarize data already given or to refme an interim diagnosis. Other details of the CENTAUR system and a full discussion of the knowledge representation and control structure problems in the rule-based systems can be found in [I]. II TME PUFF SYSTEM One such rule-based system is the PUFF system which was created using a MYCIN-like framework. PUFF’s domain- specific knowledge is represented by a set of approximately 60 production rules. The “IF” part of the productlon states a set of conditions (the premise clauses) in which the rule is applicable. The action, or “THEN” part of the production, states the appropriate conclusions. The goal in PUFF is to interpret a set of lung function tests performed on a patient, and to produce a diagnosis of pulmonary disease in that patient. Each rule clause is a LISP predicate acting on associative (object- attribute-value) triples in the data base. In PUFF there is a single object, the patient. The attributes (or clinical parameters) are the lung function tests and other information about the patient. The PUFF control structure is primarily a goal-directed, backward chaining of the production rules as it attempts to determine a value for a given clinical parameter. written. A complete description of this mechanism is given in E61. III IMPLICIT CONTROL IN THE RULES P’roduction rules, in theory, are modular pieces of knowledge, each one capturing some “chunk” of domain- specific expertise. Indeed, one of the advantages of using production rules [3] is that there need be no direct interaction of one rule with the others, a characteristic which facilitates adding rules to the knowledge base or modifying existing rules. In practice, however, there are significant interactions among rules. Executing one rule will in turn cause others to be tried when the information needed for the first rule is not already known. Therefore, the order of the premise clauses of a rule affects the order in which other rules are executed. Further, in an interactive system such as PUFF, in which the user is asked for information that can not be inferred by rules, the order of the premise clauses also determines the order in which questions are asked. 121 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. This means of controlling question order by placing premise clauses in a specific order is, in fact, exploited by experts who recognize the need for ordering the questions that are asked, but have only this implicit and indirect mechanism for achieving their goal. In this case, the production rule framework itself becomes a programming language where the rules have multiple functions; some rules represent independent chunks of expertise with premise clauses specified in an arbitrary order, while others serve a controlling function with premise clauses that cannot be permuted without altering the behavior of the system. An example of implicit control knowledge is illustrated by the PUFF rule in Figure I below. This rule invokes other rules in an attempt to determine whether there is Obstructive Airways Disease. (Clause One), and if so, to determine the subtype (Clause Two) and findings associated with the disease (Clause Three). If Clause One were inadvertently placed after either Clause Two or Three, the system’s questions of the user would probe for more detailed information about Obstructive Airways Disease without having confirmed that the disease is present. For example, by reordering the clauses in RULEOOP, PUFF might begin its consultation by asking about the patient’s smoking history, one of the findings associated with Obstructive Airways Disease, and a question that would be inappropriate in a patient without a smoking-related disease. However, this rule contains no explicit indication that the order of the clauses is critical. The problem with implicit representation of control knowledge becomes apparent in working with the knowledge base, either to modify the knowledge or to explain its use in the system. RULE882 ----B-s If: 1) 2) 3) An attempt has been made to deduce the degree of obstructive airways disease of the patient, An attempt has been made to deduce the subtype of obstructive airways disease, and An attempt has been made to deduce the findings about the diagnosis of obstructive airways disease Then : It is definite (1.8) that there is an interpretation of potential obstructive airways disease FIGURE 1. PUFF Rule--Implicit Control Modifying rules is a normal part of system development. Clauses often must be added to or removed from rules in response to perceived errors or omissions in the performance of the system. However, removing or modifying the clauses of a controlling rule can alter the system’s behavior in unexpected ways, since the implicit control knowledge also will be altered. Therefore, modifications can be safely done only by persons intimately familiar with the knowledge base. This factor not only limits the set of people who can make modifications, and of course precludes the success of automatic knowledge acquisition systems in which each rule is considered individually, but it also limits the size of the knowledge base, as even the best of knowledge engineers can retain familiarity with only a limited number of rules at a time. A system’s explanations of its own performance also suffer when information critical to performance is not represented explicitly. The rule-based systems studied generate explanations of why questions are being asked using direct translations of those rules which were being used when the question was asked. (See [2] for details.) There is no distinction made between rules that control a line of reasoning, as opposed to rules that infer a piece of information. However, users of the system should be able to ask both kinds of questions in order to obtain justifications of the system’s reasoning process as well as justifications of its inference rules. The uniform representation of control and inference knowledge in rule-based systems further confuses the user by mixing the two kinds of explanations. IV CONTROL KNOWLEDGE IN CENTAUR Control knowledge about the process of pursuing an hypothesis in CENTAUR is represented in slots associated with each prototype, separate from the inference knowledge which will actually confirm or deny the hypothesis represented as production rules. Each slot specifies one or more LISP clauses, or control tasks, that are executed at specific points during the consultation as defined by a top-level prototype representing the “typical” consultation (the CONSULTATlON Prototype). For the pulmonary function domain, prototypes correspond to specific pulmonary diseases. During a CENTAUR consultation, initial case data suggest one or more disease prototypes as likely matches. Control knowledge in these prototypes then guides the consultation by specifying what information should be sought next. Expected data values in each prototype enable CENTAUR to pinpoint inconsistent or erroneous information during the consultation. Final conclusions are presented in terms of the prototypical situations determined to be present in the case, and any inconsistencies are noted. Thus the system developer can specify “what to do” in a given prototype context as an important part of the knowledge about the domain that is distinct from the inferential knowledge used in the consultation. These control tasks are specified as LISP functions, and the system developer can define any new functions as they are required. For example, Figure 2 shows CENTAUR’s representation of the control knowledge in the PUFF rule shown in Figure I. The control knowledge is represented in two of the control slots associated with the Obstructive Airways D’isease (OAD) prototype. They specify that when OAD is confirmed (the If-Confirmed Slot), the next tasks are to deduce a degree and a subtype for OAD, and, at a later stage in the consultation (when the prototype ACTION slots are executed), to deduce and print findings associated with OAD. If-Confirmed Slot: Deduce the Degree of OAD Deduce the Subtype of OAD. Action Slot: Deduce any Findings associated with OAD Print the Findings associated with OAD FIGURE 2. OAD Prototype Control Slots 122 Prototypes not only represent the domain-specific knowledge of a particular application, but also represent domain-independent knowledge about the operation of the CENTAUR system. At the highest level in CENTAUR, the Consultation Prototype lists the various stages of the consultation (e.g., entering initial information, suggesting likely prototypes, filling in prototypes) in its control slots. The advantages of explicit representation of control knowledge thus extend to control of the consultation process itself. V ADVANTAGES OF THE CENTAUR APPROACH The association of control knowledge with individual prototypes allows control to be specific to the prototype being explored. Thus domain experts can specify a different set of control tasks for each prototypical situation. In the pulmonary domain, for example, the expert proceeds in a different way if he has confirmed OAD rather than some other disease in the patient. Further, because this control knowledge is separate from the inference rules, the expert does not have to anticipate and correct incidental interactions between control and inference knowledge. Representing the entire consultation process itself as a prototype has additional advantages. First, the system designer’s conception of the consultation process is clearly defined for all system users. Second, representing each stage of the consultation as a separate control task allows stages to be added or removed from the consultation process. For example, the Refinement Stage, which uses additional expertise to improve upon an interim conclusion, was omitted during early stages of system development for the pulmonary function problem. “Filling in” a consultation prototype with user- specified options, such as a choice of strategy for choosing the current best prototype (for example, confirmation, elimination, or fixed-order), results in a control structure that can be tailored to the desires of each individual user. The organization of knowledge into prototypical situations allows the user to more easily identify the affected set of knowledge when changes to the knowledge base are desired. Points at which specific control knowledge is used during the consultation are clearly defined, with the result that it is easier to predict the effects of any control modifications that may be made. Explicit representation of control knowledge also facilitates explanations about that knowledge. In addition to the HOW and WHY keywords available in MYCIN, a new keyword, CONTROL, has been defined so that a user of the system can inquire about the control task motivating the current line of reasoning. For example, if the user types “CONTROL” in response to a system question about the patient’s smoking history, the system would respond, The current control task Is to determine the findings associated with OAD. VI SUMMARY This paper has discussed the importance of representing control knowledge explicitly, particularly as it affects knowledge acquisition and explanation in a knowledge-based system. The representation of control knowledge as slots in a prototype in the CENTAUR system demonstrates one feasible approach. Augmenting the rule representation to include rules that function exclusively as control rules might be another. The critical lesson learned from working with the rule-based systems is that the system’s representation structures must be expressive enough to represent control knowledge explicitly, so that it will not be inaccessible to the system and to the knowledge engineer. ACKNOWLEDGMENTS Many thanks to Doug Aikins, Avron Barr, Jim Bennett, Bruce Buchanan, and Bill Clancey for their helpful advice and comments on earlier versions of this paper. REFERENCES [l] Aikins, J. Prototypes and Production W/es: A Knowledge Representation for Computer Consultations. (Forthcoming Ph. D. Thesis), Heuristic Programming Project, Dept. of Computer Science, Stanford University, 1980. [Z] Davis R. Applications of Meta Level Knowledge to the Construct/on, Maintenance and Use of Large Knowledge Bases. STAN-CS-76-552, Stanford Universlty, July 1976. [3] Davis R., and King, J. An Overview of Production Systems. In E. W. Elcock and D. Michie (Eds.), Machine lntelllgence 8. New York: Wiley & Sons, 1977. Pp. SOO- 332. [4] Kunz, J., Fallat, R., McClung, D., Osborn, J., Votteri, B., Nii, H., Aikins, J., Fagan, L., and Feigenbaum, E. A Physiological Rule Based System for lnterpretlng Pulmonary Function Test Results. HPP-78-19 (Working Paper), Heuristic Programming Project, Dept. of Computer Science, Stanford University, December 1978. [5] Newell, A. Artificial Intelligence and the Concept of Mind. In R. Schank and K. Colby (Eds.), Computer Models of Thought and Language. San Francisco: W. H. Freeman and Company,l973. Pp. l-60. f6j Shortliffe, E. H. MYCIN: A Rule-based Computer Program for Advising Physicians Regarding Antimicrobial Therapy Selection. Ph. D. dissertation in Medical Information Sciences, Stanford University, 1974. (A Iso, Computer-Based Medical Consultations: MYCIN. New York: American-Elsevier, 1976. 123	1980	70
68	DELTA-MIN: A Search-Control Method for Information-Gathering Problems Jaime G. Carbonell Computer Science Department Carnegie-Mellon University Abstract The A-MIN method consists of a best-first backtracking algorithm applicable to a large class of information-gathering problems, such as most natural language analyzers, many speech understanding systems, and some forms of planning and automated knowledge acquisition. This paper focuses on the general A-MIN search-control method and characterizes the problem spaces to which it may apply. Essentially, A-MIN provides a best-first search mechanism over the space of alternate interpretations of an input sequence, where the interpreter is assumed to be organized as a set of cooperating expert modules.’ 1. Introduction A present trend in Al is to design large systems as cooperating collections of experts, whose separate contributions must be integrated in the performance of a task. Examples of such systems include HEARSAY-II [4], POLITICS [l], PSI [5], SAM, [3]. The division of task responsibility and grain size of the experts differs markedly. For instance, the latter two systems contain few large-scale experts, which are invoked in a largely predetermined order, while the former two systems contain a larger number of smaller modules whose order of invocation is an integral part of problem solving task itself. In this paper I discuss a new search method, called A-MIN, that incorporates some of the desirable features from best-first search and some properties of gradient search. (Gradient search is locally-optimized hill-climbing.) The primary objective is to make global control decisions based on local knowledge provided by each expert module. No module is required to know either the internal structure of another module, or the overall controling search mechanism. In this way, I depart somewhat from the freerer blackboard control structure of HEARSAY-II, where search was controled by the experts themselves. The module that “shouted loudest” was given control, hence each module had to know when and how loud to shout with respect to other expert modules. In addition, there was a “focus knowledge source” [6] that helped guide forward search. This method acquires its flexibility by placing a substantial amount of global control responsibility on local experts. Moreover, it entails no externally-transparent search discipline. Finally, the primary emphasis is on forward search, not reconsidering wrong decisions in favor of choosing an alternate interpretation. In light of these considerations, I attempted to factor domain knowledge (what the experts know) from search discipline (when to pursue alternate 1 This research was sponsored in part by the (ONR) under grant number NO001 4-79-C-0661. Office of Naval Research paths suggested by different experts), so that each problem may be investigated in its own right. Here, I focus on the search control aspect, and consider the internal structure of each domain expert as a virtual “black box”. To simplify matters, I confine my discussion to tasks whose terminating condition is defined by processing an input sequence to completion without error. This class of problems is exemplified by natural language analysis, where an input sentence is processed left-to-right and the goal state is the formation of a consistent semantic representation of the input. Clearly, this is a satisficing rather than optimizing task [8], in the sense that only the first of potentially many solutions is sought. Since I want the language analysis to give the same parse of the sentence as a human would, the process must be biased to favor reaching the appropriate solution first. This biasing process, based on local decisions made by expert modules is the primary input to the A-MIN search method described below. It must be noted, however, that the left-to-right processing assumption is more restrictive than the HEARSAY paradigm, where “islands of interpretation” could grow anywhere in the input sequence and, when possible, were later merged into larger islands until the entire input sequence was covered [6, 41. 2. An Information-Gathering Search Space Consider a search space for the task of processing a finite sequence of input symbols (such as an English sentence) and producing an integrated representation incorporating all the information extracted from the input (such as a semantic representation of the meaning encoded in the sentence). The problem solver consists of a set of experts that may be applied at many different processing stages, without fixed constraints on their order of application. For instance, in the language analysis domain, one can can conceive of a verb-case expert, a morphological-transformation expert, an extra-sentential referent-identifier expert (or several such experts based on different knowledge sources), a dialog-context expert, an immediate-semantics expert, a syntactic-transformation expert, etc... A robust language analyzer must be capable of invoking any subset Of these and other experts according to dynamically determined needs in analyzing the sentence at hand. NOW, let US back off the natural language domain and consider the general class of problem information-gathering,* spaces to which an cooperating-expert approach appears useful. First, we draw a mapping between the general problem solving terminOlOgy and the expert module approach. The search space outlined below is a considerably constrained version of a general search space. This property is exploited in the A-MIN search method described in the fcllowing Section. 124 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. e The operators in the search space are the individual expert modules. Each module may search its own space internally, but I am concerned only with the macro-structure search space. Each expert has conditions of strict applicability and preference of applicability. The latter are used for conflict resolution decisions when more than one expert is applicable. o A state in the space consists of the total knowledge gathered by invoking the set of experts that caused the state transitions from the initial state to ihe present. This definition has two significant implications: An expert that adds no new knowledge when invoked does not generate a new state; therefore, it can be ignored by the search control. There is a monotonicity property in that each step away from the initial states adds information to the analysis, and therefore is guaranteed to “climb” to a potential final state. (Left-to-right, single-pass natural language analysis can exhibit such monotonic behavior.) e A final state is defined by having reached the end of the input sequence without violating a path constraint, and no expert can add more information (i.e., no transition in the state space is possible). e A path constraint is violated if either a new segment of the input cannot be incorporated or an expert asserts information that contradicts that which is already part of the current state. When this situation arises, directed backtracking becomes necessary. o The initial state is a (possibly empty) set of constraints that must be satisfied by any interpretation of the input sequence. For instance, a dialog or story context constrains the interpretation of an utterance in many natural language tasks. o Each expert can draw more than one conclusion when applied. Choosing the appropriate conclusion and minimizing backtracking on alternatives is where the real search problem lies. Choosing the next expert to apply is not a real problem, as the final interpretation is often independent of the order of application of experts. That is, since information-gathering is in principle additive, different application sequences of the same experts should converge to the same final state. The experts preselect themselves as to applicability. Selecting the expert who thinks it can add the most information (as in HEARSAY-II) only tends to shorten the path to the final state. The real search lies in considering alternate interpretations of the input, which can only be resolved by establishing consistency with later information gathered by other experts. Finally, given the possibility of focused backtracking fro& a dead end in the forward search, less effort needs to be directed at finding the “one and only correct expert” to apply. Search Metho A-MIN is a heuristic search method specifically tailored to the class of search spaces described above. It combines some of the more desirable features of gradient search and best-first search, with the modularized information sources of a cooperating-expert paradigm. Figure 3-O is the search-control strategy algorithm in an MLISP-style form. Subsequently I discuss how A-MIN works, exemplifying the discussion with a sample search-tree diagram. Every expert is responsible for assigning a likelihood-of-correctness value to each alternative in the interpretation it outputs. These values are only used to determine how much better the best alternative is than the next best alternatives, an item of information crucial to the backtrack control mechanism. There is no global evaluation function (the outputs of different experts are not directly comparable - it makes no sense to ask questions like: “Is this anaphoric referent specification better than that syntactic segmentation?“) Nor is there any mechanism to compute differences between the present state and the goal state. (Recall our definition of goal state -- only upon completion of the input processing can the goal state be established.) PROCEDURE A-MIN(initial-state. experts) altlist := NULL globaldelta := 0 state := initial-state input := READ(firsfinput) NEXTOP: IF NULL(input) THEN RETURN(state) ELSE operator := SELECTBEST(APPLICABLE(experts)) IF NULL(operator) THEN input : = READ( next input) ALSO GO NEXTOP ELSE alts := APPLY(operator, state) IF NULL(alts) THEN MARK(operator. 'NOT-APPLICABLE, 'TEMP) ALSO GO NEXTOP bestalt := SELECTMAX(alts) IF Ilaltsll > 1 THEN alts := FOR-EACH alt IN REMOVE(bestalt, alts) COLLECT <'ALT: alt, 'STATE: state, 'DELTA: globaldelta + VALUE(bestalt) - VALUE(alt)> altlist := APPEND(alts, altlist) NEWSTATE: state := MERGE-INFORMATION(state, bestalt) IF NOT(state = 'ERROR') GO NEXTOP ; if no error, continue gradient search, ; else delta-min backup below WHILE state has no viable alternatives DO BEGIN MARK(state, 'DEAD-END, 'PERM) ; delete dead ends state := PARENT(state) ; from search tree END backup-point := SELECTDELTA-MIN(altlist) state := GET(backup-point. 'STATE:) globaldelta := GET(backup-point, 'DELTA:) bestalt := GET(backup-point, 'ALT:) altlist := REMOVE(backup-point. altlist) GO NEWSTATE END A-MIN Figure 3-1: The A-MIN Search-Control Algorithm Let us see how A-MIN can be applied to an abstract example, following the diagram in figure 3-Z The roman numerals in the arcs reflect the order in which they are traversed. At the initial state, expert-4 applies and generates three alternate interpretations of the input. One alternative is ranked as most likely. A “A” value is computed for the remaining alternatives, 2 "lnformetion gathering" is a term coined by Rai Reddy to refer to encoding the difference in confidence that expert-4 had between search problems where progress towards agoalstateischaracterized by them and the most likely alternative. The more sure expert-4 is of accruing and integrating information from outside SOUrCeS. its best choice relative to the other alternatives, the larger the A 125 values. The best interpretation generated by expert-4 is integrated with the initial state constraints and found to be consistent. At this point, a new state has been generated and expert-2 applies to this state generating no new information. More input is read and expert-l applies, generating only one alternative, which is found to be consistent with the information in the present state. In a similar fashion, the rest of the tree in figure 3-xis generated. Up to now, we have witnessed an instance of gradient search, where a different evaluation function is applied at each node (The local evaluation function is, in effect, the expert who generated the likelihood values.) If no error occurs, (i.e., if the interpretation of the input remains consistent) no backup is needed. The likelihood rankings clearly minimize the chance of error as compared to straightforward depth first search. Now, let us consider the possibility of an inconsistency in the interpretation, as we continue to examine figure 3-X Figure 3-2: A-MIN Search Tree With Directed Backup The most likely interpretation generated by expert-6 was found to be inconsistent with the information in the present state. Backup is therefore necessary, given the depth-first nature of the search. But, where do we back up to? Normally, one might consider a depth-first unwinding of the search tree; but, is this the most reasonable strategy? Expert-5 was much less certain in its choice of best alternative than expert-6 (A = 2 vs A = 4). It seems more reasonable to doubt expert-5’s decision. Therefore, one wants to back up to the point where the probability of having chosen a wrong branch is highest, namely to the choice point with the minimal A (hence the name A-MIN). Continuing with figure 3-z we restore the state at expert-5 and incorporate the A = 2 interpretation. It is found to be consistent, and we apply expert-l to the new state. The best interpretation of expert-l leads to error, and backup is again required. Where to now? The minimal A is at expert-l, but this would mean choosing a non-optimal branch of a non-optimal branch. Lack of confidence in the choice from expert-5 should be propagated to the present invocation of expert-l. Hence, we add the two As in the path from the inital state and get the value: A =3, which is greater than the minimal A at expert-4 (A = 2). Therefore, we back up to expert-4. This process continues until a consistent interpretation of the entire input is found (i.e., a goal state is reached), or the search exhausts all viable alternate intepretations. Essentially, A-MIN is a method for finding one globally consistent interpretation of an input sequence processed in a predetermined order. In natural language analysis, the problem is to find a semantically, syntactically, and contextually consistent parse of a sentence. In speech understanding the constraint of formulating legal phonemes and words is added, but the nature of the problem and the applicability of the A-MIN approach remains the same. For instance, A-MIN is an alternate control structure to HARPY’s beam search [7], which also processes a sequence of symbols left to right, seeking a globally consistent interpretation. 4. Concluding Remarks To summarize, the A-MIN method exhibits the following properties: o A-MIN is equivalent to gradient search while no error occurs. Path length (from the initial state) is not a factor in the decision function. o The backtracking mechanism is directed to undo the choice most likely to have caused an interpretation error. This method compares all active nodes in the tree, as in best-first search, but only when an error occurs (unlike best-first search). o Perseverance in one search path is rewarded, as long as the interpretation remains consistent, while compounding less-than-optimal alternate choices is penalized. This behavior falls out of the way in which A values are accrued. o No global evaluation function forces direct comparisons among information gathered by different knowledge sources. Such an evaluation function would necessarily need to encode much of the information contained in the separate experts, thus defeating the purpose of a modular cooperating-expert approach. The A comparisons contrast only the differences between locally-optimal and locally-suboptimal decisions. These differences are computed by local experts, but the comparisons themselves are only between relative ratings on the desirability of alternate decisions. 126 Additional discussion of implementation, analysis, and details of the A-MIN search method may be found in [2], where an effective application of A-MIN is discussed for constraining search in a natural language processing task. 1. 2. 3. 4. 5. 6. 7. 8. References Carbonell, J. G., “POLITICS: An Experiment in Subjective Understanding and Integrated Reasoning,” in inside Computer Understanding: Five Programs Plus Miniatures, R. C. Schank and C. K. Riesbeck, eds., New Jersey: Erlbaum, 1980. Carbonell, J. G., “Search in a Non-Homogeneous Problem Space - The A-MIN Algorithm,” Tech. report, Dept. of Computer Science, Carnegie-Mellon University, 1980. Cullingford, R., Script Application: Computer Understanding of, Newspaper Stories, PhD dissertation, Yale University, Sept. 1977. Erman L. D. and Lesser, V. R., “HEARSAY-II: Tutorial Introduction & Retrospective View,” Tech. report, Dept. of Computer Science, Carnegie-Mellon University, May 1978. Green, C. C., “The Design of the PSI Program Synthesis System, ” Proceedings of the Second International Conference on Software Engineering, October 1976 , pp. 4-18. Hayes-Roth, F. and Lesser, V. R., “Focus of Attention in the Hearsay-II Speech Understanding System,” Proceedings of the Fifth international Joint Conference on Artificial Intelligence, 1977 , pp.. 27-35. Lowerre, B., “The HARPY Speech Recognition System,” Tech. report Computer Science Department, Carnegie-Mellon University, April 1976. Newell, A. and Simon, H. A., Human Problem Solving, New Jersey: Prentice-Hall, 1972. 127	1980	71
69	ON WAITING Arthur M. Farley Dept. of Computer and Information Science University of Oregon Eugene, Oregon ABSTRACT Waiting is the activity of maintaining selected aspects of a current situation over some period of time in order that certain goal-related actions can be performed in the future. Initial steps toward the formalization of notions relevant to waiting are presented. Conditions and information pertin- ent to waiting decisions are defined. Introduction Waiting is the activity of maintaining selected aspects of a current situation over some period of time in order that certain goal-related actions can be performed in the future. Those aspects of the current situation which are maintained serve as preconditions for the desired future actions. Waiting extends from a decision to maintain these preconditions until (attempted) execution of the goal-related actions (or a decision to forego them). The act of waiting is closely associated with the future actions; a problem solver is said to be "waiting to" (do) these actions. The primary function of waiting is to improve the efficiency with which problem solving plans are executed. Waiting attempts to overcome reestablisk ment of preconditions for desired, future actions. Not only may less effort be expended during problem solving, but performance can become more coherent. The problem solver can avoid frequent side- and back-tracking in multiple-goal situations. Waiting also allows for a certain degree of parallelism in problem solving. Waiting can be engaged in simul- taneously with other actions which do not destroy satisfaction of the preconditions being maintained. Waiting is an important and frequent problem solving activity in the real-world. By real-world, we mean an ongoing, continuing, schedule based con- text, within which cooperative, as well as compet- itive, efforts among groups of problem solving systems normally occur. Waiting minimally requires such a context to be effective. A decision to wait implies that other, as yet unsatisfied, precondi- tions of anticipated goal-related actions are expected to be met by means other than direct intervention by the waiting system. Waiting as a problem solving activity has been largely (if not totally) ignored by AI research to date. This is primarily due to the fact that real- world contexts as defined here have only recently been considered. This paper outlines initial steps toward formalisms within which issues of waiting can be addressed. The research represents exten- sions to a knowledge-based problem solving system previously described by the author [3,41. We briefly review important aspects of that system before describing straightforward extensions which aid our understanding of waiting. We conclude by discussing related research and suggesting future work. Knowledge-based Problem Solving The form of its representation of the environ- ment influences all other aspects of a problem solving system. Let a situation of the relevant environment at any point in time (past, present, or future) be represented by a situation state. A situation state consists of a finite set of propo- sitions which are true with respect to the specific environmental situations(s) which the state is said to represent. The current state represents the present environmental situation. A goal state is a situation state which the problem solving system desires the current state to satisfy (i.e. be con- sistent with). A problem exists for a problem solving system when the current state does not sat- isfy constraints specified by a goal state. Proh- lem solving refers to any activity undertaken in attempts to eliminate differences between current and goal states. A problem is solved when differ- ences between its goal state and the current state no longer exist. A knowledge-based problem solving system solves most problems by instantiation and execution of known general solution plans. A general solution plan describes a process which is capable of satis- fying a set of goal states from any of a set of current states. A general solution plan is repre- sented as a rooted, directed, labelled tree of plan states. The root of the tree is the goal state. Plan states of the tree are interconnected by dir- ected arcs, each labelled by an operator. An oper- ator is a description of an action (or process), represented as sets of add, delete, and precondition propositions [71. The operator labelling an arc is capable of transforming the plan state at the tail of the arc into the plan state at the head of the arc. The plan state at the tail satisfies precon- ditions of the operator, while the one at the head reflects the results of additions and deletions associated with the operator. The maximal directed path from any plan state ends at the goal state. A plan state is a situation state whose pro- positions have been partitioned into three compo- nents. For each proposition in the SELFOP compo- nent, the problem solving has one or more operators capable of satisfying the proposition. Furthermore, From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. the system normally expects (prefers) to satisfy a SELFOP proposition itself by executing one or the operators. For each proposition in the OTHEROP component, (costly) operators may exist allowing the system to satisfy the proposition itself, but the system normally expects the proposition to be satisfied by other problem solving systems in the environment. Finally, for each proposition in the NOOP component, neither the system itself nor any other problem solving system has control over sat- isfaction of the proposition (i.e., weather condi- tions). Though the structure of a general solution plan is made clear by its tree, representing the plan as a production system can facilitate its execution. A production system [61 is a collection of condition-action pairs called rules. In the pro- duction system representation of a general solution plan, each plan state serves as the condition part of a rule whose action part is the operator label- ling the arc leaving that plan state. In [41, the author describes how this representation can be useful in a selective approach to coordinating the execution of multiple plans. Rules from general solution plans associated with as yet unsatisfied goals are combined to form one production system. A rule classification scheme is defined which pro- vides for the avoidance of most inter-plan con- flicts while allowing for responsive, efficient problem solving. For example, a goal state is classified as ungS;ounded if it denies satisfaction of conditions from a (critical) rule of another plan (in a conjunctive goal set). Rules from plans for ungrounded goal states are not executed until conflicting plans have been completed. Waiting We are concerned here with the control of wait- ing during plan executions by knowledge-based pro- blem solving systems operating in real-world con- texts. We first consider the question: what situ- ations trigger consideration of waiting? A rule is self-satisfied if all propositions in its SELFOP component are satisfied. Whenever a self-satisfied rule exists during plan execution, a problem solv- ing system may consider waiting, maintaining those conditions of the rule it has satisfied until the rule can be executed. A rule whose conditions are completely satisfied, and thus can be executed during a current system cycle, is classified as ready. Ready rules can suggest waiting. The sys- tem may wait to execute one ready rule, selecting another which leaves the rule ready (through con- flict resolution). During plan execution, a rule corresponding to a plan state whose NOOP component is not satisfied is classified as irrelevant. Such a rule can only fire if uncontrollable aspects of the environment change favorably. An irrelevant rule which is otherwise satisfied is a contingent rule. Being self-satisfied, contingent rules suggest consider- ation of waiting. Any NOOP condition must be satis fied occasionally, otherwise a rule would be fan- tasy and not considered part of a problem solving system's executable plans. An important class of NOOP propositions are those which deal with time. In a general solution plan reflecting real-world time constraints (a scheduled plan), the NOOP component of a plan state will contain one or both of the propositions ISNOWBEFORE(t) and ISNOWAFTER(t), where t is a time specification. The truth of a time proposition is determined relative to the current time, considered to be an ever present aspect of the current state. Time propositions of plan states derive from time constraints placed upon goal state satisfaction. They are propogated to other states (rules) of a scheduled plan, with time parameters adjusted to reflect time estimates for traversed operators. In a scheduled plan, time propositions alone may determine the relevancy status of rules. Rules which are irrelevant solely due to unsatisfied time constraints have their NOOP components classified as early or late, depending upon which time propo- sition is not satisfied. A rule with early NOOP component, but which is otherwise ready, is classi- fied as imminent. Whenever an imminent rule exists during plan execution, the problem solving system may consider waiting while time passes until the rule becomes relevant and can be fired. One source of imminent rules are overestimations of time re- quirements for completed, prior operations from a scheduled plan. For example, you estimate a 30 minute drive to a shopping mall; but it only takes 15; you arrive before the mall opens. The rule by which you would enter the mall is imminent. Another situation which may prompt waiting is the existence of a dependent rule. A relevant, self-satisfied rule corresponding to a plan state whose OTHEROP component is not satisfied is ClaSSi- fied as dependent. Whenever a dependent rule exists during plan execution, the system may consider waiting until the expected (necessary) assistance arrives and the rule can be fired. Dependent rules can often arise during cooperative problem solving efforts. For example, you are painting a house with a friend and turn around, expecting him to hand you a new can of paint, but he hasn't finished opening it. The rule by which you would take the paint is dependent. Finally, rules which have only their SELFOP component satisfied are classified as needy. Needy rules arise frequently when coordinating with pub- lic, problem solving support systems, such as mass transportation. To board a bus, a problem solver must be at a bus stop (SELFOP), at the scheduled time (NOOP), with the bus being at the stop (OTHEROP) . If a problem solver arrives at a stop five minutes early, the rule by which it boards the bus is needy. 129 Given the existence of a self-satisfied rule, what information is pertinent to subsequent waiting decisions? With each self-satisfied rule, the pro- blem solving system can associate three values: a waiting period, a set of compatible rules, and a set of contending rules. The waiting period is an estimate of the length of time before a self- satisfied rule will become ready. A compatible rule is a ready rule which requires less (esti- mated) time than the waiting period, and would not destroy satisfaction of the self-satisfied rule's conditions. A contending rule is a ready rule from another scheduled plan which would destroy the other's self-satisfaction, requires more time than the waiting period, but will become classified late if it is not fired within the waiting period. Determining these two sets of rules would not require dramatic additional computational effort. They only contain ready rules, rules which the system always determines before selecting the next rule to execute. A simplest policy for waiting can be stated in terms of these two sets of rules. If there is a contending rule associated with a self-satisfied rule, do not wait; otherwise wait, firing a com- patible rule, if any exist. Though this may pro- duce effective behavior in many circumstances, a moment's thought suggests further considerations. A goal state is nearby to a self-satisfied rule if time estimates indicate that the system could satisfy the goal state and reestablish conditions of the self-satisfied rule within the expected waiting period. A more complex waiting policy could have a system elect to wait in the face of contending rules, especially when compatible rules and/or nearby goal states allow active waiting. Relative importances of goal states would influence such a policy. Finally, rather than an expected waiting period, a cumulative probability function representing the likelihood that a self-satisfied rule will be able to fire within a given period of time could add further sophistication to waiting policies. Conclusion Research on distributed problem solving has used waiting as a coordination primitive [ll. Coordination is realized in our model through ungrounded goal states and unsatisfied OTHEROP com- ponents. Postponing actions until others have completed their responsibilities may or may not result in waiting as defined here. Research on medical diagnosis has dealt with time duration propositions in rule conditions [2]. These suspend judgement until time lends its support, Again, time-based postponement differs from issues of waiting addressed here. In this paper we have specified formalisms which add waiting to the repertoire of capabilities of problem solving systems dealing with real-world contexts. An (incomplete) set of conditions under which waiting may be reasonably considered are defined, as are aspects of the subsequent waiting decision process. We are continuing investigation of waiting by considering a particular context -- that of satisfying a set of work and personal goals on a given day within an office and city setting. This context has received attention in recent research on plan formulation [51. It appears to be just as fruitful a source of ideas on plan execu- tion and the role of waiting. Meeting schedules and office hours, going shopping, and using public transportation are actions which require frequent, even planned, waiting. A simulation to evaluate various waiting policies is planned. r11 [21 r31 II41 [51 [61 [71 REFERENCES Corkill, D. 'Hierarchical planning in a distributed environment:, Proceedings IJCAI-79, Tokyo, 1979, p. 168-175. Fagan L. etal., "Representation of dynamic clinical knowledge", Proceedings IJCAI-79, Tokyo, 1979, p. 260-262 Farley, A.M. "The coordination of multiple goal satisfaction', Proceedings IJCAIS, MIT, 1977, p. 495. Farley, A.M. "Issues in knowledge-based problem solving" to appear in I.E.E.E. Trans- actions on Systems, Man, and Cybernetics, August, 1980. Hayes-Roth, B. and Hayes-Roth F. "A cog- nitive model of planning", Cogntive Science, 3 (19791, pp. 275-310. Newell, A. and Simon, H.A. Human Problem Solving, Prentice-Hall: Englewood Cliffs, NJ, 1972. Sacerdoti, E.D. "Planning in a hierarchy of abstraction spaces", Artificial Intelligence, 5, 1974, pp. 115-135. 130	1980	72
70	DOU&S E. &pelt Slar@rd University, Stanford, California SRI International, Menlo Park, California ABSTRACT This paper reports recent results of research on planning systems that have the ability to deal with multiple agents and to reason about their knowledge and the actions they perform. The planner uses a knowledge representation based on the possible worlds semantics axiomati7ation of knowledge, belief and action advocated by Moore [5]. This work has been motivated by the need for such capabilities in natural language processing systems that will plan speech acts and natural language utterances [1, 21. The sophisticated use of natural language requires reasoning about other agents, what they might do and what they believe, and therefore provides a suitable domain for planning to achieve goals involving belief. This paper does not directly address issues of language per se, but focuses on the problem-solving rcquircmcnts of a language-using system, and describes a working system, KAMP (Knowledge And Modalities Planner), that embodies the ideas reported herein. I. WI-IAT,A KNOWLEDGE PLANNER MUST DO Consider the following problem: A robot named Rob and a man named John arc in a room that is adjacent lo a hallway containing a clock. Both Rob and John are capable of moving, reading clocks, and talking to each other, and they each know that the other is capable of performing these actions. They both know that they are in the room, and they both know where the hallway is. Neither Rob nor John knows what time it is. Suppose that Rob knows that the clock is in the hal& but John does not. Suppose further that John wants to know what time it is, and Rob knows he does. Furthermore, Rob is helpful, and wants to do what he can to insure that John achieves his goal. Rob’s planning system must come up with a plan, perhaps involving actions by both Rob and John, that will result in John knowing what time it is. We would like to see Rob devise a plan that consists of a choice between two alternatives. First, if John could find out where the clock was, he could go to the clock and read it, and in the resulting state would know the time. So, Rob might tell John wh&re the clock was, reasoning that this information is sufficient for John to form and execute a plan that would achieve his goal. The second alternative is for Rob to move into the hall and read the clock himself, move back into the room, and tell John the time. This research was supported by ihe Defense Advanced Research Projeers Agency under contract NOOO39-79-C-0118 with the Naval Electronic Systems Command The views and conclusions contained in this document are those of the aitthor and should not be interpreted as represenlative of the oficial policies, eilher expressed or implied, of the Defense Advanced Research Projects Agency, or the I/. S. Government. Existing planning mechanisms such as NOAII [6] or SrI:It% [3] arc incapable of dealing with this sort of problem. First, to solve this problem a system must reason cffcctivcly about propositional attitudes such as knorv, believe, and want. Fxisring planning syctcms arc based on knowledge rcprcscntations that arc not adequate for that purpose. Morcovcr, they arc not equipped to handle the integration of the actions of multiple agents into a single plan. In lhc solution to the above problem, the first choice consists of actions by both John and Rob. Rob does an informing act, pnd John moves into the hall and reads the clock. This means that Rob has planned for events to occur that are beyond its immediate control and which involve knowledge about the capabiliticc of another agent. The KAMP system solves problems such as the cxamplc above. It adopts a knowledge representation based on possible worlds semantics, which is capable of representing the knowledge needed for the task. By reasoning about the knowledge and wants of other agents, KAMP dctermincs what courses of action other agents can be expected to take in the filturc, and incorporates those actions into its own plans. II. REPRESENTING KNOWT,EDGE ABOUT BELIEF It is important that a planner be based on a knowledge representation that is adcquatc for representing different kinds of facts about knowing, wanting, believing, etc. For example, it may be necessary to reprcscnt that someone knows the value of a term, without the system itself knowing what that vnluc is, or the system may need to rcprcscnt that a person knows -P as opposed to not knowing whether P. A variety of strategies have been suggcstcd for representing such knowledge, but just representing the knowledge is not sufficient. It is also necessary that the system be able to reason with the knowledge efficiently. Many of the alternatives that have been proposed for rcprcscnting knowledge are fundamentally lacking in cithcr rcprcscntational adequacy or efftcicncy. Moore [5] discusses some of the specific proposals and their shortcomings. The representation which has been selected For the KAMP system is based on Moore’s axiomatization of possible worlds semantics. This approach has a great deal of power to represent and reason ‘efficiently with modal operators, and it is particularly elegant in describing the relation between action and knowledge. Because the design of the planner is largely motivated by the design of the knowledge representation, I will briefly outline Moore’s strategy for representing knowlcdgc about belief and how it relates to action. For comparison with a system that uses a different knowledge representation for planning to influence belief, see Konolige and Nilssun [4]. 131 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. The representation consists of a modal object language that has operators s~uz11 as believe and knovv. This object langauge is translated into a m&a-language that is based on a first order axiomitization of the possible worlds semantics of the modal logic. All the planning and deduction takes place at the lcvcl of the mcta language. In this paper I will adopt the convention of writing the meta language translations of object language terms and predicates in boldface. For example, to represent the fact that John knows P, one asserts that P is true in every possible world that is compatible with John’s knowledge. If K(A, wl, w2) is a predicate that means that w2 is a possible world which is compatablc with what 4 knows in wl, and T(w, P) means that P is true in the possible world M’, and Wo is the actual world, then the statement that John knows P is represented by the formula: (I) VW K(John, W,, w) 1 T(w, P), This states that in any world which is compatablc with what John knows in the actual world, P is true. Just as knowledge defines a relation on possible worlds, actions have a similar effect. The predicate CV RWo(A PI> y. w2) rcprcsents the fact that world w2 is rclatcd to world wl by agent A performing action P in wl. Thus possible worlds can bc used in a manner similar to state variables in a state calculus. Using a combination of the K and R predicates, it is possible to develop simple, elegant axiom schemata which clearly state the relationship between an action and how it effects knowledge. For example, it would bc possible, to jxiomatize an informing action with two simple axiom schemata as follows: (3) VW, Vw2 Va Vb VP R(Do(a, Inform(t), P)), wI, w2) 3 [t/w, Ma, wl, w3) 3 T(w3, 01 (4) VW, Vw2 ‘da Vb VP R(Do(a, Inform(b, P), wl, w2) 3 VW, W, w2, w3> 3 3QC(b, wl, w4> ii R(Do(a, Inform(b, P)), w4, w+] Axiom (3) is a precondition axiom. It says that it is true in all possible worlds (i.e. that it is universally known) that when someone does an informing action, he must know that what he is informing is in fact the case. Axiom (4) says that it is universally known that in the situation resulting from an informing act, the hearer knows that the inform has taken place. If the hearer knows that an inform has taken place, then according to axiom (3) he knows that the speaker knew P was true. From making aat deduction, the hearer also knows P. III. USING THE POSSIBLE WORLDS KNOWLEDGE REPRESENTATION IN PIANNING The possible worlds approach to representing facts about knowledge has many advantages, but it presents some problems for the design of a planning system. Axiomatizing knowledge as accessiblity relations between possible worlds makes it possible for a first order logic deduction system to reason about knowledge, but it carries the price of forcing the planner to deal with infinite sets of possible worlds. Planning for someone to know something means making a proposition true in an infinite number .of possible worlds. The goal wff, instead of being a ground level proposition, as is the case in other planning systems developed to date, becomes an expression with an implication and ‘I universally quantified variable, similar to (1). Another problem arises from the way in which actions are axiomatized. l’fficicnt axioms for deduction require the assumption that people can reason with their knowledge. ‘I’hc cffccts of SOITIC actions manifest themselves through this reasoning process. For example, this occurs in the axiomatization of IWORM giVCn in (3) and (4). Speakers usually execute informing actions to get someone to know something. However, the hearer dots not know what the speaker is informing directly as a result of the inform, but from realizing that the speaker has performed an informing act. (See Searlc [7]) ‘I’hc only effect WC know of for an informing act is that the hcarcr knows that the spcakcr has pcrformcd the act. When KAMP has a goal of somconc knowing something, it must be able to dctcrmine somehow that INI:OKM is a rcasonablc thing to do, even though that is not obvious from the effects of the action. A related problem is how to allow for the possiblity that people can reason with their knowlcdgc. If the system has the goal Know(A, Q), and this goal is not achicvcablc directly, and the system knows that A knows that P > Q, then Know(A, P) should be generated as a subgoal. IIxamples of this sort of situation occur whenever Q is some proposition that is not directly obscrvablc. If P is some observable proposition that entails Q, then the planner can perform some action that will result in knowing P, from which Q can be inferred. This is the basis of planning cxpcriments. Since there is a tradeoff between being able to take full advantage of the logical power of the formalism and being able to efficiently construct a plan, a strategy has been adopted that attempts to strike a balance. The strategy is to have a module propose actions for the planner to incorporate into a plan to achieve the current goal. This module can be thought of as a “plausible move generator”. It proposes actions that are likely to succeed in achicvcing the goal. The system then uses its deduction component to verify that the suggested action actually does achieve the goal. To facilitate the action generator’s search for reasonable actions, the preconditions and effects of actions are collected into SrRrPs-like acfiott summaries. Thcsc summaries highlight the direct and indirect effects of actions that are most likely to be needed in forming a plan. For example, the action summary of the INFORM action would include the hcarcr knowing P as one of the (indirect) effects of 1NI;ORM. The effects of actions as they are represented in the action summaries can be any well formed formula. So it is possible to state effects as implications which can match the implications that arc the meta language translations of the Know operator. Using action summaries is not cquivalcnt to recasting the knowledge rcprescntation as a SrRrPS-like system. The action summaries are only used to suggesf alternatives to be tried. To allow the possibility of agents reasoning with their knowledge, KAMP follows the following process whenever a goal involving a knowlcdgc or belief state is encountered: The system tries to invoke an operator that will achieve the goal directly. If this fails, then the system’s base of consequent rules is examined to find subgoals that could be achieved and that would allow the agent to deduce the desired conclusion. Although this approach has the advantage of simplicity, the number of subgoals that the planner has to consider at each step can grow exponentially. It seems intuitively correct to consider subgoals in the order of the length of inference that the agent has to go through to reach the desired conclusion. The planner needs good criteria to prune or at 132 least postpone consideration of less likely paths is an interesting area for further research. of inference. ’ I‘his IV. UNIVERSAL KNOWLEDGE PRECONI)ITIONS When planning actions on a strictly physical level, there are few if any preconditions which can bc said to apply universally to all actions. However, when dealing with the knowledge required to perform an action, as well as its physical enabling conditions, there are a sufficient number of interesting universal preconditions so that their treatment by the planner as a special case is warranted. Universal knowlege preconditions can be summarized by the statement tl?at an agent has to have an executable description of a procedure- in order to do anything. For example, if an agent wishes to perform an INFORM action, it is necessary for him to know what it is that he is informing, how to do informing, and who the intended hearer’ is. Since thcsc preconditions apply ro all actions, instead of including them in the axioms for each action, the planner automatically sets them up as subgoals in every case. V. THE REPRESk:?\‘l-ATION OF THE PLAN WITHIN I’HE SYSTEM The KAMP planner uses a system of plan representation similar to that of Sacerdoti’s procedural networks [6]. The major difference is that CHOICE nodes (OR-SPL.H'S) arc factored out into a disjunctive normal form. Since a choice may occur within an AND- SPLIT that affects how identical goals and actions are cxpandcd after the choice, all the choices arc factored out to the top level, and each choice is treated as a spearate alternative plan to be evaluated independently, (but perhaps sharing some subplans as “subroutines” with other branches.) As in Sacerdoti’s system, nodes can specify goals to be achieved, actions to bc performed, or they may be “ph;lntoms”, goals which coincidentally happen to already be satisfied. Each node of the procedural network is associated with a world. This world represents the real world at that particular stage in the execution of the plan. At the beginning of the plan, the first node is associated with WO. If the first action is that the robot moves from A to B, then the node rcprcscnting that action would have an associated world W1 and there would be an assertion of the form R(Do(Robot, Movc(A, B)), WO, WL) added to the data base. VI. CONTROL STRUCTURE The planner’s control structure is similar to that of Sacerdoti’s NOAH system. A procedural network is created out of the initial goal. The planner then attempts to assign worlds to each node of the procedural network as follows: First, the initial node is assigned Wo, the initial actual world. Then intcratively, when the planer proposes that a subsequent action is performed in a world to reach a new world, a name is generated for the new world, and an R relation between the original world and the new world is asserted in the dcducer’s data base. Then all goal nodes that have worlds assigned are evaluated, i.e. the planner attempts to prove that the goal is true using the, world assigned to that node as the current state of the actual world. Any goal for which the proof succeeds is marked as a phantom (achicvcd) goal. Next, all the unexpanded nodes in the network that have been assigned worlds, and which are not phantoms, are examined. Some of them may bc high level actions for which a procedure exists to determine the appropriate expansion. These procedures are invoked if they exist, otherwise the node is an unstaisfied goal node, and the action generator is invoked to find a set of actions 133 PI PI [31 [41 151 Kl [71 REFERENCES Appclt, Douglas E., Problem Solving Applied to Nutural Larlgrrage Generafion, proceedings of the Annual Confcrcnce of the Association for Computational Linguistics, 1980. Cohen, Philip, On Knowing Whal 10 Say: Planning Speech Acts, University of Toronto Technical Report #LB, 1978. Fikes, Richard E., and N. Nilsson, STRIPS: A New Approach to the Application of Theorem Proving to Problem Solving, Artificial Intelligence, No. 2, 1971. Konoligc, Kurt and N. Nilsson, Planning in n Mulfiple Agent Environment, Proceedings of the First Annual Conference of the American Association for Artificial Intelligence, August, 1980. Moore, Robert C., Reasoning About Knowledge and Action, Massachusetts Jnstitute of Technology Artificial Intelligence Laboratory Technical Report ‘I’R-???, 1979. Saccrdoti, Earl, A Structure for Plans and 13chaviol; Elscvier --- North-Holland, Inc., Amsterdam, The Netherlands, 1977. Scarle, John, Sgccch Acts, Cambridge University Press, 1969. which might be performed to achieve the goal. If an action is found, it is inscrtcd into the procedural network along with its preconditions, both the universal ones and those specific to the particular action. After the nodes are expanded to the next level, a set of critic procedures are invoked, which can examine the plan for global interactions and take corrective action if needed. This cntirc process is rcpeatcd until the plan has been expanded to the point where every unexpanded node is either a phantom goal or an executable action. VII. CURRENT STATUS OF RESKARCH The KAMP planning system has been implemented and tested on several examples. It has been used to solve the problem of John, Rob and the clock cited earlier in the paper. All the critics of Sacerdoti’s NOM1 have either been implemented or are currently uncicrgoing implcmcntation. Further development of the planner will be dictated by the needs of a language planning and generation system currently under development. It is expected that this langauge generation task will make full use of the unique features of this system. KAMP is a first attempt at dcvcloping a planner that is capable of using the possible worlds semantics approach to representing knowledge about belief. Combining a planner with a very powerful knowledge representation will enable problem solving tcchniqucs to bc applied to a variety of domains such as language generation, and planning to acquire and distribute knowledge, in which they have played a relatively small role in the past. ACKNOWLEDGEMENTS The author is grateful to Barbara Cross, Gary Hendrix and Terry Winograd for comments on earlier drafts of this paper	1980	73
71	Making Judgments Hans J. Berliner Computer Science Department Carnegie-Mellon University Pittsburgh, Pa. 15213 Abstract Reasoning-based problem solving deals with discrete entities and manipulates these to derive new entities or produce branching behavior in order to discover a solution. This paradigm has some basic difficulties when applied to certain types of problems. Properly constructed arithmetic functions, such as those using our SNAC principles, can do such problems very well. SNAC constructions have considerable generality and robustness, and thus tend to outperform hand coded case statements as domains get larger. We show how a SNAC fimction can avoid getting stuck on a sub-optimal hill while hill-climbing. A clever move made by our backgammon program in defeating the World Champion is analyzed to show some aspects of the method. 1 Int reduction Problem solving research and examples usually deal with sequential reasoning toward a conclusion or required response. For such situations, criteria exist that make it possible to identify the correct response and possibly order other responses with respect to their goodness. However, in most domains such a paradigm is not possible because the number of states in the domain is so large that it is next to impossible to describe the properties of an arbitrary state with sufficient accuracy to be able to reason about it. Expertise in such domains appears to require judgment. We consider judgment to be the ability to produce graded responses to small changes in the stimulus environment. In judgment domains several responses may be considered adequate, while reasoned decisions would appear to only be correct or incorrect. The ability to reliably judge small differences in chess positions is what separates the top players from their nearest competitors. Even though a decision procedure exists for determining whether one position is better than another, it is intractable. It is this intractability or the inability to isolate features that can be used in a clear reasoning process that distinguishes the judgment domain from the reasoning domain. The boundary between the two is certainly fuzzy, and undoubtedly changes as new information about any particular domain is developed. It seems that the larger the domain and the less precise the methods of making comparisons between elements of the domain, the less adequate are reasoning techniques. 2 The Problem There are a number of techniques available to allow a program to make comparisons, i.e. to discriminate good from bad from indifferent in selecting among courses of action and among potential outcomes. However, while these techniques are fine for doing simple comparisons, most of them break down with even small additional complexity. Consider the syllogism: 1) The more friends a person has, the happier he is. 2) John has more friends than Fred. Therefore: John is happier than Fred. So far so good. However, adding just a small amount of complexity with the two additional propositions: 3) The more money a person has, the happier he is. 4) Fred has more money than John. makes it possible to derive two contradictory conclusions from the premises. This is a most unsatisfactory state of affairs. Especially so, since recoding the premises into first order predicate calculus does not help either. Neither will using productions or the branching logic of programming languages. For such rcprcsentations, the most likely formulation would be that X will be happier than Y #he is superior in This research was sponsored by the Defense Advanced Research Projects Agency (DOD), ARPA Order No. 3597, monitored by the Air Force Avionics Laboratory Under Contract F33815-78-C-1551. 134 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. all applicable categories. Another formulation would have X happier than Y if he is superior in a majority of categories (with a tie being undefined). Such “voting” techniques can be shown to be deficient if We further increase the complexity of the decision that is to be made. If premises 2 and 4 were restated as: 2a) John has 25 friends and Fred has 20. 4a) Fred has $20,000 and John has $500. Most people would agree that Fred was happier according to our definitions of happiness. Yet, the only machinery available for coming to grips with problems such as this in systems that reason is to produce a large number of additional axioms that contain compound conditions, or to define degrees of difference so that degrees of happiness can be ascertained and summed. The world of reasoning is a world of quantized perception and action. These systems are discrete and do business on a “case” basis. In order to achieve expertise it is necessary to react differentially to states that were formerly considered equivalent. Thus, the number of distinct perceptions and actions gets larger with expertise. This makes it more expensive to find the applicable rule or pattern, and creates difficulty in keeping new rules from interfering in unintended ways with the effects of older rules. Further, the possibility that more than one pattern will match grows as complexity grows, and default conditions are usually defined for states that fail to match specific rules or patterns. This makes adding new knowledge a formidable task for even moderate size domains [3]. So unless, a method is found for automatically appending viable rules to such a system, there seems to be a definite limit on the expertise it can achieve. Because it is easier to pay attention to only a few things at one time, reasoning systems seem to have more of a sub-optimization nature than is necessary in sequential problem solving. The need to solve the top level goal can obscure the fact that it could possibly by solved later with greater facility. For instance, a plan for taking a trip by car could include: 1. Get suitcase 2. Pack clothes in suitcase 3. Put suitcase in car If the raincoat is already in the car, this would involve getting it from the car only to bring it back later inside the suitcase. Conceivably, it would be simpler to bring the packed suitcase to the car and put the raincoat inside it at that time. This shows that goals need not have an immutable hierarchy. Further, there are times when achieving several low level goals is more desirable than achieving the top level goal. In addition to the above there is another problem that exists in domains that interface to the real world, where sensed parameters that have a quasi-continuous character may have to be quantized. Premature quantization of variables loses information and can cause problems when the variable is to be used later for making decisions. For instance, if day/night is a binary variable and it is advantageous to be in day, a program may arrange its problem solving behavior so that it samples the environment just before day turns to night (by the system definition), and, being satisfied with what it finds, pronounces this branch of the solution search as favorable. If it had been forced to continue the branch even a few steps, if would have come to a different conclusion as night was closing in. However, quantization of the relatively continuous day/night variable causes the blenrish effect [2], a behavior anomaly similar to the horizon effect [l], but with the step size of the variable rather than the fixed depth of the search being the culprit. This problem can be prevented by retaining a variable in its quasi-continuous state as long as possible. However, if a variable has a very large range it is impractical to create tests for each value in the range. Resorting to the testing of sub-ranges merely recreates the problem. Thus, discrete treatment of such a variable can cause problems, no matter how it is done. 3 A Better Way Arithmetic functions can do all the above things easily and cheaply if they are constructed in the right way. A polynomial of terms that represent important features in the domain is constructed. We have described our SNAC method of constructing such polynomials and shown [2, 41 that: e It is important that the values of terms vary smoothly. o Non-linearity of terms is extremely important for expertise. e Some method must exist for determining the degree to which each feature is applicable in the present situation. This is done with slowly varying variables that we call application coefficients. The SNAC method also makes it possible to avoid the previously vexing problem of getting stuck on a sub-optimal hill while hill-climbing. Figure 1 shows how getting stuck on a hill is avoided. With non-linear functions, the peaks of hills can be rounded so that retaining the peak becomes less desirable, especially if some other high ground is in view of the searching process. Further, with application coefficients it is possible to change the contour of the hill even as it is being climbed. This is shown in a - c; the arrow showing the location of the current state. As the hill is being climbed, one or more application coefficients that sense the global environment cause the goal of achieving the hilltop to become less important since it is very near being 135 ' * 3 4 5 6 Black Figure 1: Effect of SNAC on Hill Shape achieved. The change in value of the application coefficients causes the contour of the hill to begin to flatten, making the achievement of the summit less desirable, and resulting in the program looking for the next set of goals before even fully achieving the current set. Thus application coefficients can direct progress by reducing the importance of goals that are near being achieved, have already been achieved, or arc no longer important. The above is achieved mathematically as follows: The function p + Y* = C* for -C 2 X 5 C and Y> 0 will produce a semi-circle similar to Figure la. If we now change the function to be >+ A* p= Ce where A 2 1 is an application coefficient (a variable), we can flatten the semi-circle into a semi-elipse of arbitrary flatness. Here, let OLDX be OLDX increases in value. The construction is finalized by only recognizing values of A while OLDX is in the range of (say) -2C to the hill would never seem a desirable thing to do because the program could not tell the difference between getting there when it was far away or already very close. 4 An Example of SNAC Sensitivity The backgammon position in Figure 2 occurred in the Anal game of the match in which my program, BKG 9.8, beat World Champion Luigi Villa in July, 1979. In this position, BKG 9.8 had to play a $1. There are four factors that must be considered here: 1. Black has established a strong defensive points made on the 20 and 22 points. backgame position with 2. In backgame positions timing is very important. Black is counting on hitting White when he brings his men around and home. At such time he must be prepared to contain the man sent back. This can only be done if the rest of his army is not too far advanced so it can form a containing pocket in front of the sent-back man. At the moment Black would not mind having additional men sent back in order to delay himself further and improve his timing. 24 23 22 21 20 19 White 18 17 16 15 14 13 Figure 2: Black to Play a 5,l 3. There is also a possibility that Black could win by containing the man that is already back, but this is rather slim since White can escape with any 5 or 6. However, blockading this man is of some value in case White rolls no 5’s or 6’s in the near future. 4. In case the sole White back man does not escape, there is a possibility of bringing up the remainder of Black’smen not used for the backgame and trying to win with an attack against the back man. In view of the above it is very difficult to determine the right move, and none of the watching experts succeeded in finding it. The most frequently mentioned move was 13-7, which brings a man into the attack and hopes he is hit so as to gain timing (delay one’s inevitable advance). However, BKG 9.8 made the better move of playing 13-8, 3-2, breaking up its blockade somewhat in order to get more attack, and attempting to actively contain the White back man. It did not worry about the increased chance of being hit, as this only helps with later defense. This gives the program two chances to win: If the attack succeeds, and by getting more men sent back, if the attack fails it improves the likelihood of success of its backgame. I have not seen this concept in this form before in books or games. Humans tend to not want to break up the blockade that they have painstakingly built up, even though it is now the least valuable asset that Black has. It is instructive to see how the program arrived at the judgment it made; one that it had never been tested for. Black has 28 legal moves. Ibe top choices of the program were (points of the Scoring pOlYnOmid in parentheses): 13-8, 3-2 (687); 10-5, 3-2 (682); 13-8, 10-9 (672); and 136 13-7 (667). The third and fourth choices were the ones most frequently mentioned by watching experts, thus showing they missed the idea of breaking up the blockade; the thing common to the program’s top two choices. Let us see why it judged the move actually played as better than the third choice (12-17, 15-16). The program considers many factors (polynomial terms) in its judgments and quite a few of these are non-linear. The six factors on which the two moves differed were (points for each and difference in parentheses): 1. Containment of enemy man (177, 131, +46). The move made does hinder his escape more. Containment is always desirable unless one is far ahead and the end of the game is nearing. 2. Condition of our home board (96, 110, -14). It breaks up one home board point. Breaking up the board (points 1 thru 6) is never considered desirable. 3. Attack (37, 21, +16). It is the best attacking move. Attack is desirable unless we are jeopardizing a sure win in the process. 4. Defensive situation (246,260, -14). The move slows White down, thus could reduce the effectiveness of the backgame. 5. Long-term positional (-2, 11, -13). It puts a man on the 2 point, which is undesirable when the game still has a long way to go because it is too far%dvanced to be able to influence enemy men from there. 6. Safety of men (-12, -4, -8). The move made is dangerous. The program realizes this, but also understands that with a secure defensive position such danger is not serious. However, all other things being equal, it would prefer the least dangerous move Thus the better containment and attack are consider-cd to be more important than the weakening of the homcboard, the temporary slowing down of White, the long-term positional weakness, and the safety of the men. The difference between the first and second choice was that in the first choice the attack is slightly stronger. The importance of each of the above terms varies with the situation. In the example, a backgame is established; else the safety term would outweigh the attack term, and BKG 9.8 would not leave two blots in its home board. It does recognize the degree of danger, however, and will not make a more dangerous move unless it has compensating benefits. This is typical of the influence that application coefficients exert in getting a term to respond to the global situation. 5 Perspective We have been employing the SNAC method of making judgments for over two years now, and are struck with its simplicity and power. The happiness example posed earlier is solved trivially in all its forms with SNAC. If the above travel planning problem were solved as a search problem using SNAC functions that measure the economy of effort of the steps used, then undoubtedly SNAC would also do better than sequential planning based on rules, with no evaluation of outcome other than success or failure. At the moment it is difficult to determine what role, if any, SNAC like mechanisms have in human thinking. We have constructed them to simulate lower level “intuitive” type of behavior, and they appear to work admirably in capturing good judgment in the large domain of backgammon. We conjecture that as variables become more and more discrete in character and as criteria for success become more obvious, reasoning gradually replaces such judgment making. At present our backgammon program is being modified to be able to interpret its own functions with the aim of being able to explain its actions, and ultimately being able to identify its failures by type and modifying the culprit functions. 111 PI [31 [41 References Berliner, H. J. Some Necessary Conditions for a Master Chess Program. In Third Inlernational Joinl Conference on Arlificial Intelligence, pages 77-85. IJCAI, 1973. Berliner, H. On the Construction of Evaluation Functions for Large Domains. In Sixth International Joinr Conference on Arbjicial Intelligence, pages 53-55. IJCAI, 1979. Berliner, H. Some Observations on Problem Solving. In Proceeding of Ihe Third CSCSI Conference. Canadian Society for Computational Studies of Intelligence, 1980. Berliner, H. J. Backgammon Computer Program beats World Champion. Arlificial Intelligence 14(l), 1980. 137	1980	74
72	ABSTRACT MULTIPLE-AGENT PLANNING SYSTEMS Kurt Konolige Nils J. Nilsson SRI International, Menlo Park, California We analyze problems confronted by computer agents that synthesize plans that take into account (and employ) the plans of other, similar, cooperative agents. From the point of view of each of these agents, the others are dynamic entities that possess information about the world, have goals, make plans to achieve these goals, and execute these plans. Thus, each agent must represent not only the usual information about objects in the world and the preconditions and effects of its own actions, but it must also represent and reason about what other agents believe and what they may do. We describe a planning system t??at address es these is show how it solves a sample problem . sues and INTRODUCTION Certain tasks can be more advantageously performed by a system composed of several "loosely coupled," cooperating artificial intelligence (AI) agents than by a single, tightly integrated system. These multiple agents might be distributed in space to match the distributed nature of the task. Such systems are often called distributed artificial intelligence (DAI) systems Ll]. We are interested here in systems where the component agents themselves are rather complex AI systems that can generate and execute plans, communicate with each other. make infe rences, and Among the poten tial advantages of such DA1 systems are graceful (fail-soft) degradation characteristics (no single agent need be indispensable), upward extensibi lity (new agents can be added without requiring major system redesign), and communication efficiency (a message- sending agent can plan its "communication acts" carefully, taking into account the planning and inference abilities of the receiving agents). In planning its actions, each agent must consider the potential actions of the other agents. Previous AI research on systems for generating and executing plans of actions assumed a single planning agent operating in a world that was static except for the effects of the actions of the planning agent itself. Examples of such systems include STRIPS [2], NOAH [S], and NONLIN [4]. Several important extensions must be made to planning systems such as these if they are to function appropriately in an environment populated by other planning/execution systems. First, each agent must be able to represent certain features of the other agents as well as the usual information about static objects in the world. Each agent must have a representation for what the other agents "believe" about themselves, the world, and other agents. Each agent must have a representation for the planning, plan-execution, and reasoning abilities of the other agents. These requirements presuppose techniques for representing the "propositional attitudes" believe and want. Second, among the actions of each agent are ~communication actions" that are used to inform other agents about beliefs and goals and to request information. Finally, each agent must be able to generate plans in a world where actions not planned by that agent spontaneously occur. We introduce here the notion of spontaneous operators to model such actions. In this paper, we give a brief summary of our approach toward building DA1 systems of this sort. It should be apparent that the work we are describing also has applications beyond DAI. For example, our multiple agents plan, execute, and understand communication acts in a manner that could illuminate fundamental processes in natural- language generation and understanding. (In fact, some excellent work has already been done on the subject of planning "speech acts" [5-61.) Work on models of active agents should also contribute to more sophisticated and helpful "user models" for interactive computer systems. The development of multiagent systems might also stimulate the development of more detailed and useful theories in social psychology--just as previous AI work has contributed to cognitive psychology. At this early stage of our research, we are not yet investigating the effects of differing "social organizations" of the multiple agents. Our work to date has been focussed on representational problems for such systems independent of how a society of agents is organized. A MULTIPLE-AGENT FORMALISM Each agent must be able to represent other agents' beliefs, plans, goals, and introspections about other agents. Several representational formalisms might be used. McCarthy's formalism for first-order theories of individual concepts and From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. propositions [7] is one possibility, although certain problems involving quantified expressions in that formalism have not yet been fully worked out. Another candidate is Moore's first-order axiomatization of the possible world semantics for the modal logic of knowledge and action [8-g]. Appelt [lo] h as implemented a system called KAMP that uses Moore's approach for generating and reasoning about plans involving two agents. We find Moore's technique somewhat unintuitive, and it seems needlessly complex when used in reasoning about ordinary (nonattitudinal) propositions. Here we develop a representation for each agent based on Weyhrauch's notion of multiple first-order theories and metatheories [ll]. Al: Facts --w-e HOLDING(Al,A) HOLDING(Al,B) CLEAR(B) A pickup(Al,B) HANDEMPTY(A1) CLEAR(B) / putdown(A1,A) HOLDIiG(Al,A) Using Weyhrauch's termi nology, each computer individual is defined by the combination of a first-order language, - a simulation structure or partial model for that language, a set of Facts (expressed in the language), and a Goal Structure that represents a goal for the agent and a plan for achieving it. We assume that each agent has a deductive system (a combination of a theorem-prover and attached procedures defined by the simulation structure) used for deriving new facts from the initial set and for attempting to determine whether goals and subgoals follow from the set of facts. Each agent is also assumed to have a planning system (such as STRIPS) for creating plans to achieve goals. Using a typical "blocks-world" example, we diagram an agent's structure in the following way: Al (agent's name): Facts ----- Goal --s-e HOLDING(Al,A) HOLDING(Al,B) CLEAR(B) Viewed as a computational entity, an agent's structure is typically not static. Its deductive and sensory processes may expand its set of facts, or its planning system may create a plan to achieve a goal. Also, once a plan exists for achieving a goal, the agent interacts with its environment by executing its plan. In this summary, we deal only with the planning processes of agents. In the example above, the occurrence of the goal HOLDING(Al,B) in Al's goal structure triggers the computation of a plan to achieve it. Once generated, this plan is represented in the goal structure of agent Al as follows: Goal m---w Plans are represented by goal/subgoal trees composed of planning operators and their preconditions. We assume a depth-first ordering of the operators in the plan. Now let us introduce another agent, AO. Agent A0 can have the same sort of structure as Al, including its own first-order language, a description of the world by wffs in that language (facts), a simulation structure, a goal structure, and a planner and deducer. Some of AO's facts are descriptions of Al's structure and processes. By making inferences from these facts, A0 can reason about the planning and deductive activities of Al and thus take Al into account in forming its own plans. Also, a structure similar to Al's actual structure, and procedures similar to Al's deducer and planner, can be used as components of AO's simulation structure. Procedural attachment to these "models" of Al can often be employed as an alternative method of reasoning about Al. (Of course, A0 may have an incomplete or inaccurate model of Al.) Because Al's structure is a first-order language (augmented by certain other structures), A0 can use a formal metalanguage to describe it along the lines suggested, for example, by Kleene [12] and developed in FOL by Weyhrauch [Ill. A0 has terms for any sentence that can occur in Al's language or for any of Al's goals or plans; the predicates FACT and GOAL are used to assert that some sentences are in Al's Facts list or goal structure. Consider the following example: assume A0 is holding block A, block B is clear, and A0 believes these facts and further believes that Al believes A0 is holding A and that Al believes B is not clear. A0 would have the following structure: AO: Facts ----- HOLDING(AO,A) CLEAR(B) ;;;;;;;,'HOLDING(AO,A)') /-CLEAR(B) '> We use quote marks to delimit strings, which may have embedded string variables. The denotation of a ground string is the string itself. Thus, the intended interpretation of FACT(Al,'HOLDING(AO,A)') is that the wff HOLDING(AO,A) is part of the facts 139 list of Al (that is, Al "believes" that A0 is holding A). By using the FACT predicate and terms denoting other agents and wffs, any facts list for other agents can be described. (A0 can describe its own beliefs in the same manner.) We purposely use "believe" instead of "know" because we are particularly interested in situations where agents may be mistaken in their representation of the world and other agents. In the above example, AO's opinions about its own and Al's belief about whether or not block B is clear are inconsistent. We avoid formalizing "know" and thus do not take a position about the relationship between knowledge and belief (such as "knowledge is justified true belief"). We can describe some of the usual properties of belief by axioms like FACT(x,p) => FACT(x,'FACT(x,p)'); i.e., if an agent believes p, it believes that it believes p. We do not, however, use an axiom to the effect that agents believe the logical consequences of their beliefs, because we want to admit the possibility that different agents use different procedures for making inferences. In particular, we want to emphasize that the deductive capabilities of all agents are limited. While the static structure of Al is described, for AO, by FACT and GOAL predicates, the action of Al's deductive system and planner can also be axiomatized (for AO) at the metalevel (see Kowalski [13] for an example). This axiomatization allows A0 to simulate Al's deducer or planner by purely syntactic theorem-proving. Thus A0 might use predicates such as ISPROOFfx,p) and ISPLAN(x,p) to make assertions about whether certain proof or plan structures are proofs or plans for other agents (or for itself). In certain cases, A0 can find out if Al can deduce a particular theorem (or if Al can create a plan) by running its procedural model of Al's deducer (or planner) directly, rather than by reasoning with its own facts. This is accomplished by semantic attachments of models of Al's deducer and planner to the predicates ISPROOF and ISPLAN in AO's metalanguage. Semantic attachment thus allows A0 to "think like Al" by directly executing its model of Al's planner and deducer. (Here, we follow an approach pioneered by Weyhrauch [ll] in his FOL system of using semantic attachments to data structures and programs in partial models.) The same kind of attachment strategy can be used to enable A0 to reason about its own planning abilities. The usual problems associated with formalizing propositional attitudes [8,14] can be handled nicely using the FACT predicate. For example, the atomic formula FACT(Al,'CLEAR(A) v CLEAR(B)') asserts the proposition that Al believes that A is clear or B is clear, and is not confused with the formula [FACT(A~,'CLEAR(B)') v FACT(Al,'CLEAR(A)')], which asserts the different proposition that Al believes that A is clear or Al believes that B is clear. Furthermore, semantic attachment methods confer the advantages of the so- called "data base approach" [8] when appropriate. Of particular importance among statements concerning AO's beliefs about Al are those that involve "quantifying in," i.e., where a quantified variable appears inside the term of a FACT predicate. We follow the general approach of Kaplan [15] toward this topic. For example, the sentence (Ex)FAOT(A~,'H~LDING(A~,X)') occurring among AO's facts asserts that Al is holding an identified (for Al) block without identifying it (for AO). AN EXAMPLE We can illustrate some of the ideas we are exploring by a short example. Suppose that there are two agents, A0 and Al, each equipped with a hand for holding blocks. Initially Al is holding a block, A, and A0 wants to be holding A. Suppose that A0 believes these initial facts, but (to make our example more interesting) A0 has no information about whether or not Al itself believes it is holding A. Thus, the initial structure for A0 is: AO: Facts Goal ------ me--- HANDEMPTY(A0) HOLDING(AO,A) HOLDING(Al,A) Let us assume the following planning operators (for both A0 and Al). We use standard STRIPS notation [16]. ('P&D' denotes the precondition and delete lists; 'A' denotes the add list.) putdown(x,b) agent x puts block b on the table P&D: HOLDING(x,b) A: ONTABLE & CLEAR(b) & HANDEMPTY pickup(x,b) agent x picks up block b P&D: CLEAR(b) & HANDEMPTY A: HOLDING(x,b) asktoachieve(x,y,g) agent x gives agent y the P: T goal denoted by string g A: GOAL(y,g) tell(x,y,s) agent x tells agent y the P: FACT(x,s) expression denoted by string s A: FACT(y,s) Agents take into account the possible actions of other agents by assuming that other agents generate and execute plans to achieve their goals. The action of another agent generating and executing a plan is modelled by a "spontaneous operator." A spontaneous operator is like an ordinary planning operator except that whenever its preconditions are satisfied, the action corresponding to it is presumed automatically executed. Thus, by planning to achieve the preconditions of a spontaneous operator, a planning agent can incorporate such an operator into its plan. Let us assume that agent A0 can use the operator "achieve" as a spontaneous operator that models the action of another agent generating and executing a plan: achieve(x,g) agent x achieves goal g by creating and executing a plan to achieve g. PC: GOAL(x,g) & ISPLAN(x,p,g,x) & ISPLAN(x,p,g,AO) D: the delete list is computed from the plan, p A: FACT(x,g) FACT(AO,g) The expression ISPLAN(x,p,g,f) is intended to mean that there is a plan p, to achieve goal g, using agent x‘s planner, with facts belonging to agent f. Our precondition for achieve ensures that before A0 can assume that condition g will be spontaneously achieved by agent x, A0 has to prove both that agent x can generate a plan from its own facts and that agent x could generate a plan from AO's facts (to ensure that the plan is valid as far as A0 is concerned). Here are some axioms about ISPLAN that A0 will need: 1) FACT(x,w) => ISPLAN(x,NIL,w,x) (If x already believes w, then x has a plan, namely NIL, for achieving w.) 2) [ISPLAN( X,U,Y,d & PC(Z,Y) & OP(x,z,g)] => ISPLAN(x,extend(u,z),g,x) (If x has a plan, namely u, to achieve the preconditions, y, of its operator, z, with add list containing g, then x has a plan, namely extend(u,z) for achieving g. The functional expression, extend(u,z), denotes that plan formed by concatenating plan z after plan u.) The planning tree in Figure 1 shows a possible plan that A0 might generate using its facts (including axioms about ISPLAN) and operators. The sequence of operators in this plan is {tell(AO,Al,'HOLDING(Al,A)'), asktoachieve(AO,Al,'CLEAR(A)'), achieve(Al,'CLEAR(A)'), pickup(AO,A)j. Note that semantic attachment processes were used in several places in generating this plan. We leave to the control strategy of the system the decision about whether to attempt to prove a wff by semantic attachment or by ordinary syntactic methods. We are now in the process of designing a system for generating and executing plans of this sort. Space prohibits describing some additional features of our system, including its control strategy for generating plans. We plan to experiment with a,, complex of several agents, each incorporating planning systems like that briefly described here. We gratefully acknowledge helpful discussions with Doug Appelt, Bob Moore, Earl Sacerdoti, Carolyn Talcott and Richard Weyhrauch. This research is supported by the Office of Naval Research under Contract No. N00014-80-C-0296. 1 . 2. 3. 4. 5. 6. 7. 8. 9. 10. REFERENCES Sacerdoti, E. D., "What Language Understanding Research Suggests about Distributed Artificial IntelligenceT,, in Distributed Sensor Nets, ppm 8-11. Paper presented at the DARPA Workshop, Carnegie-Mellon University, Pittsburgh, Pennsylvania (December 7-8, 1978). Fikes, R. E. and N. J. Nilsson, "STRIPS: A New Approach to the Application of Theorem Proving to Problem Solving,,, Artificial Intelligence, 2(3/4), pp. 189-208 (1971). Sacerdoti, E. D., A Structure for Plans and Behavior, (New York: Elsevier, 197r Tate, A., "Generating Project Networks,,, in IJCAI-5, pp. 888-893 (1977). Searle, J. R., "A Taxonomy of Illocutionary Acts," in Language Mind and Knowledge, K. Gundersm,(University of Minnesota Press, 1976). Cohen, P. R. and C. R. Perrault, "Elements of a Plan-Based Theory of Speech Acts,,, Cognitive Science, _ 3(3), pp. 177-212 (1979)e McCarthy, J., "First Order Theories of Individual Concepts and Propositions,,, in Machine Intelligence 9, pp.-120-147, -ayes and D. Michie (Eds.), (New York: Halsted Press, 1979). Moore, R. C., "Reasoning About Knowledge and Action,,, in IJCAI-5, pp. 223-227 (1977)' Moore, R. C., "Reasoning About Knowledge and Action,,, Artificial Intelligence Center Techincal Note 191, SRI International, Menlo Park, California'(lY80). 141 13. Kowalski, R., Logic for Problem Solving, (New 15. Kaplan, D., "Quantifying In," in Reference and York: North-Hom,T'79). Modality, L. Linsky (Ed.), pp. 112-144, (London: Oxford University Press, 1971). 14. Quine, W. V. O., "Quantifiers and Propositional Attitudes," in Reference and 16. Nilsson, N. J., Principles of Artificial Modality, L. Linsky (Ed.), pp. 101-111, Intelligence, (Menlo Park: Eoga Publishing (London: Oxford University Press, 1971). co., 1980). Al: HOLDING(AO,A) bkup(AO,A) HANDEMPTY(A0) (initial fact) achieve(Al,'CLEAR(A)') plan axiom {extend(u,z)/pj ISPLAN(Al,p,'CLEAR(Al)',AO) (verified by proc. attach. to a model of Al's planner using AO's facts after substituting for p) GOAL(Al,'C EAR(A)') [ISPLAN(Al,u,y,Al) & PC(z,y) & OP(Al,z,'CLEAR(A)')] asktoachieve(AO,Al, I 'CLEAR(A)') proc. attach. to PC and OP x {'putdown(Al,A)'/z, 'HOLDING(AI,A)'/y~ T ISPLAN(Al,u,'HOLDING(Al,A)',Al) I plan axiom b'IL/uj FACT(Al,'HOLDING(Al,A)') I tell(AO,A1,'HOLDING(A1,A)') FACT(AO,'HOLDING(Al,A)') (verified by proc. attach. to AO's "fact finder".) Figure 1 ' 142	1980	75
73	SCOUT: A SIMPLE GAME-SEARCHING ALGORITHM WITH PROVEN OPTIMAL PROPERTIES Judea Pearl Cognitive Systems Laboratory School of Engineering and Applied Science University of California Los Angeles, California 90024 ABSTRACT This paper describes a new algorithm for searching games which is conceptually simple, space efficient, and analytically tractable. It pos- sesses optimal asymptotic properties and may offer practical advantages over a-6 for deep searches. I. INTRODUCTION We consider a class of two-person perfect information games in which two players, called MAX and MIN, take alternate turns in selecting one out of d legal moves. We assume that the game is searched to a depth h, at which point the terminal positions are assigned a static evaluation function VO* The task iS to evaluate the minimax value, Vh, of the root node by examining, on the average the least number of terminal nodes. SCOUT, the algorithm described in this paper, has evolved as a purely theoretical tool for ana- lyzing the mean complexity of game-searching tasks where the terminal nodes are assigned random and independent values [l]. With the aid of SCOUT we were able to show that such games can be evaluated with a branching factor of P/(l-P), where P* is the root of xd+x-1 = 0, and that no directional algorithm (e.g., ALPHA-BETA) can do better. We have recently tested the performance of SCOUT on a 'real' game (i.e., the game of Kalah) and were somewhat surprised to find that, even for low values of h, the efficiency of SCOUT surpasses that of the a-6 procedure [2]. The purpose of this paper is to call attention of game-playing practi- tioners to the potentials of SCOUT as a practical game-searching tool. Section II describes the operation of SCOUT in conceptual terms avoiding algorithmic details. Section III presents, without proofs, some of the mathematical properties of SCOUT and compares them to those of the a-~ procedure. Finally empirical results are reported comparing the performances of SCOUT and a-6 for both random and dynamic orderings. * Supported in part by NSF Grants MCS 78-07468 and MCS 78-18924. II. THE SCOUT ALGORITHM SCOUT invokes two recursive procedures called EVAL and TEST. The main procedure EVAL(S) returns V(S), the minimax value of position S, whereas the function of TEST(S, v, >) is to validate (or refute) the truth of the inequality V(S) > v where v is some given reference value. Procedure: TEST(S, v, >) To test whether S satisfies the inequality V(S) > v, start applying the same test (calling itself) to its successors from left to right: If S is MAX, return TRUE as soon as one suc- cessor is found to be larger than v; return FALSE if all successors are smaller than or equal to v. If S is MIN, return FALSE as soon as one suc- cessor is found to be smaller than or equal to v; return TRUE if all successors are larger than v. An identical procedure , called TEST(S, v, z), can be used to verify the inequality V(S) 1 v, with the obvious revisions induced by the equality sign. Procedure: EVAL(S) EVAL evaluates a MAX position S by first eval- uating (calling itself) its left most successor S,, then 'scouting' the remaining successors, from left to right, to determine (calling TEST) if any meets the condition V(Sk) > V(S1). If the inequal- ity is found to hold for Sk:, this node is then evaluated exactly (calling EVAL(Sk)) and its value V(Sk) iS used for subsequent 'Scoutings' tests. Otherwise Sk iS exempted from evaluation and Sk+1 selected for a test. When all successors have been either evaluated or tested and found unworthy of evaluation, the last value obtained is issued as V(S). An identical procedure is used for evaluating a MIN position S, save for the fact that the event V(Sk) 1 V(S1) now constitutes grounds for exempt- ing S from evaluation. Flow-charts describing both 6 COUT and TEST in algorithmic details can be found in [l]. At first glance it appears that SCOUT is very wasteful; any node Sk which is found to fail a test criterion is submitted back for evaluation. The terminal nodes inspected during such a test may 143 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. (and in general will) be revisited during the eval- uation phase. An exact mathematical analysis, however, reveals that the amount of waste is not substantial and that SCOUT, in spite of some dupli- cated effort, still achieves the optimal branching factor P/(l-P), as will be demonstrated in Sec- tion III. Two factors work in favor of SCOUT: (1) most tests would result in exempting the tested node (and all its descendents) from any further evalua- tion, and (2) testing for inequality using the TEST(S, v) procedure is relatively speedy. The speed of TEST stems from the fact that it induces many cutoffs not necessarily permitted by EVAL or any other evaluation scheme. As soon as one suc- cessor of a MAX node meets the criterion u(s ) > v, all other successors can be ignored. EVAL, y I: contrast, would necessitate a further examination of the remaining successors to determine if any would possess a value higher than V(Sk). Several improvements could be applied to the SCOUT algorithm to render it more efficient. For example, when a TEST procedure issues a non-exempt verdict, it could also return a new reference value and some information regarding how the decision was obtained in order to minimize the number of nodes to be inspected by EVAL. However, the analysis presented in Section III, as well as the simulation tests, were conducted on the original version described above. These studies show that, even in its unpolished form, SCOUT is asymptotically opti- mal over all directional algorithms and is somewhat more efficient than the a-6 procedure for the game tested (i.e., Kalah). Recently, Stockman [3] has also introduced an algorithm which examines fewer nodes than a-6. However, Stockman's algorithm requires an enormous storage space for tracing back a large number of potential strategies. SCOUT, by contrast, has storage requirements similar to those of a-6; at any point in time it only maintains pointers along one single path connecting the root to the current- ly expanded node. III. ANALYSIS OF SCOUT'S EXPECTED PERFORMANCE In this section we present, without proofs, some mathematical results related to the expected number of nodes examined by SCOUT and a-B. Addi- tional results, reference [l]. including proofs, can be found in The model used for evaluating these algorithms consists of a uniform tree of height h (h even) and branching factor d, where the terminal positions are assigned random values, independently drawn from a common distribution F. We shall refer to such a tree as a (h, d, F)-tree. Theorem 1: The root value of a (h, d, F)-tree with continuous strictly increasinq terminal dis- tribution F converges, as h -+ CQ (in probability) to the 1-P)-fractile of F, where P is the solution 6 of x +x-l = 0. If the terminal values are discrete: v, < v2 < . . . < VM, then the root value converges to a definite limit iff l-P* # F(v.) for all i, in which case the limit is the smallelt vi satisfying l-P* < F(Vi). Definition: Let A be a deterministic algo- rithm which searches the (h, d, F)-game and let IA(h, d, F) denote the expected number of terminal positions examined by A. The quantity: r&b F) = 'im [IA(h, d, F)]l'h h-too is called the branching factor corresponding to the algorithm A. Definition: Let C be a class of algorithms capable of searching a general (h, d, F)-tree. An algorithm A is said to be asymptotically optimal over C if for all d, F, and BEC, rA(& F) 5 rg(d, I=). Definition: An algorithm A is said to be directional if for some linear arrangement of the terminal nodes it never selects for examination a node situated to the left of a previously examined node. Theorem 2: The expected number of terminal positions examined by the TEST algorithm in the orooosition "V(S)> v" for the root gf-a testing (h, h, F)-tree, has a branching v * v* and P/(l-P) if v = v* F(v) = l-P and P* is the root factor d112 if , where v* sati of xd+x-1 = 0. sfies Theorem 3: TEST is asymptotically optimal over all directional algorithms which test whether the root node of a (h, d, F)-tree exceeds a speci- fied reference v. Corollar 1: Any procedure which evaluates a (h, debust examine at least 2dh/2-1 nodes. Corollary 2: The expected number of terminal positions examined by any directional algorithm which evaluates a (h, d)-game tree with continuous terminal values must have a branching factor great- er or equal to P/(l-P). The quantity P/(l-P) was shown by Baudet [3] to be a lower bound for the branching factor of the a-6 procedure. Corollary 2 extends the bound to all directional game-evaluating algorithms. Theorem 4: The expected number of terminal examinations performed by SCOUT in the evaluation of (h, d)-game trees with continuous terminal values has a branching factor of P/(l-P). Theorem 5: The expected number of terminal examinations performed by SCOUT in evaluating a (h, d, F)-game with discrete terminal values has a branchina factor dl/L, with exceptions only when one of the discrete values, v, satisfies F(v) = 1-p. 144 Corollary 3: For games with discrete terminal values satisfying the conditions of Theorem 5, the SCOUT procedure is asymptotically optimal over all evaluation algorithms. The improvement in efficiency due to the dis- crete nature of the terminal value manifests itself only when the search depth h is larger than log M/log [d(l-P)/P], where M is the quantization density in the neighborhood of VO = v. The branching factor of a-8 is less tractable than that of SCOUT. At the time this paper was first written the tightest bounds on r,,B were those delineated by Baudet [3] giving the lower bound r,-@ 2 P/(l-P) (a special case of Ccrollary 2) and an upper bound which is about 20 percent higher over the range 2 5 d I 32. Thus, SCOUT was the first algorithm known to achieve the bound P/(l-P) and we were questioning whether a-6 would enjoy a comparable asymptotic performance. More- over, it can be shown that neither SCOUT nor a-6 dominate one another on a node-by-node basis; i.e., nodes examined by SCOUT may be skipped by a-6 and vice versa [l]. The uncertainty regarding the branching factor of the a-0 procedure has recently been resolved [5] Evidently, U-B and SCOUT are asymptotically equiva- lent; r e uals P/(l-P) for continuous valued trees a%'dl92 for games with discrete values. P For low values of h the branching factor is no longer an adquate criterion of efficiency and the comparison between SCOUT and a-6 must be done empirically. The following table represents the number of node inspections spent by both algorithms on the game of Kalah (l-in-a-hole version) [2]: It appears that as the search depth increases SCOUT offers some advantage over cl-@. Experiments with higher numbers of stones in each hole indicate that this advantage may deteriorate for large d. We suppose, therefore, that SCOUT may be found useful in searching games with high h/d ratios. Cl1 PI [31 [41 [51 REFERENCES Pearl, J. "Asymptotic Properties of Minimax Trees and Game-Searching Procedures." UCLA- ENG-CSL-7981, University of California, Los Angeles, March 1980, to be published in Artificial Intelligence. Noe, T. "A Comparison of the Alpha-Beta and SCOUT Algorithms Using the Game of Kalah." UCLA-ENG-CSL-8017, University of California, Los Angeles, April 1980. Stockman, G. "A Minimax Algorithm Better Than Alpha-Beta?" Artificial Intelligence 12, 1979, 179-196. Baudet, G. M. "On the Branching Factor of the Alpha-Beta Pruning Algorithm." Artificial Intelligence 10, 1978, 173-199. Pearl, J. "The Solution for the Branching Factor of the Alpha-Beta Pruning Algorithm." UCLA-ENG-CSL-8019, University of California, Los Angeles, May 1980. II Random Orderina II Dvnamic Orderina Search Depth 2 SCOUT 82 a-B 70 % Improvement -17.0 .# J % SCOUT a-6 Improvement 39 37 -5.4 I 3 II 394 I : 380 -3.7 62 61 -1. 4 1173 1322 +11.3 91 96 -1. 5 2514 4198 +40.1 279 336 +17. 6 5111 6944 +26.4 371 440 +15. .45	1980	76
74	Problem Solving in Frame-Structured Systems Using Interactive Dialog Harry C. Reinstein IBM Palo Alto Scientific Center 1530 Page Mill Road Palo Alto, Ca. 94304 ABSTRACT This paper provides an overview of the process by which problem solving in a particular frame-like knowledge-based system is accomplished. The inter-relationship between specialization traversal and entity processing is addressed and the specific role of the user interaction is described. I INTRODUCTION Semantic networks Cl1 and frame-like systems have emerged as powerful tools in a variety of problem domains [Z&l. In many of these systems an initial knowledge base is used to drive an interactive dialog session, the goal of which is the instantiation of the particular knowledge base elements which represent a solution to the problem being addressed. In a system developed at the IBM Scientific Center in Palo Alto [3,41, a dialog is generated from a KRL-based c51 semantic network for the purpose of generating a well-formed definition of a medical sensor-based application program. It is intended that the user of the system be conversant with the problem to be solved by the application but not that they be a computer programmer. The overall logic of this process is the subject of this paper. II THE DIALOG LOGIC --- The ultimate goal of the problem-solving dialog session is the complete instantiation of all entities (knowledge units) relevant to the problem solution. To do this the system must be able to create new work contexts (in our case entities) from existing ones and be able to traverse the specialization hierarchies rooted at these entities to accomplish complete instantiation. The logic governing the interrelationships between these two tasks, and the methods of user interaction tend to characterize frame-based dialog systems. One could, for example, choose to pursue a path of 'least commitment' by processing all relevant references at their highest levels in the specialization hierarchies before attempting deeper specialization traversal [6,71. This approach seems well-suited to problem domains where the solutions are highly dependant on the interaction of constraints between the processed entities. In the case of our application development system it was felt that the solution process would be enhanced if individual entities were completely specialized as they were encountered, with outside references posted as pending work contexts for subsequent processing. This 'greatest commitment' approach seems well-suited to semantic networks in which specialization traversal in one hierarchy provides increasingly more specialized references to other hierarchies, and where the constraints between hierarchies are not potentially inconsistent. An example of this can be seen in the relationship between our SENSOR hierarchy, which provides descriptive knowledge about the analog, digital, and keyboard-entry sensors available in the laboratory and the DATA hierarchy, which describes the kinds of data which can be processed. In these interdependant hierarchies, one finds that BLOOD PRESSURE is measurable only by a subset of PRESSURE SENSORS, and that ARTERIAL BLOOD PRESSURE (a specialization of BLOOD PRESSURE) makes even further specializations on that set. Traversal of either hierarchy will implicitly specialize the related hierarchy. Downward traversal of specialization trees is the main driving force of the dialog system. III MECHANICS OF SPECIALIZATION TRAVERSAL - Entities intended for inclusion in the problem solution are dynamically replicated and inserted into the knowledge structure as an immediate specialization of the entity from which they are copied. As more becomes known about them, these new entities are moved down the specialization hierarchy, always appearing as a specialization of the most constrained model available in the initial knowledge base. These dynamic entities have exactly the same representation as the initial entities and differ from them only in that their constraints can be overwritten in the process of instantiation. If, for example, one of the attributes of a DEVICE is its user supplied name, then the value obtained for that name during the dialog would be placed in the dynamic entity while the corresponding 146 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. entity/attribute in the initial knowledge base is only constrained to be a name. V SUMMARY The mechanism for migrating a dynamic entity down the associated specialization hierarchy may require user interaction. This interaction is accomplished using a video character display with full screen data entry facilities, a light pen, and program function keys. It has been our experience that non-computer-trained users are very sensitive to the level of human factors provided and it is well worth any effort one can make to facilitate interaction. First the user is prompted to supply values for attributes which have been declared in the knowledge base to be user-sourced. (This is equivilent to the 'lab data' assertion in MYCIN c81). Having obtained these, a pattern-matching search is performed to see if specialization is possible. If not, the next step is to attempt specialization by allowing the user to choose from a list of the names of the immediate descendants at the current level in the hierarchy. If the user is unable to select the specialization by name, he or she is interrogated for selected attribute values which, if known, would determine a specialization path. This process continues until a terminus in the hierarchy is reached. During the traversal process any references to other entities must be resolved, and these references generate additional work contexts for the system. It is particularly important that the resolution process be able to determine if the reference should resolve to an already existing dynamic entity or if it should resolve to an entity in the initial knowledge base. Some considerations relevant to this problem are discussed below. Processing entities to their most specialized form is a valid driving function in some knowledge bases, and generating user interaction specifically for this traversal can be a sufficient involvement for the user in the problem solving process. Representing the results of the problem solving session in the same form as, and in direct association with the initial knowledge base has many positive features. Included among these is the ability to use a single search/resolution mechanism to select from either set of entities when building the problem solution. In general, frame-structured knowledge bases, in conjunction with user interaction can provide a powerful problem solving facility. REFERENCES Cl1 Fikes, R. and Hendricks, G. "A Network Based Representation and Its Natural Deduction System" In Proc. IJCAI-77. Cambridge, Massachusetts, August, 1977, pp. 235-245. c21 Waterman, D.A. and Hayes-Roth, F., (Eds.) Pattern-Directed Inference Systems. New York, New York: Academic Press, 1978. [31 Hollander, C.R. and Reinstein, H.C. "A Knowledge-based Application Definition System" In Proc. IJCAI-79. Tokyo, Japan, August, 1979, pp. 397-399. [41 Reinstein, H.C. and Hollander, C.R. "A Knowledge-based Approach to Application Development IV PROCESSING ENTITY REFERENCES for Non-programmers", Report G320-3390, IBM Scientific Center, Palo Alto, California, July 1979. When a reference resolves to a single entity one of three situations prevails: 1) the reference [51 Bobrow, D. and Winograd, T., "An Overview of is to an exactly matching dynamic entity, 2) the m, a Knowledge Representation Language." reference is to an existing dynamic entity which is Cognitive Science 1:l (1977), 3-46. less constrained than desired, or 3) the reference is to an entity in the initial knowledge base. In L61 Martin, N. et al "Knowledge Management for the first of these cases no further processing is Experiment Planning in Molecular Genetics" In Proc. required. In the second case, the more constrained IJCAI-77. Cambridge, Massachusetts, August, 1977, form of the attribute is forced into the dynamic pp. 882-887. entity and a search is performed to see if this new form permits further migration down the [71 Stefik, M.J., Planning With Constraints, Ph.D. specialization hierarchy. In this way values of Thesis, Stanford University, 1980, (available from attributes obtained from specialization down one Computer Science Dept., Stanford University, Report hierarchy can implicitly cause specialization STAN-CS-80-784). acitivity in related hierarchies. In the third case a new dynamic entity would be created. C8l Shortliffe, E.H. MYCIN: Computer-based Medical Consultations, New York, New York: American Elsevier, 1976. 147	1980	77
75	REPRESENTING KNOWLEDGE IN AN INTERACTIVE PLANNER Ann E. Robinson and David E. Wilkins Artificial Intelligence Center SRI International Menlo Park, California 94025 ABSTRACT This note discusses the representation for actions and plans being developed as part of the current planning research at SRI. Described is a method for uniformly representing actions that can take place both in the domain and during planning. The representation accommodates descriptions of abstract (hypothetical) objects. I. INTRODUCTION A principal goal of current planning and plan- execution research at SRI is development. of a planning system that interacts with a person, allowing that person to: (1) explore alternative plans for performing some activity, (2) monitor the execution of a plan that has been produced, and (72 modify the plan as needed during execution. Described here is the knowledge representation being developed. Our research builds' directly on previous planning research and on research in representing the domain knowledge necessary for participating in natural-language dialogs about tasks. In particular, some of our representation ideas are based on the process model formalism described in [2] and [S]. The basic approach to planning is to work within the hierarchical planning paradigm, representing plans in procedural networks, as has been done in NOAH [4] and other systems. Unlike its predecessors, our new system is being designed to allow interaction with users throughout the planning and plan-execution processes. The user will be able to watch and, when desired, guide and/or control the planning process. During execution of a plan, some person or computer system monitoring the execution will be able to specify what actions have been performed and what changes have occurred in the world being modeled. On the basis of this, the plan can be interactively updated to accommodate unanticipated occurrences. Planning and plan-execution can be intermingled by producing a plan for part of an activity and then executing some or all of that plan before working out remaining details. We are extending planning research in several major directions. One of the key directions, the one discussed here, is a method for representing -------- + The research reported here is supported by Air Force Office of Scientific Research Contract F49620-79-C-0188 and by Office of Naval Research Contract N00014-80-C-0300. actions that can take place both in the domain and during planning. Action descriptions (often referred to as operators), procedural networks, and knowledge about domain objects and their interrelationships are represented in the same formalism -- a hierarchy of nodes with attributes. This uniform representation provides the ability to encode partial descriptions of unspecified objects as well as objects in the domain model. Thus, operator descriptions referring to abstract (unbound) objects can be represented in the same formalism as procedural network nodes referring to specific objects in the domain model. (Partial descriptions of unspecified objects will be described here as constraints on the possible values of a variable representing the object.) Operators can be encoded at several levels of abstraction. Each one contains information for planning at the next level of detail. We have already encoded many domain operators for a construction task; planning operators will be encoded shortly. The domain operators provide the planning system with information about producing a plan in the domain. The planning operators provide the planning system with information so it can reason about its own planning process (meta- planning). They also provide a major part of the interface between the planning system and the user, who will be able to direct the planning process via the planning operators. The uniformity of representation for domain knowledge, specific plans of action, and all operators will facilitate both the user's ability to interact with and control the planning system, and the system's ability to incorporate (learn) new operators from plans it has already produced. We will describe the representation in more detail below. 148 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. II. THE FORMALISM The formalism for representing knowledge about actions, plans, and domain objects consists of typed nodes linked in a hierarchy. Each node can have attributes associated with it. There are four node types for representing objects: CLASS, INSTANCE, INDEFINITE, and DESCRIPTION. These will not be discussed in more detail here since they are similar to those occurring in re resentation formalisms such as KRL, FRL, and UNITS -f 51. The node types for representing actions can be grouped into four categories: OPERATOR, for encoding operators; PNET, for representing specific actions (nodes in the procedural network); PLOT, for describing how to expand a given OPERATOR, i.e., a description of an action in greater detail; PNET.ACTION, for encoding plan steps (procedural network nodes) that have been 'executed' and thus represent actions assumed to have occurred in the world being modeled. Nodes can have lists of attributes and can be connected into a hierarchy through CLASS and SUBCLASS links. Attributes of nodes for representing actions include the resources and arguments of the action (i.e., the objects that participate in the action), the action's goal, the action's effects on the domain when it is performed, and the action's preconditions. OPERATOR nodes have a plot attribute which specifies PLOT nodes for carrying out the operator. The PLOT of an operator can be described not only in terms of GOALS to be achieved, but also in terms of PROCESSes to be invoked. (Previous systems would represent a PROCESS as a goal with only a single choice for an action to perform.) The ability to describe operators in terms of both GOALS and PROCESSes will help simplify encoding of operators and will allow the planning system to reason about alternative action sequences more efficiently. Figure 1 shows a sample operator and a PNET it might produce. The figure illustrates the uniformity across different types of nodes in our formalism. The nodes are expressed in the same formalism, and, for the most part, have the same attributes (e.g., resources, shared-resources, arguments, preconditions, purpose) with similar values for these attributes. "Similar values" means that the values refer to the same types of objects -- often the value of an attribute for some node will be more constrained than the value of the same attribute in a corresponding node of a different category. The next two paragraphs illustrate this in detail, after which we describe two instances where the uniformity of the representation is advantageous. Attributes in OPERATOR and PLOT nodes generally refer to variables rather than specific objects, since these are uninstantiated operators that may be instantiated into PNET nodes in different ways during planning. For example, in Figure 1 resource variables meat1 and vegl in the operator FIX.MEAL refer to objects of the meat and vegetable class, respectively. In the expansion of FIX.MEAL, meat1 has been constrained to be a fish (denoted by calling it "fishl") since it was so constrained in the node being OPERATOR expanded. FIX.MEAL RESOURCES. meat 1, vegl - SPLIT GOAL (PREPARED meat 1) RES. meat1 GOAL: (PREPARED vegl) RES: veg 1 f PROCESS: SERVE JOIN - RES meatl, vegl A PNET node which represents usmg FIX.MEAL ma plan. PROCESS. FIX.MEAL . . . 17 RESOURCES. . . 0 flshl veg 1 Expansion of thts PNET node at the next level usmg FIX.MEAL. Figure 1 In our formalism, such variables are described by INDEFINITE nodes with constraints on their possible values. For PNET nodes, attributes frequently refer both to variables (which will often be more constrained in this case) and to completely specified objects. For PNET.ACTION nodes, attributes generally refer to specific objects in the domain model. The system's ability to use INDEFINITE nodes to partially describe objects is important for representing objects with varying degrees of abstractness in the same formalism. Few previous planning systems have used this approach (e.g., NOAH cannot partially describe objects and has different formalisms for describing operators and procedural nets). Stefik's [51 system does allow abstract descriptions and constraints on partially described arguments, but arguments are required to be fully instantiated before the constraints can be evaluated. (See also Hayes-Roth et al. cw The uniformity of representation between PLOT and PNET nodes permits the description of operators as what amounts to generalized fragments of procedural network. This turns problem solving into a process of incremental instantiation. During planning, PNET nodes are incrementally expanded to a greater level of detail by selecting 149 an appropriate operator, determining which of its variables match those in the node being expanded, creating new variable records for those variables not matched, adding any new constraints to these variables, and following the operator's plot description to create new procedural network nodes. The uniformity of representation facilitates this production of PNET nodes from PLOT nodes. Once a plan has been successfully constructed, it may be desirable to save it for subsequent planning activities, incorporating it into the system as a new operator. We expect to develop algorithms for doing this, i.e., producing an operator (with its associated PLOT nodes) from PNET fragments. For each control node and each action- oriented node in a procedural network, a corresponding PLOT node can be easily created for the operator because of the uniformity of representation. The major task remaining in producing an operator would be generalizing the constraints on values for variables in the procedural network nodes into looser constraints in the new operator. An additional uniformity between descriptions of specific actions and operators facilitates the matching of an operator to the node it is to expand. Thus PROCESS and GOAL nodes in a procedural network or plot will have attributes similar to those of the OPERATOR node which represents a more detailed description of their corresponding action. The similarities of representation of all action-oriented nodes facilitates interaction with the user who can talk in the same way about operators, steps in operator plots, and nodes in the procedural network. Similarly, description of actions is facilitated by this uniformity. Organizing the representation as nodes with attributes is, of course, not new and is not essential. The representation could also be expressed in a formal logic (a translation to logic would be fairly straightforward). We have chosen to represent things as nodes with attributes because this blends well with our plans for interaction with the user. III. PARTIAL DESCRIPTION USING CONSTRAINTS Stefik's system [5], one of the few existing planning systems with the ability to construct partial descriptions of an object without identifying the object, contains a constraint- posting mechanism that allows partial descriptions similar to those described above. Our system also provides for partial description using constraints, and extends Stefik's approach in two ways. Unlike Stefik's system, our system permits evaluation of constraints on partially described objects. Both CLASSes and INSTANCES can have constraints. For example, a set can be created which can be constrained to be only bolts, then to be longer than one inch and shorter than two inches, and then to have hex heads. Our system also provides with the for partial descriptions that vary context, thus permitting consideration of alternative plans simultaneously. A context mechanism has been developed to allow for alternative constraints on variables relative to different plan steps. The constraints on a variable's value as well as the binding of a variable to a particular instance (possibly determined during the solution of a general constraint-satisfaction problem) can only be retrieved relative to a particular context. This permits the user to easily shift focus back and forth between alternatives. Hayes-Roth et al. [I] describe the use of a blackboard model for allowing shifting of focus between alternatives. Such focus shifting can not be done in systems using a backtracking algorithm where descriptions built up during expansion of one alternative are removed during the backtracking process before another alternative is investigated. Most other planning systems either do not allow alternatives (e.g., NOAH[4]), or use a backtracking algorithm (e.g., Stefik [5], Tate [6]). IV. CONCLUSION We have described some properties of the knowledge representation developed for our new planning system. Most of the planner is still under development (e.g., critics, reasoning about resources, and search control have yet to be implemented). The central idea discussed here is the uniform representation of the domain operators, planning operators, procedural networks, and knowledge about domain objects. Ways to exploit this uniformity are pointed to. These include a rich interaction with the user, meta-planning, and having the system learn new operators from plans it has constructed. REFERENCES 1. Hayes-Roth, B., F. Hayes-Roth, S. Rosenschein, S. Cammarata, "Modeling Planning as an Incremental, Opportunistic Processll, In Proc. IJCAI-79. Tokyo, Japan, August, 19791 pp. 375- 383. 3. Robinson, A.E., D. Appelt, B. Grosz, G. Hendrix, and J. Robinson, "Interpreting Natural- Language Utterances in Dialogs About Tasks", Technical Note 210, SRI International, Menlo Park, California. March, 1980. 4. Sacerdoti, E., A Structure for Plans and Behavior. Elsevier North-Holland, NewYork. 1977. 5. Stefik, M., Planning With Constraints. Report STAN-CS-80-784, Computer Science Department, Stanford University, Ph.D Dissertation. 1980. 6. Tate, A., "Generating Project Networks", In Proc. IJCAI-77. Cambridge, Mass. August, 1977. 150	1980	78
76	INFEHENCEWITHRECURSIVERULES Stuart C. Shapiro and Donald P. McKay Department of Computar Science State University of New York at Buffalo Amherst, New York 14226 ABSTRACT Recursive rules, such as "Your parents' ances- tors are your ancestors", although very useful for theorem proving, natural language understanding, questions-answe ring and information retrieval systems, present problems for many such systems, either causing infinite loops or requiring that arbitrarily many copies of them be made. We have written an inference system that can use recursive rules without either of these problems. The solu- tion appeared automatically from a technique designed to avoid redundant work. A recursive rule causes a cycle to be built in an AND/OR graph of active processes. Each pass of data through the cycle resulting in another answer. Cycling stops as soon as either the desired answer is produced, no more answers can be produced, or resource bounds are exceeded. Introduction Recursive rules, such as "your parents' ances- tors are your ancestors", occur naturally in inference systems used for theorem proving, question answering, natural language understanding, and information retrieval. Transitive relations, v(x,y,z) [ANCES'lloR(x,y) & ANCESToR(y,z)-+ %%SToR(x,z)], inheritance rules, e.g. ~z(x,y,p) [ISA(x,y) & HAs(y,p) -+ HAS(x,p)l, circular defini- tions and equivalences are all occurrences of recursive rules. Yet, recursive rules present problems for system implemantors. Inference systems which use a %aive chaining" algorithm can go into an infinite loop, like a left-to-right top-down parser given a left recursive grammar [41. Sme systems will fail to use a recur- sive rule more than once, i.e. are incomplete [6,121. Other systems build tree-like data structures (connection graphs) containing branches thelengthofwhichdependonthenlrmberoftimes the recursive rule is to be applied [2,131. Since scme of these build the structure before using it, the correct length of these branches is problematic. Son-e systems eliminate recursive rules by deriving and adding to the data base all implications of the recursive rules in a special pass before normal inference is done [91. The inference system of SNePS [13] was designed ti use rules stored in a fully indexed data base. when a question is asked, the system retrieves --7-s- Theworkwas supported inpartbytheNationa1 Science Foundation under Grant No. MCS78-02274. relevant rules and builds a data structure of processes which attempt to derive the answer frcm the rules and other information stored in the data base. Since we are using a semantic network to represent all declarative information available in the system, we do not make a distinction between "extensional" and "intensional" data bases, i.e. non-rules and rules are stored in the same data base. More significantly, we do not distinguish "base" frm "defined" relations. Specific instances of ANCESTOR maybe storedaswell as a ruledefining ANCESTOR. This point of view contrasts with the basic assurrption of several data base question answering systems [3,8,9]. In addition, the inference system described here does not restrict the left hand side of rules to con- tain only one literal which is a derived relation [3], does not need to recognize cycles in a graph [3,8] and does not require that there be at least one exit frm a cycle [8l. The structure of processes may be viewed as an AND/OR problem reduction graph in which the process working on the original question is the mot, and rules are problem reduction operators. Partly influenced by Kaplan's producer-consumer model [53, we designed the system so that if a process working on some problem is about to create a process for a subproblem, and there is another process already working on that subproblem, the parentprocess canmake useof the extant process and so avoid solving the same problem again. The method we employ handles recursive rules with no additional mechanism. The structure of processes may be viewed as an active connection graph, but, as will be seen below, the size of the resulting structure need not depend on the number of times a recursive rule will be used. This paper describes hm our system handles recursive rules. Aspects of the system not directly relevant to this issue will be abbreviated orcmitted. Inparticular,details of thematch routine which retrieves formulas unifiable with a given formula will not be discussed (but see [lOI). The Inference Systeq The SNePS inference system builds a graph of processes [7,11] to answer a question (derive instances of a given formula) based on a data base of assertions (ground atomic formulas) and rules (non-at&c formulas). Each process has a set of 151 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. registers which contain data, and each process may send messages to other processes. Since, in this system, the messages are all answers to scnte ques- tion, we will call a process P2 a boss of a process PI if PI sends messages to P2. Sane processes, called data collectors, are distinguish& by two features: 1) they can have store than one boss: 2) they store all massages they have sent to their bosses. The stored messages are used for two purposes: a) it allows the data collector to avoid sending the sama message twice; b) it allows the data collector to be given a new boss, which can inmediately be brought up to date by being given all the messages already sent to the other bosses. Four types of processes are important to the discussion of recursive rules. They are called INFER, CHAIN, SWITCH and FILTER. INFER and CHAIN are data collectors, SWITCH and FILTER are not. Four Processes An INFER process is created to derive instances of a formula, Q. It first matches Q against the data base to find all formulas unifiable with Q. The result of this match is a list of triples, <T,-r,W, where T is a retrieved formula called the target. and 'c and 0 are sub- stitutions called ti?get binding and source binding respectively. -- Essentially -c and o are factorea versions of the mst general unifier (mgu) ofQ and T. Pairs of the rqu whose variables are in Q appear in G, while those whose variables areinTappearinT. Any variable in term position is taken from T. Factoring the q-u obviates the need for renaming variables. For example if Q=P(x,a,y) and *P(b,y,x), we would have o=(b/x,x/y) and T={a/y,x/x) (the pair x/x is included to make our algorithms easier to describe). Note that Qa = T-r = P(b,a,x), the variables in the variable position of the substitution pairs of 0 are all and only the variables in Q, the variables in the variable position of 'c are all and only the variables in T, all terms of cr ce from T, and the non-variables in T came from Q. for For each match Q, there are two <TJ,o> that an INFER finds possiblities we shall consider. First, T might be an -assertion in the data base. In this case, G is an answer (Qcr has been derived). If the INFER has already stored 0, it is ignored. Otherwise, CT is stored by the INFER and the pair <Q,o> is sent to all the INFER's bosses. For CT to be a reasonable answer, it is crucial that all its vai&les occur in Q. The other case we shall con- sider is the one in which T is the consequent of some rule of the form Al&...&An1T. (our system allcrws other forms of rules, but consideration of this one will suffice for explaining how we handle recursive rules). In this case, the INFER creates two other processes, a SWITCH and a CHAIN to derive instances of T-r. The SWITCH is made the CHAIN's boss, and the INFER the SWITCH's boss. It maybe the case that an already extant CHADJ maybeused insteadof anewone. This will be discussed below. The SWITCH process has a register which is set to the source binding, O. The answers it receives from the W are substitutions B, signifying that TTB has been derived. SW'PK!H sends to its boss the application 46. the sub- , stitution derived from o by replacing each term t in o by tB. The effect of the SWITCH is to change the answer from the context of the vari- ables of T to the context of the variables of Q. 1n0u.r example, themmightsendthe answer B=k/xl. SWITCH would then send a\B = {b/x,x/y}\ {c/xl= (b/x,c/y) to the INFER, indicating that Q~\B = P(x,a,y)(b/x,c/y) = P(b,a,c) has been derived. The importance of the factoring of the mguof Q andT into the sourcebinding 0 and the targetbinding T - a separation which the SWITCH repairs -- is thattheCHAIN canwork onT in the context of its original variables and report to many bosses, each through its own SWITCH. A CHAIN process is created to use a particu- lar substitution instance, T, of a particular formula, Al&... GAk1T to deduce instances of T-r. Its answers, which will be sent to a SWITCH, will be substitutions B such that T-rB has been deduced using the rule. For each Ai, ISilk, the CHAIN tries to discover if ANT is deducible by creating an INF'ER process for it. However, an INFER process might already be working on Aia. If ~.=TT, the already extant INFER is just what the CHAIN wants. It takes all the data the INFER has already collected, and adds itself to the INFER's bosses so that itwill also get future answers. If a is more general than -r, the INFER will produce all the data the CHAIN wants, but unwanted data as well. In this case the CHAIN creates a FILTER process to standbetween it and the INFER. The FILTER stores a substitution consisting of those pairs of T for which the term is a constant, and when it receives an answer substitution from the INFER, it passes it along to its CHAIN only if the stored substitution is a subset of the answer. For example, if T were (a/x,y/z,b/w} and a were (u/x ,v/z ,v/wl , a FILTER would be created with a sutstitution of {a/x,b/w), insuring that unwanted answers such as (c/x,d/z,b/w) produced by the more general INFER were filtered out. If a is not compatible with T, or is less general than T, a new INFER must be created. However, if a is less general than T, the old INFER might already have collected answers that the new one can use. These are takenby the new DJFERand senttoits bosses. Also, since thenew INFERwillprcduce all the additional answers that the old one would (plus others), the old INFER is eliminated and its bosses given to the new INFER with intervening FILTEXs. The net result is that the same structure of processes is created regardless of whether the mre general or less general question was asked first. A CHAIN receives answers from INFERS (possibly filtered) in the form of pairs <Ai,Bi> indicating that AiPi, an instanceof the antec&entAi,has beendeduced. Whenever the CHAIN collects a set of consistent substitutions (Bi,...,Bn), one for each antecedent, it sends an answar to its bosses consisting of the ccgnbination of Bl,...,Bk (where the ambination of 61 = ~tll/vll,...,tln,/vlnlI ,...,Bk = ctkl/vkl ,...,tknk/vknkl is the mgu of the expressions (vll,...,vlnl,...,vkl,...,vknk) and (tll ,. . . ,tlnl,. . . ,tkl ,...,tknk) [l, p.1871). 152 Recursive Rules Cause Cycles Just as a CHAIN can make use of already exist- ing INFERS, an INFER can r&e use of already existing ms, filtered if necessary. A recursive rule is a chain of the form Al&... &Ak~Bl,Bl&...&Bn~...X, with C unifiable with at least one of the antecedents, Al say. When an INFER operates on Al, it will find that C matches Al, and it may find that it can use the CHAIN already created for C. Since this CHAIN is in fact the INFER's boss, this will result in a cycle of processes. The cycle will produce more answers as new data is passed around the cycle, but no infinite loop will be created since no data collector sends any answernore thanonce. (If an infinite set of Skolem constants is generated, the process will still terminate if the root goal had a finite number of desired answers specified [ll, p.1941). Figure 1 shows a structure of processes which we consider an active connection graph. It is built to derive instances of ANCESToR(William,w) form the rules vCrx,y) [PARENT(x,y)~ANCESToR(x,y) 1 and V‘(x,y,z) [ANcESToR(x,y) & PAREMT(y,z)x ANCESToR(x,z)]. The notation for the rule instances is similar to thatpresented in [31. Note particularly the SWITCH in the cycle which allows newly derived instances of the goal ANCESToR(William,w) to be treated as additional instances of the antecedent ANCESToR(William,y). A similar structure would be built regardless of the order of asserting the two rules, the order of anteaedents inthetwoantecedentrule, the order of execution of the processes, whether the query had zero, either one, or both variables ground, or if the twoanteceiientruleused ANCESTOR for both antecedents. In the SNePS inference system, recursive rules cause cycles to be built in a graph structure of processes. The key features of the inference system which allow recursive rules to be handled are: I) the processes that produce derivations (INFER and CHAIN) are data collectors: 2) data collectors never send the same answer nore than once: 3) a data collector may report to mre than one boss: 4) a new boss may be assigned to a data collector at any time -- it will miately be given all previously collected data; 5) variable contexts are localized, SWITCH changing contexts dynamically as data flows around the graph: 6) FILTERs allow more general producers to be used by less general consumers. 1. chang, C.-L., and Lee, R.C.-T., Symbolic Logic and Mechanical Theorem Proving, Acadtic Press, New York, 1973. 2. chang, C.-L., and Slagle, J.R., Using rewriting rules for connection graphs to prove theorems, Artificial Intelligence 12, 2 (August 1979), 159-180. 3. chang, C.-L., On evaluation of queries contain- ing derived relations in a relational data base. In normal Bases for Data Bases, Gallaire, H., Minker, J. and Nicolas, J. (edS.), Plenum, New York, i980. 4. Fikes, R.E., and Hendrix G.G., The deduction component. In Understanding Spoken Language, Walker, D.E., ed., Elsevier North-Holland, 1978, 355-374. 5. Kaplan, R.M., A multi-processing approach to natural language understanding. Proc. National Computer Conference, AFIPS Press, Montvale, NJ, 1973, 435-440. 6. Klahr, P., Planning techniques for rule selec- tion in deductive question-answering. In Pattem- Directed Inference Systems, Waterman, D.A., and Hayes-Roth, R., eds., Academic Press, New York, 1978, 223-239. 7. McKay, D.P. and Shapiro, S.C., MULTI--A LISP based muitiprocessing system. Proc. 1980 LISP Conference. Stanford Universitv. 1980. 8. Naqvi,'S.A., and Henschen,-L.J., Performing inferences over recursive data bases. Proc. First AAAI Conference, Stanford Univers- 1980. 9. Reiter, R., On structuring a first order data base, Proc. SecondNational Conference, Canadian Society for Computational Studies of Intelligence, 1978, 50-99. - 10. Shapiro, S.C., Representing andlocatingde- duction rules in a semantic network. Proc. Work- shop on Pattern-Directed Inference Systexns. In SIGART Newsletter, 63 (June 1977), 14-18. 11. Shapiro, S.C., The SNePS sepnanticneixork processing system. In Associative Networks: The Representation and Use of Kncwledse bv Camrsuters. F&dler, N.V., ed., Acadenic Press, NW York, 1979, 179-203. 12. Shortliffe, E-H., Computer Based Medical Con- sultations: MYCIN, &rerican Elsevier, New York, 1976. 13. Sickel, S., A search technique for clause interconnectivity graphs, IEEE Transactions on Computers Vol. C-25, 8 (August 1976) I 823-835. ANCZSMR(William,z) PARETJT(William,y) AN(TlZSMR(William, y) Figure 1. 153	1980	79
77	ON PF0VING LAWS OF TBE ALGEBRA OF FP-SYSTE2% INEDINB~LCF Jacek Leszczykcwski Polish Acadeq of Sciences Institute of Computer Science P.O.BOX 22, 00-901 Warszawa PKiN, PQLMD I INTFODUCX'ION J.Backus, in CACM 21/8, defined a class of ap- plicative prograrrunin g systems called FP /functional prograrmcing/ systems in which a user has: l.objects built recursively frcxn atcms, UU /an undefined ele- ment/ and objects by a strict /i.e. a UU-preserving/ "list" operator, 2. elementary functions over obje- cts, 3. tools for building functions out of already defined functions. One can think of a machine support while work- ing with FP systems and proving facts about FP sys- tems as well as facts concerning the functions be- ing defined. The choice of EDINBURGH KF is rather natural because it is an interactive ccmputer sys- tem /implemented in LISP/ for reasoning about fun- ctions /see &I/. It consists of two parts. The first part is a family of calculi each of which is characterized by four factors: 1. type operators /representing domains in the sense of Scott's theo- ry; see m/, 2. constants /representing continuo- us functions/, 3. axicms, 4. inference rules. One of them, PPLAMBDA, is given as the "initial" calcu- lus, and other calculi may be built by users as e.x- tensions of existing calculi. The second part is a high level prograrmxin g language ML which is fully higher order and is strongly typed. Its polymorphic types make it as convenient as typeless languages. This paper is a short report on the application of EDINBUZH LCF to proving the laws of the algebra of FP systems listed by Backus in [ll. Actualiy, we generalized FP-systems and the laws are formula- ted in stronger fonn than it was done by Backus. We briefly describe /sec.II/ the style of proving with the system, then /sec.III/ ccxtnnent the strtegies u- sed in the proofs giving only their specifications. The summing up remarks will be given in sec.IV. r@- re detailed report on the project is given in D>. II STYLEOFPROVING As mentioned there are inference rules associa- ted with each of the calculi of the system; the in- ferece rules of PPLAMBDA are primitive, and derived rules may be prcgramned by users. The inference ru- les are represented as ML-functions taking theorems /being a data type in ML/ as arguments and giving values which are theorems as well; an example is the computational induction rule INDUCT /it is the Scott induction rule; for more detailes see c21/. We could prove our theorems applying the inference rules in appropriate order but it is not a convenient style of proving. We base our proofs on partial subgoaling meth- ods, called tactics; these mean that given a formula to be proved-t to transform it into "simpler" formulae /which in turn have to be proved/ and the proof justifying the "transformation". The system can support this kind of proving via predefined ty- pes: goal, tactic and proof, defined as follatiJs: goal = form # siqset f form list proof = tlxn list -7 thm tactic = goal -> goal list# proof The first element of the Cartesian product defining the type goal is for stating the formula which is going to be proved, the third one for listing assum- ptions, the second one is /frcm the user's point of view/ an abstract type consisting of simplification rules; these are /possibly conditional/ equivalen- ces of terms to be used as left-to-right rewriting rules. We shall explain new the use of the subgoaling methods. Let us define when a theorem A> f' /A'- hypotheses, f'- conclusion/ achieves a goal f,ss,A. This is the case when, up to renaming of bound va- riables, f' is identical with f and each member of A' is either in A or is in the hypotheses of a member of the simplification set ss . Then, we say, a theorem list achieves a goal list if the first element of the theorem list achieves the first element of the goal list, etc. Thus, a tactic T will work "properly" if for any goal g and any goal list gl and any proof p such that T(g) = gl,p if we have a theorem list thml which achieves the goal list gl , then p(thml) will achieve the goal g - An important special case is when gl is empty for then we need only apply p to the empty theorem list - i.e. evaluate p(ni1) to obtain a theorem achieving our original goal g . We shall use two of the standard tactics. The first is SIMPTAC; applied to any goal (w,ss,A) , SIMPTAC produces a singleton list of goals E(w', ss ,Afl andproof p , where w' is the simplification of w by rewriting rules in ss and p justifies all the simplifications made. In the case where w' is a tautology the goal list is null. The second one is CONDCASESTAC ; for any goal (w,ss,wl) it 84 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. finds a term t of the type tr /truth values do- main/ which is free in w and occurs as the con- ditional expression and then produces three subgo- als with formula w as the their first element and the simplification sets extended by the new assump- tions /caseswhen t equal UU, true and false respectively/. NW we introduce the last tool for desigining proofs: tacticals - mechaninsms for ccxnbining tactics to form larger ones. We shall use only one of the standard tacticals of the system called THEN For any tactic Tl and T2 the canposed tactic Tl THEN T2 applies Tl to the'goal and then applies T2 to all resulting subgoals produced by Tl ; the proof function returned is the ccmposition of the proof functions produced by Tl and T2 . III CQMMENTSONTHEPROOFS There are tree Foups of the proofs of the laws of the FP-system algebra listed by Backus in 09. The first one is based on SIMPTAC ; for exam- ple, to prove II.1 of fl) we used the tactic: AFTAC THEN SIMPTAC where AFTAC is one of the pro gramned tactic and is specified as follms: AFTAC ("u =<v",ss,wi) = I"!X. u x = v x",ss,wl-J , p where: ! - universal quantification = < stands for equality or inequality of terms X- anewvariable The second group of proofs is based on COND- CASESTAC . The ccxrrposed tactic used in the proofs was: APTAC THEN CONDCASESTAC THEN SIMFTAC . For example in proofs of the laws: II.2 and 11.3.1 presented in PI . The third group of proofs involves the use of the ccmputational induction rule INDUCT in an in- derectway. The proofs are based on the programmed tactic INDTAC which is an :'inverse" of the struc- tural induction rule over the type of lists which in turn is derived frcm the ccxnputational induction rule INDUCT. The specification of INDATC is the following: where: X- anew variable ss/ is ss extended by the asswtion w simplified by ss P uses the structural induction ru- le on lists derived form INDUCT nil and cons are the constants over lists. Let us presentoneof the proofs using INIYTAC Suppose we want to prove: "ToALL(G o F)= (ToALL G)o @CALL F) 11 where : 0 - the ccmposition of functions ToAILis takes two arguments:H-being a function and a list and produces a list of the results and the fol- lowing axicms are satisfied: "!H. TbALL H UU = UU II "!H. ToALL H nil = nil It "!H.!X.!L. ToALL H (cons X L)= cons(HX)(ToALLHL)i' . Frcxn the shape of subgoals produced by INDTAC we knm that we need to simplify the formulae by the above axicxns as well as the definition of the ccxnposition operator. Suppose we created the simp- lification set ss including the desired rewriting rules; new we can specify our goal in the following way: g = "ToALL (+G o F)= (Jl?oALL G)o (TbALL F)",ss,nil L&T = APTACTHENINDTACTHENSIMPTAC bea tactic we want to use to tackle our problem. If we apply our tactic T to our goal g we get a pair as result. Let us store the value on the variables P and gl ; we can do it in ML in the following way: gl,p := Tg It terns out that our tactic was fully successful and the system respond that gl is the empty list. Thus I see sec.11, we can apply p to the empty theorem list and the produced value is the theorem we wanted to prove. In general cases when we are not so clever, the subgoal list is not empty and we have to apply another tacit to solve the subgoals. Iv . FINAL- As we mentioned in sec.7: the aim of the paper was to present an application of EDINBURGH ID But the aim of any application itself is to find general tactics which can be used in totally diffe- rent examples. Why? Because EDINBURGH LCF is es- sentially a tool placed samewhere between a theorem prover and a proof checker; this is why we cannot rely on the built in strategies which is the case with the theorem provers but we are interested in looking for general purpose ones which when found canbeprogramnedin ML andanbeused to tackle other goals. The tactic used to prove the laws of the FP-system algebra are of a ganeral use and we- re inwolved in proving properties of the functions defined in FP-systems; see [9J. The generalization 85 of FP systems is briefly described in 12 arad presented in mr>re detailed version in v 9 . Exam- ples of more powerful and ccmplex tactics can be found in t8) . Let us canpare EDINBUXHLCF with other im- plemented systems. On one hand the explicite pre - sence of the logic of the systemmakes DXNBURGH LCF "hL11Tkzn-oriented" and easy to extend. On the other hand, for example, the Bayer-Moore theorem prover /see L3]/ is very efficient in its use of built in strategies, but difficult to extend ; by contrast the need to conduct all inferences through the basic inference rules /as ML-procedu- res/ , which appears necessary if we wish to allaw users to extend the system reliably by prqxxrming leads to sm inefficiency in LCF. This we have found tolerable and indeed it can be reduced sig - nificantly by direct implementation of ML /which at present is ccmpiled into LISP / . Another way of making the system quick is by running it on rrcul- tiprocessor machines which is done for example at Royal Technical University , Stockholm, Sweden . For a nice general ccmparison of these two system see [5J . The EDINBURGH ICF style of proving which con sists in solving the problems by means of program- med proof strategies seems to be natural. It took the athor 2 months to be able to work with the system. This style fits pretty well to doing large proofs with machine assistance. By this we mean neither that a large proof is submitted step by step and merely checked by the machine /see D] /, nor that the system discovers the large proof by itself, but that the problem may be split into sma- ler parts, each of which is tackled semiautcmatica- ly by a subgoaling method. A nice example of such application of EDINBURGH XF is the ccmpiler co- rrectness proof presented in E41 . I wish to thank Awa Cohn, Mike Cordon, Robin Milner and Chris Wadsworth for their friendly help duringmy stay inFdinburgh andespecially Robin Milner for his support while preparing the draft version of the paper. REFERENCES BackusJ. lrCan programming be liberated frcxn the von Neumann style? A functional prcgram- ming style and its algebra of programs", C&-m ACM 21,8 , 1978 . Bird R., "Programs and Machines; an introduc- tion to the theory of camputation", Wiley 1976 Boyer R.S., mre J S., "A ccmputational Logic" Academic Press, New York 1979 . Texas, 1979 . Cohn A., "Remarks on Machine Proof", manusc:- ript, 1980 . Gordon M., Milner R., Wadsworth C., 'EDINBURGH Ia?" , Springer Verlag , 1979 . van Benthem Jutting L.S., "Checking Landau's 'Grundlagen'in the AUTCMATH system', Tech. Hoghschule, Eidhoven, The Netherlands, 1977 . IeszczyXmski J., ~~AI-I experiment with EDINBURGH LCF" , Proc. CAUE-5. , Les Arsc, France , 1980 . Leszczylmski J., "Theory of FP systems in EDINBUIGH LCF", Internal Report, Ccanp. Sci. Dept., Edinburgh University, Edinburgh, Scot- land, 1980 . Milner R., "LCF: a way of doing proofs with a machine", Proc. MFCS-8 Symposium, OlcEnouc, Chechoslcrwakia, 1979 . Milner R., "Implementation and application of Scott's logic for ccxnputable functions", Proc. ACM Conference on Proving Assertions abouT Programs, SIGPLAN Notices , 1972 . Leszczy3xwski J., "EDINBUZH IXF supporting FTI? systems", Proc. Annual Conference of the Geselschaft fur Informatik, Universitat des Saarlandes, 1980 . [41 Cohn A., "High level proof in LCF", Proc. 4th Workshop on Autcmated Deduction , Austin , 86	1980	8
78	Mapping Image Properties into Shape Constraints: Skewed Symmetry, Affine-Transformable Patterns, and the Shape-from-Texture Paradigm John R. Kender and Takeo Kanade Computer Science Department Carnegie-Mellon University Pittsburgh, Pa. 15213 1. Introduction Certain image properties, such as parallelisms, symmetries, and repeated patterns, provide cues for perceiving the 3-D shape from a 2-D picture. This paper demonstrates how we can map those image properties into 3-D shape constraints by associating appropriate assumptions with them and by using appropriate computational and representational tools. We begin with the exploration of how one specific image property, “skewed symmetry”, can be defined and formulated to serve as a cue to the determination of surface orientations. Then we will discuss the issue from two new, broader viewpoints. One is the class of Affine-transformable patterns. It has various interesting properties, and includes skewed symmetry as a special case. The other is the computational paradigm of shape-from-texture. Skewed symmetry is derived in a second, independent way, as an instance of the application of the paradigm. This paper further claims that the ideas and techniques presented here are applicable to many other properties, under the general framework of the shape-from-texture paradigm, with the underlying meta-heuristic of non-accidental image properties. 2. Skewed Symmetry In this section we assume the standard orthographic projections from scene to image, and a knowledge of the gradient space (see [4]). Symmetry in a 2-D picture has an axis for which the opposite sides are reflective; in other words, the symmetrical properties are found along the transverse lines perpendicular to the symmetry axis. The concept skewed symmefry is introduced by Kanade [l] by relaxing this condition a little. It means a class of 2-D shapes in which the symmetry is found along lines not necessarily perpendicular to the axis, but at a fixed angle to it. Formally, such shapes can be defined as 2-D Affine transforms of real symmetries. Figures 2-1 (a)(b) show a few key examples. Stevens [5] presents a good body of psychological experiments which suggests that human observers can perceive surface orientations from figures with this property. This is probably because such qualitative symmetry in the image is often due to Figure 2-1: Skewed symmetry (b) real symmetry in the scene. Thus let us associate the following assumption with this image property: “A skewed symmetry depicts a real symmetry viewed from some unknown view angle.” Note that the converse of this assumption is always true under orthographic projection. We can transform this assumption into constraints in the gradient space. As shown in Figure 2-1, a skewed symmetry defines two directions: let us call them the skewed-symmetry axis and the skewed-transverse axis, and denote their directional angles in the picture by (r and /?, respectively (Figure 2-l(c)). Let G = (p,q) be the gradient of the plane which includes the skewed symmetry. We will show that p’cos(~) - q’sin(y) = -cos(a-/?) (1) where p’ = pcosX + qsinX q’ = -psinX + qcosX h = (a + /3)/2. Thus, the (p,q)‘s are on a hyperbola. That is, the skewed symmetry defined by (Y and fl in the picture can be a projection of a real symmetry if and only if the gradient is on this hyperbola. The skewed symmetry thus imposes a one-dimensional family of constraints on the underlying surface orientation (p,q). 3. Affine-Transformable Patterns In texture analysis we often consider small patterns (texel= texture element) whose repetition constitutes “texture”. Suppose we have a pair of texel patterns in which one is a 2-D Affine transform of the other; we call them a pair of Affine-transformable patterns. Let us assume that “A pair of Affine-transformable patterns in the picture are projection of similar patterns in the 3-D space (i.e., they can be overlapped by scale change, rotation, and translation)“. Note that, as in the case of skewed symmetry, the coiiverse of this assumption is always true under orthographic projections. The above assumption can be schematized by Figure 3-l. Consider two texel patterns P, and P, in the picture, and place the origins of the x-y coordinates at their centers, respectively. The transform from P2 to P, can be expressed by a regular 2x2 matrix A = (aij). PI and P2 are projections of patterns P’, and P’, which are drawn on the 3-D surfaces. We assume that P’, and P’, are small enough SO that we can regard them as being drawn on small planes. Let us denote the gradients of those small planes by G, = (~1 ,ql) and G2 = (p2,q2), respectively; i.e., P’, is drawn on a plane-z=p,x+q,yandP’20n-z=p2x+q12y. 4 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Now, our assumption amounts to saying that P’, is transformable from .P’, by a scalar scale factor u and a rotation matrix R = (zynsz iz,“d,“,). (W e can omit the translation from our consideration, since for each pattern the origin of the coordinates is placed at its gravity center, which is invariant under the Affine-transform). Thinking about a pattern drawn on a small plane, -z= px + qy, is equivalent to viewing the pattern from directly overhead; that is, rotating the x-y-z coordinates so that the normal vector of the plane is along the new z-axis (line of sight). For this purpose we rotate the coordinates first bv cp around the y-axis and then by 8 around the x-axis. We have the following relations among v, 8, p, and q: sincp = p/m cosql = 1 /J&i (2) sine =9/&7&i c0se = J;;T;-i/ The plane which was represented as -z= px +qy in the old coordinates is, of course, now represented as -z’ =O in the new coordinates. Let us denote the angles of the coordinate rotations to obtain P’, and P’, in Figure 3-l by (cp, ,8,) and (‘p2,e2), individually. The 2-D mapping from P’i (xl-y’ plane) to Pi (x-y plane) can be conveniently represented by the following 2x2 matrix Tt which is actually a submatrix of the usual 3-D rotation matrix. Ti =( cyo -Si~W~“) Now, in order for the schematic diagram of Figure 3-1 to hold, what relationships have to be satisfied among the matrix A = (at)), the gradients Gi = (pilqi) for i = 1,2, the angles (pi, ei) for i = 1,2, the scale factor u, and the matrix R ? We equate the two transforms that start from P’, to reach at P,: one following the diagram counter-clockwise P’,->P->P, , the other clockwise P’,->P’,->P,. We obtain AT2 = T, aR. By eliminating u and a, and substituting for sines and cosines of pi and Bi by (2), we have two (fairly complex) equations in terms of pi, qi, and the elements of A. We therefore find that the assumption of Affine-transformable patterns yields a constraint determined solely by the matrix A. The matrix is determined by the relation between P, and P, observable in the picture: without CT/?= CT ( coti -5ind Sind ~0x4 > Figure 3-1: A schematic diagram showing the assumptions on Affine transformable patterns. knowing either the original patterns relationships (a and R) in the 3-D space. (P’, and P’) or their The Affine transform from P, to P, is more intuitively understood by how a pair of perpendicular unit-length vectors (typically along the x and y coordinate axes) are mapped into their transformed vectors. Two angles (a and /I) and two lengths (T and p) can characterize the transform. Components of the transformation matrix A = (aij) are represented by: a,, =7cosa a12=Pcos/? (3) a21 = Tsina a2 = psi@ Let us consider the case that a and p are known, but T and p are not. Using (3), eliminate T and p. Then, we obtain (Pl cosa + q 1 sinaNp, cosp + q, sir@) + cos(a-P) = 0 which is exactly the same as the hyperbola (1). . The Shape-from-Texture Paradigm This section derives the same skewed-symmetry constraints from a second theory, different from the Affine-transformable patterns. The shape-from-texture paradigm is briefly presented here; a futler discussion can be found in [3]. The paradigm has two major portions. In the first, a given image textural property is “normalized” to give a general class of surface orientation constraints. In the second, the normalized values are used in conjunction with assumed scene relations to refine the constraints. Only two texels are required, and only one assumption (equality of scenic texture objects, or some other simple relation) to generate a well-behaved one-dimensional family of possible surface orientations. The first step in the paradigm is the normalization of a given texel property. The goal is to create a normalized texture property map (NTPM), which is a representational and computational tool relating image properties to scene properties. The NTPM summarizes the many different conditions that may have occurred in the scene leading to the formation of the given texel. In general, the NTPM of a certain property is-a scalar-valued function of two variables. The two input variables describe the postulated surface orientation in the scene (top-bottom and left-right slants: (p,q) when we use the gradient space). The NTPM returns the value of the property that the textural object would have had in the scene, in order for the image to have the observed textural property. As an example, the NTPM for a horizontal unit line length in the image summarizes the lengths of lines that would have been necessary in 3-D space under various orientations: at surface orientation (p,q), it would have to be m. More specifically, the NTPM is formed by selecting a texel and a texel property, back-projecting the texel through the known imaging geometry onto all conceivable surface orientations, and measuring the texel property there. In the second phase of the paradigm, the NTPM is refined in the following way. Texels usually have various orientations in the image, and there are many different texel types. Each texel generates its own image-scene relationships, summarized in its NTPM. If, however, assumptions can be made to relate one texel to another, then their NTPMs can also be related; in most cases only a few scenic surface orientations can satisfy both texels’ requirements. Some examples of the assumptions that relate texels are: both lie in the same plane, both are equal in textural 5 measure (length, area, etc.), one is k times the other in measure, etc. Relating texels in this manner forces more stringent demands on the scene. If enough relations are invoked, the orientation of the local surface supporting two or nrore related texels can be very precisely determined. What we now show is that the skewed symmetry method is a special case of the shape-from-texture paradigm; it can be derived from considerations of texel slope. To normalize the slope of a texel, it is back-projected onto a plane with the postulated orientation. The angle that the back-projected slope makes with respect to the gradient vector of the plane is one good choice (out of many) for the normalized slope measure. Under perspective, the normalized value depends on the image position and camera focal length; under orthography it is much simpler. Using the construction in Figure 4-1, together with several lemmas relating surfaces in perspective to their local vanishing lines, slope is normalized as follows. Assume a slope is parallel to the p axis; the image and gradient space can always be rotated into such a position. (If rotation is necessary, the resulting NTPM can be de-rotated into the original position using the standard two-by-two orthonormal matrix.) Also assume that the slope is somewhere along the line y = ys, where the unit of measurement in the image is equal to one focal length. Then, the normalized slope value--the normalized texture property map -- is given by Es - Ys(P2 + cl211 ~dtl + P2 + s>1. This normalized value can be exploited in several ways. Most important is the result that is obtained when one has two slopes in the image that are assumed to arise from equal slopes in the scene. Under this assumptions, their normalized property maps can be equated. The resulting constraint, surprisingly, is a simple straight line in the gradient space. Under orthography, nearly everything simplifies. The normalized slope of a texel becomes q / [P/(1 f P2 + q>1. (4) lt is independent of Y,; in effect, all slopes are at the focal point. Consider Figure 2-1 (a). Given the angle that the two texels form (/?-a), rotate the gradient space so that the positive p axis bisects the angle. Let each half-angle be 6, so 6 = (/I-a)/2. Calculating the normalized value of either slope is obtained directly from the standard normalized slope formula, corrected for the displacement of + S and -6 respectively. That is, for the slope at the positive 6 orientation, instead of formula (4), we use the Figure 4-1: Back-projecting an image slope onto a plane with gradient (p, q). formula under the substitution pcosS + qsin6 for p, for q. We do similarly for the slope at -8. -psin6 qcosd The fact that the normalized slopes are assumed to be perpendicular in the scene allows us to set one of the normalized values equal to the negative reciprocal of the other. The resultant equation becomes p2cos2&q2sin26 = sin&cosS = -cos26. This is exactly the hyperbola (1) with 26 =/l-a. 5. Conclusion The assumptions we used for the skewed symmetry, the Affine-transformable patterns, and texture analysis can be generalized as “Properties observable in the picture are not by accident, but are projections of some preferred corresponding 3-D properties.” This provides a useful meta-heuristic for exploiting image properties: we can call it the meta-heuristic of non-accidental image properties. It can be regarded as a generalization of general view directions, often used in the blocks world, to exclude the cases of accidental line alignments. Instances that can fall within this meta-heuristic includes: parallel lines in the picture vs. parallel lines in the scene, texture gradients, and lines convergent to a vanishing point. One of the most essential points of our technique is that we related certain image properties to certain 3-D space properties, and that we map the relationships into convenient representations of shape constraints. We explicitly incorporate assumptions based either on the meta-heuristic or on apriori knowledge of the world. The shape-from-texture paradigm provides a computational framework for our technique. In most part of our discussion we assumed orthography. Similar--though more involved and less intuitive--results can be obtained under perspective projections. This work is further same title as this paper. .discussed in a technical report with PI [31 [41 6 References Kanade, T. Recovery of the 3-Dimensional Shape of an Object from a Single View. Technical Report CMU-CS-79-153, Computer Science Department, Carnegie-Mellon University, Oct., 1979. Kender, J.R. Shape from Texture. PhD thesis, Computer Science Department, Carnegie-Mellon University, 1980. Mackworth, A.K. Interpreting Pictures of Polyhedral Scenes. Artificial intelligence 4(2), 1973. Stevens, K.A. Surface Perception from Local Analysis of Texture and Contour. Technical Report AI-TR-512, MIT-AI, 1980.	1980	80
79	WHAT SHOULD BE COMPUTED IN LOW LEVEL VISION SYSTEMS William B. Thompson Albert Yonas University of Minnesota Minneapolis, Minnesota 55455 ABSTRACT Recently, there has been a trend towards developing low level vision models based on an analysis of the mapping of a three dimensional scene into a two dimensional image. Emphasis has been placed on recovering precise metric spatial information about the scene. While we agree with this approach, we suggest that more attention be paid to what should be computed. Pschophysical scaling, adaptation, and direct determination of higher order relations may be as useful in the per- ception of spatial layout as in other perceptual domains. When applied to computer vision systems, such processes may reduce dependance on overly specific scene constraints. L* Introduction The following is a position paper directed at several aspects of low-level visual processing. The current trend towards focusing on the determi- nation of exact, three-dimensional form in a scene is questioned. Both analysis of representative scene domains and experience with human vision sug- gest that less precise form properties may be suf- ficient for most problems. Several computational issues are also briefly discussed. 2. Alternate Approaches to "Low-Level" Analysis - -- Computer vision systems have traditionally been divided into segmentation and interpretation components. A multiplicity of image features have been investigated in the hope that they would facilitate the partitioning of an image into regions corresponding to "objects" or "surfaces" in the original scene. Only after this two- dimensional segmentation operation was completed would procedures be applied in an attempt to deter- mine the original three-dimensional structure of the scene. Recently, an alternative approach to implementing the lower levels of a computational vision model has been developed. Its basic premise This research was supported in part by the Nation- al Science Foundation under Grant MCS-78-20780 and by the National Insitute of Child Health and Human Development under Grants HD-01136 and HD-05027. is that the determination of three-dimensional structure is such an integral part of the scene description processes that it should be carried out at all levels of the analysis [1,2,3,4,51. Proponents of this approach usually employ a well structured methodology for developing computational models of form perception: Precisely describe a carefully con- strained scene domain. Identify important scene properties. Determine the function which maps these scene properties into an-image. Develop computationally feasible mechan- isms for recovering the "important" scene properties from the image. Great emphasis is placed on determining what scene properties are computable, given a set of con- straints on the scene. Scene properties normally considered essential to the analysis include object boundaries, three- dimensional position, and surface orientation. In many cases, the determination of these features requires that properties such as surface reflec- tance and illumination must also be found. A key distinction between the classical techniques and this newer approach is that in the latter, analysis procedures are developed analytically from an understanding of how scene properties affect the image rather than ad hoc assumptions about how image properties might relate to scene structure. The representational structures which have been used to implement form based analysis have, for the most part, been iconic. The features represented are almost always metric properties of the corresponding point on the surface: distance from the observer, orientation with respect to either the observer or a ground plane, reflectance, incident illumination, and so on. To determine relative effects (eg. which of two points is farther away), absolute properties are compared. The determination of these metric scene pro- perties requires that the possible scenes be highly constrained. Usually, the analysis depends on res- trictions both on the types of objects allowed and on the surface properties of the objects. For From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. example, a "blocks world" assumption (or alter- nately, the assumption of a "Play-Doh" world made entirely of smooth surfaces) might be made. In addition, it is commonly assumed that surfaces are all lambertian reflectors and that, for a given surface, the coefficient of reflectance is constant. Illumination is often limited to a sin- gle distant point source, possibly coupled with a diffuse illuminator. Secondary illumination effects are usually presumed to be negligible. 20 Absolute Scene Properties Are Not Always Needed -m The proponents of form based analysis presume the need for finding exact shape properties of a scene. They concentrate on investigating how con- straints on scenes affect what properties are com- putable and how they can be determined. We suggest that more attention be paid towards what properties should be computed. We argue that for a wide variety of problem areas, absolute metric informa- tion about scene shape is not required. Instead, relative properties such as flat/curved, convex/concave, farther-away/closer, etc. are both sufficient and easier to compute. Most tasks involving description of a visual environment depend on generalized shape properties. In fact, much effort has been spent searching for shape characterizations that embody those relation- ships useful for description but not the enormous amount of irrelevant detail contained in any representation based on specific position. Even in task domains such as object manipulation and obsta- cle avoidance, precise positional information is frequently not necessary. Both these task areas contain significant sub-problems involving object identification - a descriptive task often possible with approximate and/or relative information about shape. Even when actual position is needed, feed- back control can be used to minimize the need for highly accurate positional determinations. A second argument for emphasizing the diffi- culty of determining metric properties comes from our experience with human perception. The psycho- logical literature contains many references to the effects of the scaling process that relates the physical domain to the psychological [6,71, the effects of adaptation to stimulation [8], and the effects of practice on variable error [91. By investigating the competence of the human visual system in determining primitive shape effects, we can gain insight into sufficient (but not neces- sary) properties for more complex analysis. In our own work on perception of surfaces, preliminary results from one set of experiments seem relevant to the development of computational models. We synthesized a frontal view of a surface the profile of which is shown in figure 1. Lighting was assumed to be a combination of a single distant point source and a perfectly diffuse source. A simple reflectance model was used and secondary illumination effects were not considered. A series of synthesized images was produced with the inten- tion of examining the perception of single displays and the ability to determine differences between displays. The "object" in our images was an ellip- soid with semi-axes A, B, and C. (A was in the horizontal direction as seen by the viewer, B was in the vertical direction, and C was in the direc- tion along the line of sight.) The object was presented against a black background, and thus no cast shadows were present. In one set of experi- ments, A and B were held constant producing a cir- cular occluding contour. Subjects were asked to estimate the value of C for a number of different displays, with true values of C ranging from one half of A to four times A. On initial trials, sub- jects tended to see the same shape independently of the actual value of C. On subsequent trials, per- formance improved, but with a significant, sys- tematic underestimation of the true value. As a final note, when subjects were asked to qualita- tively describe the changes in the scene as C was varied, they often indicated that they felt that the change was due to differences in illumination, not shape. It is certainly premature to make any defini- tive conclussions from our results. Nevertheless, we suggest the following conjecture: Subjects appear to see a specific shape (as opposed to sim- ply a "round" object); however, the metric proper- ties they estimate for that shape are not neces- sarily consistent with the "true" values. The sub- jects do appear to be better at ranking displays based on different values of C. 4* --- Non-metric Scene Properties We suggest that requiring specific, accurate determination of scene properties may be unneces- sarily restrictive. Less precise and/or purely relative qualities are sufficient for many situa- tions. By concentrating on these characteristics, we may be able to significantly relax the con- straints under which our computational vision models must operate. Finally, human vision is often quite inaccurate in determining metric values for these same properties. Rather than indicating a deficiency in human vision, this suggests that alternative (and presumably more useful) charac- teristics are being computed by the human per- ceiver. Two approaches to structuring computer vision models based on these observations seem relevant. First of all, it may be possible to directly com- pute properties of interest, rather than deriving them from more "primitive" characteristics (see [lO,lll). For example, we might look for ways of estimating surface curvature that do not depend on first determining depth and then taking the second derivative. A second possibility is to presume that esti- mation of shape properties is subject to the same scaling processes as most other perceptual phenomena. Thus, our model would estimate some non-linear but monotonic transformation of charac- teristics such as depth. The transformations would be adaptive, but in general not known by higher level analysis procedures. Thus, the precise metric three-dimensional structure can not be recovered. For many tasks, the scaled values are sufficient and the need for highly constrained, 8 photometric analysis of the image is reduced. With luminence gradient, illumination direction, and appropriate standardization, precise scene proper- surface curvature. For a given gradient, knowing ties may be determined. Without standardization, either illumination or curvature allows determina- relative characteristics are still computable. tion of the other. The model must be able to Ordinal relationships are detrminable over a wide account for this symmetry. range while quantatative comparisons are possible over a more limited range. (eg. it may be possible 6. Conclusions to judge that A is twice as far as B but not that C is 100 times as far as D.) When attempting to construct computational models of low-level vision systems, we need to pay 50 Computational Models as much attention to what should be computed as we do to how it is computed. We may investigate this Recently, much attention has been focused on problem in at least three ways. The first is a using parallel process models to specify the compu- computational approach: we can determine what is tational structure of low-level vision systems. An computable given a set of constraints about the image is partitioned into a set of neighborhoods, scene and the imaging process. The second is an with one process associated with each region. The ecological approach: we catalog the range of prob- processes compute an estimate of scene properties lem domains in which our system is expected to corresponding to the region using the image function and then determine the primitive scene features in the region and whatever is known about properties needed for analysis. The third is meta- surrounding scene structure. The circularity of phorical: study a working visual system (eg. form estimation for one point depending on the form human) in order to determine which low-level scene of neighboring points can be dealt with in several properties it is able to perceive. These proper- ways. A variable resolution technique may be ties then define a sufficient set for analysis. employed. First large, non-interacting neighbor- hoods are used. Then, progressively smaller neigh- Much current work focuses on estimating exact borhoods are used, each depending on scene proper- positional information about a scene* We argue ties computed using previously analyzed, larger that in many cases, these metric properties cannot regions. (Marr's stereo model is an example [121.) be easily determined. Even more importantly, how- Alternately, an iterative technique can be used to ever, they often need not be determined. Simple find crude estimates of scene properties and then relative properties may be sufficient for analysis those values are fed back into the process to pro- and be much easier to compute. duce more refined estimates. (Examples include many "relaxation labeling" applications [131.) In either case, the determination of absolute scene properties usually requires a set of boundary values - image points at which the scene con- BIBLIOGRAPHY straints allow direct determination of the proper- ties. The computational process must then pro- pagate these constraints to other image regions. [II D. Marr, "Representing and computing visual The robustness of these parallel process information", Artificial Intelligence: An MIT -- models may be significantly increased if they are Perspective, P.H. Winston and R.H. Brown, only required to compute relative properties. The ed-, PP. 17-82, 1979. need for accurately propagating scene information is greatly reduced. Furthermore, photometric [21 H.G. Barrow and J.M. Tennenbaum, "Recovering analysis of the image will usually not be required. intrinsic scene characteristics from images," For instance, general characteristics of the inten- in Computer Vision Systems, A.R. Hanson and sity gradient may be all that is required for E-M. Riseman, eds., New York: Academic Press, analysis. As an example, for a reasonably general 1978. class of scene types, a discontinuity in the luminence gradient will usually correspond to a [31 B. Horn, "Obtaining shape from shading shadow, an occlusion boundary, or the common boun- information," in The Psychology of Computer dary between two surfaces. Continuous but non-zero Vision, P.H. Winston, ed., New York= McGraw- gradient values indicate either surface curvature Hill, 1975. or illumination variation. In neither case is the actual magnitude of the gradient required. [41 S. Ullman, The Interpretation of Visual Motion, Cambridge: MIT Press, 1979: Finally, many of the problems in low-level vision are underspecified. No single "correct" [51 K-A. Stevens, "Surface perception from local solution exists because insufficient information is analysis of texture and contour", Ph.D. available to derive the original scene properties. Thesis, MIT, Feb. 1979. Thus, computational models must either naturally embody default assumptions or allow for ambiguous 161 GOT. Fechner, Elemente der Psychophysik, representations. (There is reason to expect that Leipzig: Breitkopf and Hartel, 1860. both approaches are useful.) Even more important, (Reissued Amsterdam: Bonset, 1964.) the control structures used by the models must not impose any arbitrary input/output assumptions. For example, consider again the relationship between 9 [73 S.S. Stevens, "Perceptual magnitude and its measurement," in Handbook of Perception, &. II, Psychophysical JudgemeT;;& Measurement, Carterette and Friedman, eds., New York: Academic Press, 1974. [al H. Helson, Adaptation Level Theory, New York: Harper, 1964. Dl E.J. Gibson, Perceptual Learning and Develop- ment, New York: Appleton-CenGy-Crofts, 1969. [lo] J.J. Gibson, The Senses Considered as Visual -- Systems, Boston: Houghton Mifflin, 1966. [ll] J.J. Gibson, The Ecological Approach to Visual Percept=, Boston: Houghton Miffli. 1979. [121 DO Marr and TO Poggio, "A theory of human stereo vision," MIT AI Lab. MEMO 451, Nov. 1977. [131 A. Rosenfeld, R. Hummel, and S. Zucker, "Scene labeling by relaxation operations," IEEE Trans. Systems, &, PP and Cybernetics, vol. 6, ppe 420-433, June 19x .’ 1, - . 'i 7 0 . /I I' P Figure 1. 10	1980	81
80	INTERPRETING LINE DRAWINGS AS THREE-DIMENSIONAL SURFACES Harry G. Barrow and Jay M. Tenenbaum Artificiai Inteiiigence Center SRI Internationai, Menio Park, CA 94025 ABSTRACT We propose a computationai modei for interpreting iine drawings as three-dimensionai surfaces, based on constraints on iocai surface orientation aiong extremai and discontinuity boundaries. Specific techniques are described for two key processes: recovering the three- dimensionai conformation of a space curve (e.g., a surface boundary) from its two-dimensionai projection in an image, and interpoiating smooth surfaces from orientation constraints aiong extremai boundaries. INTRODUCTION Our objective is the deveiopment of a computer modei for interpreting two-dimensionai iine drawings, such as Figure 1, as three-dimensionai surfaces and surface boundaries. Line drawings depict intensity discontinuities at surface boundaries, which, in many cases, are the primary source of surface information avaiiabie in an image: i.e., in areas of shadow, compiex (secondary) iiiumination, or specuiar surfaces anaiytic photometry is inappropriate. Understanding how iine drawings convey three- dimensionaiity is thus of fundamentai importance. Given a perspectiveiy correct iine drawing depicting discontinuities of smooth surfaces, we desire as output arrays containing vaiues for orientation and reiative range at each point on the impiied surfaces. This objective is distinct from that of eariier work on interpretation in terms of object modeis (e.g. [I]) and more basic. No knowiedge of piants is required to understand the three-dimensionai structure of Figure 1, as can be demonstrated by viewing fragments out of context (through a mask, for exampie). Ambiguity and Constraints The centrai probiem in perceiving iine drawings is one of ambiguity: in theory, each two- dimensionai iine in the image couid correspond to a possibie projection of an infinitude of three- dimensionai space curves (see Figure 2). Yet peopie are not aware of this massive ambiguity. When asked to provide a three-dimensionai interpretation-of an eiiipse, the overwheiming response is a tiited circie, not some bizarrely twisting curve (or even a discontinuous one) that has the same image. What assumptions about the scene and the imaging process are invoked to constrain to this unique interpretation? * This research was supported by funds from DARPA, NASA, and NSF. We observe that aithough aii the iines in Figure 1 iook fundamentaiiy aiike, two distinct types of scene event are depicted: extremai boundaries (e.g., the sides of the vase), where a surface turns smoothiy away from the viewer, and discontinuity boundaries (e.g., the edges of the ieavss), where smooth surfaces terminate or interswt. Each type provides different constraints on three-dimensionai interpretation. At an extremai boundary, the surface orientation can be inferred exactiy; at every point aiong the boundary, orientation is normai to the iine of si ht and to the tangent to the curve in the image d. A discontinuity boundary, by contrast, does not dirsctiy constrain surface orientation. However, its iocai curvature in the image does provide a statisticai constraint on the three- dimension& tangent of the corresponding space curve. The iocai surface normai is constrained oniy to be orthogonai to this tangent, and is thus free to swing about it as shown in Figure 3. The abiiity to infer 3-D surface structure from extremai and discontinuity boundaries suggests a three-step modei for iine drawing interpretation, anaiogous to those invoived in our intrinsic image modei [2]: iine sorting, boundary interpretation, and surface interpoiation. Each iine is first ciassified according to the type of surface boundary it represents (i.e., extremai versus discontinuity). Surface contours are interpreted as three-dimensionai space curves, providing reiative 3-D distances aiong each curve; iocai surface normais are assigned aiong the extremai boundaries. Finaiiy, three-dimensionai surfaces consistent with these boundary conditions are constructed by interpoiation. (For an aiternative model, see Stevens [3].) This paper addresses some important aspects of three-dimensionai recovery and interpoiation (see [l] and [4] for approaches to iine ciassification). INTERPRETATION OF DISCONTINUITY BOUNDARIES To recover the three-dimensionai conformation of a surface discontinuity boundary from its image, we invoke two assumptions: surface smoothness and generai position. The smoothness assumption impiies that the space curve bounding a surface wiii aiso be smooth. The assumption that the scene is viewed from a generai position impiies that a smooth curve in the image resuits from a smooth curve in space, and not from an accident of viewpoint. In Figure 2, for exampie, the sharpiy receding curve projects into a smooth eiiipse from oniy one viewpoint. Thus, such a curve wouid be a highiy improbabie three-dimensionai interpretation of an eiiipse. 11 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. The problem now is to determine which smooth space curve is most iikeiy. For the speciai case of a wire curved in space, we conjectured that, of aii projectiveiy-equivaient space curves, humans perceive that curve having the most uniform curvature and the ieast torsion [2]; i.e., they perceive the space curve that is smoothest and most planar. Consistent findings were reported in recent work by Witkin [5] at MIT on human interpretation of the orientation of pianar ciosed curves. Measures of Smoothness - The smoothness of a space curve is expressed quantitativeiy in terms of intrinsic characteristics such as differentiai curvature (k) and torsion (t), as weii as vectors giving intrinsic axes of the curve: tangent (T), principai normai (N), and binormai (B). A simpie measure for the smoothness of a space curve is uniformity of curvature. Thus, one might seek the space curve corresponding to a given image curve for which the integrai of k' (the spatiai derivative of k) was minimum. This aione, however, is insufficient, since the integrai of k' couid be made arbitrariiy smaii by stretching out the space curve so that it approaches a twisting straight iine (see Figure 4). Uniformity of curvature aiso does not indicate whether a circuiar arc in the image shouid correspond to a 3-D circuiar arc or to part of a heiix. A necessary additionai constraint in both cases is that the space curve corresponding to a given image curve should be as pianar as possibie, or more preciseiy that the integrai of its torsion shouid aiso be minimized. Integrai 1 expresses both the smoothness and pianarity of a space curve in terms of a singie, iocaiiy computed differentiai measure d(kB)/ds: d(kB/ds)*ds = (kg2 + k2t2)ds (1) Intuitiveiy, minimizing this integrai corresponds to finding the three-dimensionai projection of an image curve that most cioseiy approximates a pianar, circular arc, for which k' and t are both everywhere zero. Recovery Techniques A computer modei of this recovery theory was impiemented to test its competence. The program accepts a description of an input curve as a sequence of two-dimensionai image coordinates. Each input point, in conjunction with an assumed center of projection, defines a ray in space aiong which the corresponding space curve point is constrained to iie. The program can adjust the distance associated with each space curve point by siiding it aiong its ray iike a bead on a wire. From the resuiting 3-D coordinates, it can compute iocal estimates for curvature k, intrinsic axes T, N, and B, and the smoothness measure d(kB)/ds. An iterative optimization procedure then adjusts distance for each point to determine the configuration of points that minimize the integrai in (1). The program was tested using input coordinates synthesized from known 3-D space curves so that resuits couid be readiiy evaiuated. Correct 3-D interpretations were produced for simpie open and ciosed curves such as an eiiipse, which was interpreted as a tiited circle, and a trapezoid, which was interpreted as a tiited rectangie. However, convergence was siow and somewhat dependent on the initiai choice of z-vaiues. For exampie, the program had difficuity converging to the "tiited-circie" interpretation of an eiiipse if started either with aii z-vaiues in a piane paraiiei to the image piane or aii randomized to be highiy nonpianar. SURFACE INTERPOLATION Given constraints on orientation aiong extremai and discontinuity boundaries, the next probiem is to interpoiate smooth surfaces consistent with these boundary conditions. The probiem of surface interpoiation is not pecuiiar to contour interpretation, but is fundamentai to surface reconstruction, since data is generaiiy not avaiiabie at every point in the image. We have impiemented a soiution for an important case: the interpoiation of approximateiy uniformiy-curved surfaces from initiai orientation values and constraints on orientation [6]. The input is assumed to be in the form of sparse arrays, containing iocai estimates of surface range and orientation, in a viewer-centered coordinate frame, ciustered aiong the curves corresponding to surface boundaries. The desired output is simpiy fiiied arrays of range and surface orientation representing the most iikeiy surfaces consistent with the input data. These output arrays are anaiogous to our intrinsic images [2] or Marr's 2.5D sketch [7]. 12 For any given set of input data, an infinitude of possibie surfaces can be found to fit arbitrariiy weii. Which of these is best (i.e., smoothest) depends upon assumptions about the nature of surfaces in the worid and the image formation process. For exampie, surfaces formed by eiastic membranes (e.g., soap fiims) are constrained to minimum energy configurations characterized by minimum area and zero mean curvature; surfaces formed by bending sheets of ineiastic materiai (e.g., paper or sheet metai) are characterized by zero Gaussian curvature; surfaces formed by many machining operations (e.g., pianes, cyiinders, and spheres) have constant principai curvatures. Uniformiy Curved Surfaces We concentrate here on surfaces that are iocaiiy sphericai or cyiindricai (which have uniform curvature according to any of the above criteria). These cases are important because they require reconstructions that are symmetric in three dimensions and independent of viewpoint. Many simpie interpoiation techniques faii this test, producing surfaces that are too fiat or too peaked. An interpoiation aigorithm that performs correctiy on sphericai and cyiindricai surfaces can be expected to yieid reasonabie resuits for arbitrary surfaces. Our approach expioits an observation that components of the unit normai vary iineariy across the images of surfaces of uniform curvature. Consider a three-dimensionai sphericai surface, as shown in Figure 5. The radius and normai vectors are aiigned, and so from simiiar figures we have: Nx = x/R, Ny = y/R, Nz = z/R . A simiiar derivation for the right circuiar cyiinder is to be found in [6]. The point to be noted is that for both the cyiinder and the sphere, Nx and Ny are iinear functions of x and y, and Nz can be derived from Nx and Ny. An Interpoiation Technique - We have impiemented an interpoiation process that expioits the above observations to derive the orientation and range over a surface from boundary vaiues. It uses paraiiei iocai operations at each point in the orientation array to make the two observabie components of the normai, Nx and Ny, each vary as iineariy as possibie in both x and y. This couid be performed by a standard numericai reiaxation technique that repiaces the vaiue at each point by an average over a two-dimensionai neighborhood. However, difficuities arise near surface boundaries where orientation is discontinuous. We decompose the two-dimensionai averaging process into severai one-dimensionai ones, by considering a set of iine segments passing through the centrai point, as shown in Figure 6a. Aiong each iine we fit a iinear function, and thus estimate a corrected vaiue for the point. The independent estimates produced from the set of iine segments are then averaged. Oniy the iine segments that do not extend across a boundary are used: in the interior of a region, symmetric iine segments are used (Figure 6a) to interpoiate a centrai vaiue; at boundaries, an asymmetric pattern aiiows vaiues to be extrapoiated (Figure 6b). The interpoiation process was appiied to test cases in which surface orientations were defined around a circuiar outiine, corresponding to the extremai boundary of a sphere, or aiong two paraiiei iines, corresponding to the extremai boundary of a right circuiar cyiinder. Essentiaiiy exact reconstructions were obtained, even when boundary vaiues were extremeiy sparse or oniy partiaiiy constrained. Resuits for other smooth surfaces, such as eiiipsoids, seemed in reasonabie agreement with human perception. Current work is aimed at extending the approach to partiaiiy constrained orientations aiong surface discontinuities, which wiii permit interpretation of generai soiid objects. REFERENCES 1. K. Turner, "Computer Perception of Curved Objects Using a Teievision Camera," Ph.D. thesis, Department of Machine Inteiiigence and Perception, University of Edinburgh, Edinburgh, Scotiand (1974). 3. K. Stevens, "Constraints on the Visuai Interpretation of Surface Contours," A.I. Memo 522, M.I.T., Cambridge, Massachusetts (March 1979). 4. I. Chakravarty, "A Generaiized Line and Junction Labeiing Scheme with Appiications to Scene Anaivsis." IEEE Transactions on Pattern ” , -- Anaiysis and MachTInteiiigence, Voi. PAMI- 1, No. 2,(Aprii,79). 59 A. Witkin, Department of Psychoiogy, M.I.T., Cambridge, Massachusetts (private communication). 6. H. G. Barrow and J. M. Tenenbaum, "Reconstructing Smooth Surfaces from Partiai, Noiss Information," Proc. ARPA Image Understanding Workshs.S.C.,LosAngeies, Caiifornia (Fam. 7. D. Marr, "Representing Visuai Information," in Computer Vision Systems, A. Hanson and E. Riseman, eds.(Academic Press, New York, New York, 1978). 13 FIGURE 1 LINE DRAWING OF A THREE-DIMENSIONAL SCENE (Surface and boundary structure are dis- tinctly perceived despite the ambiguity inherent in the imaging process.) FIGURE 2 THREE-DIMENSIONAL CONFORMATION OF LINES DEPICTED IN A LINE DRAWING IS INHERENTLY AMBIGUOUS (All of the space curves in this figure project into an ellipse in the image plane, but they are not all equally likely FIGURE 3 AN ABSTRACT THREE-DIMENSIONAL SURFACE CONVEYED BY A LINE DRAWING (Note that surface orientation is constrained to one degree of freedom along discontinuity boundaries.) FIGURE 4 AN INTERPRETATION THAT MAXIMIZES UNIFORMITY OF CURVATURE y AXIS x AXIS FIGURE 5 LINEAR VARIATION OF N ON A SPHERE (a) symmetric (b) asymmetric FIGURE 6 LINEAR INTERPOLATION OPERATORS interpretations.) 14	1980	82
81	SHAPE ENCODING AND SUBJECTIVE CONTOURS’ Mike Brady, W. E. L. Grimson Artificial Intelligence Laboratory, MIT. and D. J. Langridge Division of Computing Research, CSIRO, Canberra, ACT, Australia. 1. Int reduction Ullman [15] has investigated the shape of subjective contours (see for example [7]. [4]. [5], (121). In fact, the work is more generally ap- plicable to other cases of pcrccptual shape completion, in which the visual system is not constrained by actual physical intensity changes. Examples include patterns foimcd from dots line drawings and alphabetical characters. and incomplctcly Ullman proposes that subjective contc,.us consist of two circles which meet smoothly and which arc tangential to the contrast bound- aries from which they originate. The foim of the solution derives from a number of premises, one of which Ullman calls “the locality hypothesis”. ‘Ihis is “based in part on cxpcrimcntal obscrvntions, and partly on a theoretical consideration” ([I 51 ~2). ‘I’hc “cxpcrimcntal ob- servation” rcfcrrcd to is the following: suppose that A’ is a point near A on the filled-in contour AB as shown in Figure 1. If the process by which AB was constructed is applied to A’B, it is claimed that it gcneratcs the portion of AJ3, bctwccn A’ and B. Let us call this property “cxtcnsibility”. Ullman argues that cxtcnsibility, togcthcr with the propcrtics of isotropy, smoothness, and having minimal integral cur- vature, logically entails a solution consisting of two circles which meet smoothly. In the fit11 version of this paper, we analyze the two-cti solution and formulate the condition for minimal integral curvature. This can bc solved by any descent method such as Newton-Raphson. A program has been written which computes the minimum integral curva- ture two-circle solution given the boundary angles 4, 8, and AB, and which returns as a result the point at which they meet and at which the curvature is discontinuous (the ‘knot point’). A tortuous version of this simple proof and program recently appcarcd in [13]. We then show by example that the two circle solution is not in fact extensible. Figure 1. Ullman’s extensibility property. If the process by which AD was constructed is applied to A’B, portion of AB between A’ and B. it is claimed that it generates the lntcrestingly, Knuth [S]. in his discussion of mathematical typog raphy, initially proposes that “most plcasing curves” should be ex- tensible, isotropic, cyclically symmetric (that is, the solution should not depend on the order of the contributing points), smooth, and be generable by a local (in his case four point) process. In fact, Knuth ([8]. Lemma 1) shows that these requirements arc mutually inconsistent, and hc argues that it is most reasonable to drop the assumption of ex- tensibility. He chooses the mathematically convenient process of cubic spline fitting as the basis of his character typography. This is an in ing choice in the light of what follows here. We conclude from the above that either Ullman’s solution is rect, despite the erroneous justification he gives, in which case should like to be able to offer an alternative derivation, or else incorrect, in which case we should like to propose an alternative. For a longer discussion of the various suggestions regarding completion which have appeared in the litcraturc, the reader is rc to [ll. Often one finds that they are singularly lacking in justifi beyond the vague claim that they seem to generate “rcasonablc” sham On a more precise note, Pavlidis([ 111, chapters 7 and 8) surveys the many versions of polygonal approximation, spline fitting, and Fourier domain descriptors which have been proposed in the considcra tern recognition literature Two criteria have tended to domit selection of curves for shape completion, namely mathcmatic tability and computational efficiency, the latter implicitly assuming some level of computational power on conventional computers. The enor- mous computational power of the human perceptual system advocates caution in the application of such arguments. Indeed, a somewhat different approach is to base the sclcction of a class of curves on an analysis of some problem that the human visual system is trying to solve, and to isolate and examine the constraints within which that problem is posed. Several examples of this approach have app over the past few years, mainly in the work of Marr and his cd- laborators (see [9]). The work of Grimson, described briefly in Section 2. is of this sort. Ullman ([I61 section 3.3) makes some prclimi comments about the biological feasibility of pcrccptual computations, One of the difficulties which surrounds the choice of a class of curves in the specific case of subjective contours is that the differences between altcrnativc choices is often quite small at the angular extent of most examples of subjective contours. This in turn means that the issue is very difficult to resolve by psychophysical experimentation of the sort used by Ullman ([lS] page 3). In [I], we pursue an alternative approach, which was inspired by the second section of Ullman’s paper [ 151. He develops a local algorithm to compute the two circle solution which minimizes integral absolute 15 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. curvature. This naturally suggests dispensing with the cxtcnsibility as- sumption cntircly, just as Knuth did, and proceeding to find a solu- tion which minimizes some “pcrformancc index” related to curvature n, such as abs(n) or $. In order to test this idea, we apply the ideas of modern control theory (see [14],[2]). close approximation to curvature when the gradient fi is small. This is Since IC and abs(K) are non-conservative, WC consider minimizing $. We develop the Hamiltonian and show that it leads to a particularly intractable diffcrcntial equation namely, a reasonable condition to impose in the case of point which we argue at greater length in PI. subjective contours, a WC proceed to set up the Hamiltonian as usual. The boundary conditions arc as follows: (1) z = 0, y = 0, w = 4 (2) z = t, y = 0, w = -0 = yt. where w = 2, ?~n = 9, and the plant equations are: Yl = w WI =u (= &) (1) while the given by pcrformancc index is $. ‘I’hc Clamiltonian state function is where h is the Lagrange multiplier, and u is a constant of integration. This almost certainly does not have a closed form analytical solution. One possible iinc of approach at this juncture would be to base a local U2 Setting $$& equal to zero in the usual way, and solving 8$ = XI and g = -pl, leads to computation on one of the known techniques for the numerical solution of ordinary differential equations. Although shooting methods [3] are an obvious method on which yet explored the idea. to base such a local computation, WC have not ~=AxZ-p, In order to proceed, WC suggest in Section 2 that the shape com- pletion problem considcrcd hcrc is a two-dimcnsionnl analoguc of the problem of interpolating a three-dimensional surface, for example from the rclativcly sparse set of disparity points gcncratcd by the Marr- Poggio [IO] stereo algorithm. ‘I’his problem has rcccntly been studied where p is a constant of integration. This integrates easily to yield the cubic solution 3 2 y+-$+ux+T, where u and r are further constants of integration. Inserting the bound- ary conditions enables us to solve for h, p, u, and r. We get finally by Grimson [6]. He proposes that the pcrformancc index should be a semi-norm and shows that many of the obvious pcrformancc indices rclatcd to curvature, such as $ and abs(lc), do not have this property. He notes that the quadratic variation, defined by f2,* in two dimensions, is not only a semi-norm but is a close approx imation to curvature when the gradicn t fi is small (subscripts indicate partial dcrivativcs). ‘This is a rcasonablc condition to i::lpose in the case of subjcctivc contours. Accordingly, WC set LIP the 1 lomiltonian for the quadr‘rtic variation and show that it lcads to a cubic, which rcduccs to a parabola in the case of equal angles. This is particularly intcrcsting in view of the comments x3 x2 y = t;l(tanfI - tanf$) + t(tan$ - 2 tan@ + ztan0. In Figure 2, WC showed Ullman’s solution for the case 0 = 30”, 4 = 20’. In Figure 3 WC show the curve generated by our method, and in Figure 4 WC overlay the two. Clearly, the diffcrencc between the solutions is quite small at the angular extent shown in the figures. (Further examples arc given in [I]). As WC pointed out in the Introduction, this limits the usefulness of the kind of psychophysical experimentation used by Ullman to decide which solution is adopted by humans. made about Knuth’s work above. 2. Minimizing quadratic variation In order to proceed, we suggest that the shape completion problem In cast the angles 4 and 0 are equal, solution reduces to a parabola, namely the cubic term is zero and the considered here is a two-dimensional analoguc of the problem of inter- polating a three-dimensional surface, for example from the relatively 2 y= -Z-tanB+ztanf?. t sparse set of disparity points generated by the Marr-Poggio [lo] stereo algorithm. More generally, WC suggest that the process by which subjec- tive contours are gcncratcd results from the “mis-application” to two- The apex is at (1, 6 tan 8) and the focal length is &. Hence the focus is below the lint AB so long as 0 < 4, which is normally the case for subjective contours. dimensional figures of a process whose main purpose is the interpola- tion of three-dimensional surfaces. This idea requires some justification, which space here does not permit (see [l]). It also gives a different perspective on subjective contours, and leads to some fascinating, tes- table hypotheses. In the cast 0 = 4, it is straightforward to compute the difference between the parabola gcncrated by our method and the circle predicted by Ullman, which is The three-dimensional interpolation problem has recently been studied by Crimson [6] in the context of human stcreopsis. He observes that a prcrcquisite to finding the optimal fimction satisfying a given set of boundary conditions (namely that they should all pass through the given set of sparse points), is that the functions be comparable. Translating this into mathematics, hc proposes that the performance in- dex should be a semi-norm. Most importantly, hc shows that many of the obvious performance indices related to curvature, such as $ and Using the approximation (1 - abs(K), do not have this property. He notes that the quadratic variation, defined by f2,, in two dimensions, is not only a semi-norm but is a 16 this reduces to [4] Coren “Subjective contours and apparent depth,” Psychol. Review 79 (1972) 359-367. v - 4 II= -ztanB+ wz2tan30. t The diffcrencc bctwcen the solutions is essentially given by the second term whose maximum is at z = 4. Dividing by the horizontal extent t of the curve, the relative difference is bounded by w. For 8 < 1 this is negligible. 3. Acknowledgements This report describes research done in part at the Artificial Intelli- gence Laboratory of the Massachusetts lnstitute of Technology. Support for the laboratory’s artificial intelligence research is provided in part by the Advanced Research Projects Agency of the Dcpartmcnt of Defense under Office of Naval Research contract N00014-75-C-0643. The authors would like to thank the following people for valuable discussions at various stages of the research described here: John Greenman, Pat Hayes, David Marr, Slava Prazdny, Chris Rowbury. Robin Stanton, and Shimon Ullman. * This paper will appear as MIT AI memo 582. Here intelligibility and content have been sacrificed in the intcrcsts of strict page limit brevity. 4. References [I] Rrady, Crimson, and Langridge “Shape encoding and subjective con- tours,” (1980) to apEear. 1[2] Hryson and Ho Applied oplimal confrol, Ginn, Waltham MA, 1969. [3) Come and de Boor Elementary numerical analysis , McGraw Hill, Tokyo, 1972. [5) I;risby and Clatworthy “Illusory contours: curious cases of simul- tancous brightness contrast., 3 ” Perceprion 4 (1975), 349-357. [6] Grimson Computing shape using a theory of human stereo vision, Ph.D. ‘I’hcsis, Department of Dept. of Mathematics, MIT, 1980. [7] Kanisza “Subjective contours,” Sci. Artrer. 234 (1976), 48-52. ]$I Knuth “Mathematical typography,” Bull. Amer. M&h. Sot. (new series) 1(1979), 337-372. 19) Marr Vision, Freeman, San Francisco, 1980. [LO) Marr and Poggio “A theory of human stereo vision,” Proc. R. Sot. Lond B 204 (1979) 301-328. [Ill Pavlidis Slrucrural Partern Recognition , Springer Verlag, Berlin, 1977. ]I4 Rowbury Apparent contours, Univ. of Essex, UK,, 1978. ]13] Wutkowski “Shape completion,” Computer Graphics and Image Processing 9 (1979), 89-101. 1141 Schultz and Melsa State functions and linear control systems , McGraw Hill, New York, 1967. [IS] Ullman “Filling the gaps: The shape of subjective contours and a model for their generation,” Biof. Cyb. 25 (1976) l-6. 1161 Ullman “Relaxation and constrained optimisation by local processes,” Compuler Graphics and Image Processing IO (1979). 115-125. . . . . . . . . . Figure 2. Ullman’s solution for the case of boundary angles 0 = 30”, 4 = 20”. Figure 3. The solution generated by the method proposed in this paper. The boundary conditions are the same as those of Figure 2. ........ ....... ........ .... ..... .... ..... ... ... ...... . ... . ......... . . . . ........... . , ............. .............. . . ........... . . . . . Figure 4. The solutions of Figure 2 and Figure 3 are overlayed to demonstrate the difference between them. 17	1980	83
82	A STATISTICAL TECHNIQUE FOR RECOVERING SURFACE ORIENTATION FROM TEXTURE IN NATURAL IMAGERY Andrew P. Witkin Artificial Intelligence Center SRI International, Menlo Park, CA 94025 ABSTRACT A statistical method is reported for inferring the shape and orientation of irregularly marked surfaces using image geometry. The basis for solving this problem lies in an understanding of projective geometry, coupled with simple statistical models of the contour generating process. This approach is first applied to the special case of surfaces known to be planar. The distortion of contour shape imposed by projection is treated as a signal to be estimated, and variations of non-projective origin are treated as noise. The resulting method is next extended to the estimation of curved surfaces, and applied successfully to natural images. The statistical estimation stratefl is then experimentally compared to human perception of orientation: human observers' judgements of tilt correspond closely to the estimates produced by the planar strategy. I INTRODUCTION Projective geometry lawfully relates the shape of a surface marking, the orientation of the surface on which the marking lies, and the shape of the marking's projection in the image: given any two, the third can usually be recovered. This paper addresses the problem of recovering the shape and orientation of an irregularly marked surface from the shapes of the projected markings. Of the three constituents of the projective relation, one--projected shape--is given, and another-- surface orientation--is to be recovered. Since two of the constituents must be known to recover the third, the problem has no solution unless something is assumed about the unprojected shapes of the surface markings. For instance, if those shapes were known exactly, recovering surface orientation would usually be a straightforward exercise in projective geometry. More interesting and more general is the case of a surface about which nothing specific is known in advance. To recover surface orientation in this case, some assumption must be made about the geometry of surface markings that is general enough to apply to a broad range of surfaces, powerful enough to to determine a solution for surface orientation, and true enough to determine the right solution. To meet these requirements, the inference of surface shape will be treated as a problem of statistical estimation, combining constraints from projective geometry with simple statistical models of the processes by which surface markings are formed. The distortion imposed on shapes by projection will be treated, quite literally, as a signal, and the shapes themselves, as noise. Both the "signal" and the "noise" contribute to the geometry of the image, and statistical models of the noise permit the projective component to be isolated. II ESTIMATING THE ORIENTATION OF PLANAR SURFACES The estimation problem will first be considered subject to the artificial restriction that the surface is known to be planar. While not realistic, this limited case provides the groundwork from which more general methods will be developed. The shape of a surface marking is the shape of its bounding curve. The shapes of curves are naturally described by tangent direction as a function of position. Since tangent direction in the image is subject to projective distortion, this representation is suitable for estimating projective distortion. The mapping between a tangent direction on a surface and the corresponding direction in the image is readily expressed as a function of surface orientation, by means of simple projective geometry. By examining a region of the image, a distribution of projected tangent directions may be collected. The form of this distribution depends in part on the distribution of directions over the corresponding surface region, but undergoes systematic distortion depending on the orientation of the surface: the projection of an inclined shape is foreshortened, i.e. compressed in the direction of steepest inclination (the tilt direction.) The amount of compression varies with the angle between the image plane and the plane of the shape (the slant angle.) The effect of this distortion on the distribution of tangent directions may be illustrated by the simple case of a circular marking. The distribution of tangent directions on the original circle, measured by arc length, is 1 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. uniform. The orthographic projection of a circle is an ellipse, whose minor axis lies parallel to the tilt direction, and whose eccentricity varies with the slant angle. The distribution of tangent directions on an ellipse is not uniform, but assumes minima and maxima in the directions of the minor and major axes respectively. The degree of nonuniformity increases with the eccentricity of the ellipse. In other words, projection systematically "pushes" the image tangents away from the direction of tilt. The greater the slant angle, the more the tangents are "pushed." This systematic distortion of the tangent distribution is the projective "signal" that encodes surface orientation. Its "phase" (the direction toward which the tangents gravitate) varies with surface tilt, and its "amplitude" (th e amount of distortion,) with surface slant. To estimate the phase and amplitude of the projective signal is to estimate surface orientation. Statistical estimation of the projective component requires a model of the "noise", i.e. of the expected distribution of tangent directions prior to projective distortion. With no prior knowledge of the surface, there is no reason to expect any one tangent direction to be more likely than any other. In other words, it is natural to assume that all tangent directions on the surface are equally likely. (Note that the amount of projective distortTon increases with surface slant, effectively increasing the signal-to-noise ratio. Therefore, as slant increases, the exact form assumed for the noise distribution becomes less critical.) Together with the geometric relation, this simple statistical model defines a probability density function for surface orientation, given a set of tangent directions measured in the image. The surface orientation value at which this function assumes a maximum is the maximum likelihood estimate for surface orientation, given the model. The integral of the function over a range of surface orientations is the probability that the actual orientation lies in that range. Conceptually, the distribution of tangent directions in the image can be projected onto an arbitrarily oriented planar surface. We seek that orientation for which the distribution of projected tangent directions is most nearly isotropic. The estimator was first applied to geographic contours: projections of coastlines drawn from a digitized world map. This choice of data circumvents the problem of contour detection, and allows the actual orientation to be precisely controlled. The overall accord between estimated and actual orientation was excellent, and, equally important, the confidence measures generated by the estimator effectively distinguished the accurate estimates from the inaccurate ones. I- -1 --- Figure 1 - Two photographs of roughly planar surfaces, and the orientation estimates obtained from them. The estimated orientations are indicated by ellipses, representing the projected appearance a circle lying on the surface would have if the maximum likelihood estimate were correct. III EXTENTION TO CURVED SURFACES To apply the methods developed in the planar case to curved surfaces without additional assumptions, it would be necessary to obtain at each point in the image a measure of the distribution of tangent directions. But such a local measure is never available, because the density of the contour data is limited. On the other hand, a distribution can be taken at each point of the data in a surrounding region, as small as possible, but large enough to provide a reasonable sample. This spatially extended distribution may be represented as a three dimensional convolution of the image data with a summation function. The same technique was then applied to natural To understand how such a distribution should images, using zero-crossing contours in the be applied to estimate surface disposition, it is convolution of the image with the Laplacian of a Gaussian [I], [2]. helpful to distinguish the intuitive, perceptual While the veridical notion of surface orientation from the strict orientations were not independently measured, the definition of differential geometry. It may be maximum likelihood estimates are in close accord argued that surface orientation is not a unique with the perceived orientations (see Figure 1). 2 property of the surface, but must be regarded as a function of scale. The scale at which orientation is described corresponds to the spatial extent over which it is measured. Thus, by measuring orientation over a'large extent, the surface is described at a coarse scale. The scale at which the,surface is estimated therefore depends on the spatial extent over which the distribution is computed. Since that extent must be sufficiently large compared to the density of the data, the density effectively limits the spatial resolution of the estimate. On this view, close parallels can be drawn to Horn's [3] method for inferring shape from shading. The local estimator for orientation is a geometric analogue to the photometric reflectivity function. The strategy was implemented, and applied to natural images. Contours were extracted as in the planar case, and the spatially extended distribution approximated by a series of two- dimensional convolutions with a "pillbox" mask. The estimated surfaces were in close accord with those perceived by the human observer (see Figure 2). IV RELATION TO HUMAN PERCEPTION A psychophysical experiment was performed to examine the relation between human observers' judgments of the orientations of curves perceived as planar, and the estimates obtained from the estimation strategy outlined above. A series of "random" curves were generated using a function with pseudorandom parameters. Although such curves have no "real" orientation outside the picture plane, they often appear inclined in space. Observer's judgments of orientation were obtained by matching to a simple probe shape. The judgments of tilt (direction of steepest descent from the viewer) were highly consistent across observers, while the slant judgments (rate of descent) were much more variable. Orientation estimates for the same shapes were computed using the planar estimator, and these estimates proved to be in close accord with those of the human observers, although the shapes had no " real" orientation. While no conclusion may be drawn about the mechanism by which human observers judge orientation from contours, or about the measures they take on the image, this result provides evidence that the human strategy, and the one developed on geometric and statistical grounds, are at the least close computational relatives. 1. 2. 3* REFERENCES Marr, D. C & Poggio, T., "A computational theory of human stereo vision," Proc. Roy. Sot. Lond., Vol. 204, pp.301-328 (1979) Marr, D. C & Hildreth, E., "A Theory of edge detection," MIT AI Memo 518 (1979) Horn, B. K. P, "Understanding image intensities," Artificial Intelligence, Vol. 21, No. 11, pp.201-231 (1977) Figure 2 - A complex image, and the estimate obtained. The ellipses represent the appearance the estimated surface would have were it covered with circles. Note that the estimate correctly distinguishes the highly slanted foreground from the nearly frontal background. The upward pitch of the right foreground is also detected. 3	1980	84
83	INFORMATION NEEDED TO LABEL A SCENE Eugene C. Freuder Dept. of Mathematics and Computer Science University of New Hampshire Durham, NH 03824 ABSTRACT I analyze the information content of scene labels and provide a measure for the complexity of line drawings. The Huffman-Clowes label set is found to contain surprisingly little additional information as compared to more basic label sets. The complexity of a line drawing is measured in terms of the amount of local labeling required to determine global labeling. A bound is obtained on the number of lines which must be labeled before a full labeling of a line drawing is uniquely deter- mined. I present an algorithm which combines local sensory probing with knowledge of labeling constraints to proceed directly to a labeling analysis of a given scene. "r INTRODUCTION Huffman [4] and Clowes [2] developed (inde- pendently) a basic labeling scheme for blocks world picture graphs. Given a basic labeling set: + (convex), - (concave), -+ (occluding, region on arrowhead side), and a standard set of simplifying restrictions on scene content and viewpoint, the physically realizable junction labelings are just those shown in the last column of Fig. 1. Waltz [5] explored the richer label sets obtained by including additional information in the labels (and loosening the scene restrictions). In this paper I explore weaker, cruder label sets. I identify three stages of scene labels, of which the standard set is the third and richest. I then explore the increase in information content embodied in each successive stage. Rather sur- prisingly I find that there is very little real information gain as we move from stage to stage. A first stage scene labeling may well determine a unique second stage labeling. If it does not, it will come quite close to doing so, and I am able to identify precisely the additional information that is necessary and sufficient to complete the second stage labeling. Similar results are obtained for the transition from Stage II to Stage III. These results supply some theoretical insight into the nature and strength of the basic line labels and physical constraints. I go on in Section III to analyze the amount of information required to obtain a Stage I labeling. The information is measured in terms of the number of line labels which must be determined in order for labeling constraints to unambiguously STAGE I STAGE II STAGE III Fork A 2 - - 2 2 A 2 2 2 A + + + Arrow I I T T 2 - +I+ 2 - T +i Fig. 1. Junction labelings ot each stage imply a unique labeling of the entire line draw- ing. In practice the required line labels can be obtained by local sensory probing of the physical scene. I obtain a bound on the number of labels required to imply a full labeling of an arbitrary line drawing. Finally I discuss an algorithm that effectively combines sensory probing for labels with knowledge of labeling constraints. The algorithm proceeds directly to a full labeling, reflecting a presented physical scene, while requiring neither a complete sensory scan for every line label nor a consideration of all possible physical realizations of the line drawing. 18 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. II LABELING STAGES The standard label set is a refinement of a cruder categorization of lines as representing a physical edge of either one or two visible faces. I consider three labeling stages. In Stage I the only labels are the numbers 1 and 2, indicating the number of associated faces. In Stage II, the number 1 is replaced by occlusion labels (+) indicating which side the single face is on. They will be termed Stage II labels. A Stage II label- ing will be one that utilizes -f and 2 labels. In Stage III the number 2 is replaced by + and - as the distinction is made between convex and concave edges. The labels + and - will be termed Stage III labels. - A Stage III labeling is one that utilizes +, - and -f labels. At Stage I there are only 9 distinct junction labels. At Stage II the L labelings are differ- entiated, at Stage III the fork and arrow labels are differentiated. Fig. 1 shows the physically realizable labelings at each stage. T labelings are added at each stage, but notice that fork and arrow labelings do not increase in moving from Stage I to Stage II, and the number of L labelings does not increase in moving from Stage II to Stage III. Thus we really do not know any more about forks and arrows at Stage II than we do at Stage I, nor more about L's at Stage III than at Stage II. Once we have labelled a fork 2,1,1 for example, we really know already that it can be labelled 2,+,-t. My interest in Stage ,J labeling was aroused by the work of Chakravarty [1] who utilized inform- ation about the number of regions associated with lines and junctions, in connection with a more elaborate labeling scheme. A. The Picture Graph -- A blocks world line drawing is, of course, a graph. For the purposes of our analysis we will modify picture graphs by "separating" T junctions, removing T junctions by pulling the shafts away from the crossbars. After labeling a scene separated in this fashion the T junction labelings are easily recovered by rejoining the T junctions to form the original scene. The separation reflects the fact the information does not pass through a T junction , and will permit us to identify independent segments of the scene as connected components of the (separated) picture graph. The segments are independent in the sense that each can be labeled independently, a label in one segment can have no bearing on possible labelings for the other segment. The connected components of a graph are the maximal connected subgraphs, where a graph is connected if there is a chain of edges between any two vertices. B. Stage I to Stage II Theorem 1. Given a picture graph with a Stage I labeling (separated at T junctions). Further separate the graph by separating L junc- tions that have a 2 label on one line, i.e. pulling the two sides of each such L apart to remove the junction. The Stage I labeling uniquely deter- mines a Stage II labeling on all connected compon- ents of the resulting graph except those consist- ing solely of l-labeled lines, none of which is a crossbar of a T. A unique labelinq for the ex- ceptions may be determined by specifying the Stage II label of a single line in each such component. c31 l For proofs of the theorems in this paper see C. Stage II to Stage III -- Theorem 2. Given a picture graph with a Stage II labeling (separated at T junctions). The Stage II labeling uniquely determines a Stage III labeling on all connected components except those consisting solely of 2-labeled lines. A unique labeling for the exceptions may be determined by specifying the Stage III label of a single line in each component, III OBTAINING A STAGE I LABELING Given labels for a sufficient number of lines the physical constraints on the labeling process will imply a unique labeling for the remainder of the scene. A bound on this "sufficient number" will provide a bound on the complexity and poten- tial ambiguity of the picture graph, and on the effort required to label it. The limitations on physically realizable labelings summarized in Fig. 1 easily give rise to a set of "implication rules" for completing a junction labeling given the labelings of one or two of its lines. 2. These rules are listed in Fig. Note first that labels for shafts of arrows and crossbars of T's can be derived immediately (2's and l's respectively), without any previous labels. Labels for two of the lines of a fork or arrow imply the third. (Thus, in effect, a single line label, other than for the shaft, determines an arrow labeling.) => 2 I\ I I => T p => yp yp => ‘;1‘ Fig. 3. Implication rules. 19 I will say that the labelina of a subset of picture graph iines implies the labeling of the entire graph, if repeated application of the implication rules of Fig. 2;'starting with the given subset, leads to a complete, unique labeling of t,he graph. The labeling of a subset of lines is sufficient if the labeling implies the labeling of the graph. A subset of lines is sufficient if any consistent lamoftheubset is suffi- cient. The minimal number of lines in a suffi- cient subset will be called the sufficient number of the picture graph. -- The sufficient number must, of course, be determined relative to a specified label set. We will be dealing with sufficiency for Stage I labeling in this paper. In Section A I obtain an upper bound on the sufficient number of a picture graph. In [3] I discuss means of obtaining sufficient sets of lines (and thus, bounds on sufficient numbers) for individual graphs; I modify one of these methods to provide a test for the sufficiency of a set of lines or line labels, and a labeling algorithm. This algorithm is discussed in Section B. Note that deduction can "propagate": we may deduce a label for a line which in turn permits deduction of the label for an adjoining line, etc. There is room for heuristic tuning in choosing which lines to probe for labels. Fig. 3 demonstrates thi on a simple scene. The fort labeled L line to probe for get away with a single senso the choice need not be entir reasonable to suspect that t be 2-labeled.) m + + I I+ I w.2 a. s labeling algorithm uitous choice of a 2- labeling permits us to ry probe. (Actually ely fortuitous. It is he "bottom lines" will A. A Bound on the Sufficient Number of a Picture ---- ---- Graph Theorem 3. The sufficient number of a picture graph is no more than the number of forks and arrows plus the number of connected components of @ => l@l O=>probed the graph separated at T's and L's. The bound provided by the theorem is tight, in the sense that I can exhibit a picture graph with sufficient number equal to the given bound. A simple triangular figure may consist of all occluding lines, or one of the lines may have a 2 label. Knowing the labeling of two lines is not sufficient to imply the label of the third for all possible labelings. Thus, the sufficient number Fig. 3. Labeling algorithm. REFERENCES [1] Chakravarty, I. A generalized line and junction labeling scheme with applications to scene analysis, IEEE Trans. PAMI- (1979) 202-205. is three, which equals the number of forks and arrows (0) plus the number of connected components in the graph separated at T's and L's (3). (In general the sufficient number of a picture graph will be considerably less than the bound provided by the theorem). PI c31 B. A Labeling Algorithm II41 Algorithm: 1. Obtain any labels that can be deduced by repeated application of the implication rules of Figure 2. (N t o e arrow shafts and T crossbars can be labeled immediately.) II51 2. While unlabeled lines remain: 2.1. Pick an unlabeled line and "probe" the physical scene to determine its label. (This information could be obtained from visual, tactile, or range finding data.) Clowes, M.B. On seeing things, Artificial Intelligence 2 (1971) 79-116. Freuder, E. On the Knowledge Required to Label a Picture Graph. Artificial Intelli- Qence, in press. Huffman, D.A. Impossible objects as nonsense sentences, in: Meltzer, B. and Michie, D. (Eds.), Machine Intelligence 5 (Edinburgh Univ. Prminburgh, 1971) 295-324. Waltz, D. Understand ing line drawings of scenes wit h shadows, in: Winston, P.H. (Ed.), The Psychology (McGraw-Hill, New York, 1975 19-91. of Computer Vision 2.2. Deduce any further labels that can be obtained by repeated applications of the implication rules. 20	1980	85
84	INTERPRETIVE VISION AND RESTRICTION GRAPHS Rodney A. Brooks and Thomas 0. Binford Artificial Intelligence Laboratory, Computer Science Department Stanford University, Stanford, California 94305 ABSTRACT We describe an approach to image interpretation which uses a dynamically determined interaction of prediction and observation. We provide a representational mechansim, built on our geometric modeling scheme which facilitates the computational processes necessary for image interpretation. The mechanism implements generic object classes and specializations of models, enables case analysis in reasoning about incompletely specified situations, and manages multiple hypothesized instantiations of modeled objects in a single image. It is based on restriction nodes and quantified variables. A natural partial order on restriction nodes can be defined by comparing the satisifiability of their constraints. Nodes are arranged in an incomplete restriction graph whose arcs represent relations of nodes under the partial order. Predictions are matched to descriptions by finding maximal isomorphic subgraphs of a prediction graph and an observation graph 183 subject to a naturally associated infimum of restriction nodes being satisifiable. In this manner constraints implied by local two dimensional matches of image features to predicted features are propagated back to the three dimensional model enforcing global consistency. I INTRODUCTION A. Image Interpretation A descriptive process is one which takes an image and produces a description of image features and their relations found within that image. A predictive process is one which uses models of objects expected in an image to predict what features and their relations will be present in the image. We have constructed a working model-based vision system called ACRONYM 183. The interaction between prediction and description is stronger than in previous systems. A central aspect of ACRONYM is that it interprets images at the level of three dimensional models. Here we describe a new layer of representation built on the ACRONYM system. We briefly describe a mechanism for reasoning about incompletely specified geometric situations. All this has been implemented. We describe how matching of two dimensional features can be mapped back to constrain geometric uncertainties in three dimensions in order to obtain a three dimensional understanding of an image. This aspect of the new system is still being implemented (June 1980). 21 B, The ACRONYM System ACRONYM itself covers a wider range of tasks than vision (see 143, [lo]). The reader is referred to 181 for a complete overview of the system and a description of geometric models based on generalized cones as the nodes of a subpart tree. We give here a brief overview of ACRONYM’S vision related modules. A user interacts with the system via a high level modeling language and an interactive editor to provide three dimensional descriptions of objects and object classes which are viewpoint independent, and to partially mode1 general scenes. The result is the object graph. A library provides useful prototypes and a graphics module provides valuable feedback to the user. A rule-based module, the predictor and planner, takes models of objects and scenes and produces the prediction graph which is a prediction of the appearance of objects expected within the scene. It predicts observables in the image over the expected range of variations. It provides a plan, or instructions, for lower level descriptive processes and the matcher to find instances of the objects within the image. The process of prediction and planning is repeated as first coarse interpretations are found, more predictions are carried out, and finer, less ambiguous interpretations are produced. The descriptive aspect of ACRONYM is currently provided by the edge mapper IS] which describes monocular pictures as primitive shape elements (ribbons) and their spatial relationships. It is goal-directed and is thus programmed by the predictor and planner. The observation graph is the result. As with the predictor and planner, the edge mapper may be invoked many times during the course of an interpretation as finer levels of description become desirable. ACRONYM will incorporate stereo and advanced edge-based description modules (Baker [23 and Arnold and Binford [II). This will provide three dimensional cues directly. The matcher interfaces description and prediction. It finds maxima1 subgraphs of the observation graph isomorphic with subgraphs of the prediction graph, which also meet global consistency requirements. In the new implementation the matching process is mapped back to three dimensional models. Such higher level understanding ensures global consistency and enables deductions about three dimensional structures from a single monocular image. The matcher re-invokes the predictor and planner and the edge mapper to extend the two graphs which it is matching in the context of successfully completed submatches. This provides direction to both prediction and From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. description and reduces each must consider. dramatically the number of possibilities Lowe 191 has implemented a system which determines parameters of models including articulations from correspondences of a match. This module provides predictions from a tentative interpretation to guide detailed verification. We have concentrated on two classes of images in our development work; aerial images of airport scenes, and scenes of industrial workstations. Together they provide a wide range of challenges and force us to look for general methods and solutions to problems, since they are sufficiently dis-similar that special purpose solutions will fail on one of the two. II REPRESENTATION A. Requirements We have chosen to describe the world to ACRONYM in terms of three dimensional models of objects, and their expected three dimensional spatial relationships ([Sl contains details). The representation is a coarse to fine description of objects as subpart trees of generalized cones 133. In this paper we are concerned with ways to represent variations within models, and how to keep track of multiple inconsistent instances which may arise during image interpretation. Thus what follows does not rely on generalized cones as the representational primitive. In structured situations the exact dimensions of objects may be known in advance, but their orientation may be uncertain. For instance it may be known that a bin is full of a particular type of part, but the parts may have been dropped in with arbitrary orientations. In an industrial automation domain such descriptions may already be available from a Computer Aided Design data-base. In less structured environments, not even the dimensions of objects will be known exactly. For instance for some aerial image interpretation tasks it is desirable to represent both the class of wide bodied passenger jet aircraft, and particular types of aircraft such as a Boeing 747, and even more particular models such as a 747-B. Thus it is necessary to represent constrained variations in shape, size and even structure (e.g. different aircraft have different engine configurations), and constrained variations in spatial relations between objects. Consider also that an F-l 11 aircraft can have variable wing geometry. A manipulator arm has even more complex variations in spatial relations between its subparts. The appearance of an object may change qualitatively, rather than merely quantitatively, over the allowed variations in its size, shape, structure or orientation relative to the camera. Thus it will often be necessary to carry out some case analysis in prediction of appearance, and put further constraints on models, and to keep such mutually confliciting hypotheses in the prediction graph, until such time as they can be confirmed or denied by descriptive processes. The prediction graph represents case as combinations of components instead of explicit enmueration of all cases. As interpretation of an image proceeds, constraints on the exact values of variations within a model will be derived from the matches made in the image. However there may be multiple instances of a modeled object within the image. Parts on a conveyor belt will have different orientations. Aircraft at a passenger terminal will have different lengths and wing spans. Thus multiple instances of objects must be representable in the interpretation graph. B. Representing Variations In the following discussion we will consider the problem of modeling both the generic class of wide-bodied passenger jet aircraft, and specific wide-bodied passenger jet aircraft, such as the Boeing 747, Lockheed L-101 1, McDonnell-Douglas DC-10 and the Airbus Consortium A-300. We will then discuss a wider situation where such aircraft are on runways and taxiways, and there are undetermined variables in the camera model. We need to represent both variations in size (e.g. different aircraft subclasses will have different fuselage lengths), and variations in structure (e.g. different aircraft subclasses will have different engine configurations). In both cases we want to represent the range of allowable variations. We consider the broader problem of quantification of sets. Furthermore, there will sometimes be interdependencies between these variations (e.g. a scaling between fuselage length and wing span). Node: FUSELAGE-CONE NAME : SIMPLE-CONE SPINE: 20005 SWEEPING-RULE: CONSTANT-SWEEPING-RULE CROSS-SECTION: 20004 Node: Z0005 NAME: TYPE: LENGTH: SPINE STRAIGHT FUSELAGE-LENGTH Node: CONSTANT-SWEEPING-RULE NAME : SWEEPING-RULE TYPE: CONSTANT Node: 20004 NAME : TYPE: RADIUS: CROSS-SECTION CIRCLE FUSELAGE-RADIUS Generalized cone representation Figure 1. of fuselage. The primitive representational mechanism used in ACRONYM is that of units and slots. Objects are represented by units, as are generalized cones, cross-sections, sweeping-rules, spines, rotations and translations to name the more important ones. Figure 1 shows four units with their slots and fillers from a particular ACRONYM model. They describe the generalized cone representing the fuselage of the generic wide-bodied passenger jet aircraft. Note that units are referred to as “Nodes” because they are nodes of the Object graph of figure I. The NAME slot is a distinguished slot which all units possess. It describes the entity represented by the unit and corresponds to the SELF slot of KRL units E51. Units identified by “Z” followed by a four digit number are those which were given no explicit identifier by the user who modeled the object. The modeling language parser has generated unique identifiers for 22 The value of a slot is given by its filler. Slot fillers may be explicit, such as “2” or “STRAIGHT”. They can also be symbolic constants in the same sense as constants are used in programming languages such as PASCAL. Such fillers are fine for representing specific completely determined objects and situations. Slots may be filled by a quantifier, or any evaluable expression involving quantifiers. A quantifier is an identifier with constraint “FUSELAGE-LENGTH” and system (quantification). “FUSELAGE-RADIUS” are examples of such quantifiers in figure 2. The following constraints might be imposed upon FUSELAGE-LENGTH and FUSELAGE-RADIUS when modeling the class of wide-bodied passenger jet aircraft: Node: JET-AIRCRAFT NAME 8 OBJECT SUBPARTS: (STARBOARD-W I NG PORT-W I NG FUSELAGE 1 QUANTIFIERS: (F-ENG-QUANT ENGINE-LENGTH ENGINE-RADIUS WING-ATTACHMENT ENG-OUT ONE-WING-SPAN WING-SWEEP-BACK WING-LENGTH WING-RATIO WING-WIDTH WING-THICK) Node : STARBOARD-W I NG (s 40.0 FUSELAGE-LENGTH) (s FUSELAGE-LENGTH 70.0) fs 2.5 FUSELAGE-RADIUS) (s FUSELAGE-RADIUS 3.5) (s 15.6 (QUOTIENT FUSELAGE-LENGTH FUSELAGE-RADIUS)) NAME : OBJECT SUBPARTS: ( (SP-DES F-ENG-QUANT e STARBOARD-ENGINE)) CONE-DESCRIPTOR: STARBOARD-WING-CONE These constrain the range of allowable length and radius, and express a lower bound on the ratio of length to radius. Quantifiers express allowable variations in dimensions of objects and in the structure of objects. Figure 2 gives the complete subpart tree for a model of generic wide-bodied passenger jet aircraft. For brevity, not all the slots of the OBJECT units are shown here. The QUANTIFIERS slot is explained later. The SUBPARTS slot of an OBJECT unit is filled with a list of subparts giving the next level of description of the object. Entries in the list can be simple pointers to other OBJECT units (e.g. JET-AIRCRAFT has three substructures: STARBOARD-WING, PORT-WING and FUSELAGE). They can also be more complex such as the single entry for the subparts of STARBOARD-WING, which speiifies a quantification of subparts called STARBOARD-ENGINE. In this case the quantification is the quantifier F-ENG-QUANT. Note that PORT-WING has a quantification of PORT-ENGINES as subparts, which is represented by the same quantifier F-ENG-QUANT. This explicitly represents one aspect of the symmetry of the aircraft: it has the same number of engines attached to each wing. Constraints on this quantifier and on R-ENG-QUANT, the number of rear engines might be: Node: STARBOARD-ENGINE NAME I OBJECT CONE-DESCRIPTOR8 PORT-ENGINE-CONE Node : PORT-W I NG NAME r OBJECT SUBPARTS: ( (SP-DES F-ENG-QUANT . PORT-ENGINE) 1 CONE-DESCRIPTOR: PORT-WING-CONE Node: PORT-ENGINE NAME : OBJECT CONE-DESCRIPTOR: PORT-ENGINE-CONE Node: FUSELAGE NAME : OBJECT SUBPARTS: (RUDDER STARBOARD-STABILIZER PORT-STABILIZER) QUANTIFIERS: (STAB-ATTACH STAB-WIOTH STAB-THICK STAB-SPAN STAB-SWEEP-BACK STAB-RAT 101 CONE-DESCRIPTOR: FUSELAGE-CONE (5 1 F-ENG-QUANT) ts 2 F-ENG-QUANT) (s 0 R-ENG-QUANT) (I 1 R-ENG-QUANTI (b 3 (PLUS F-ENG-QUANT R-ENG-QUANTI) These say that there must be either one or two engines on each wing, -zero or one at the rear of the aircraft, and if there are two on each wing then there are zero at the rear. Symmetry of size (such as length of the wings) can likewise be represented by using the same quantifier as a place holder in the appropriate pair of slots. Node : RUDDER NAME : SUBPARTS : OBJECT ( (SP-DES R-ENG-QUANT . REAR-ENGINE)) CONE-DESCRIPTORI RUDDER-CONE Node: REAR-ENG I NE NAME : OBJECT CONE-DESCRIPTORr REAR-ENGINE-CONE Node: STARBOARD-STABILIZER NAME : OBJECT Our compjete model for a generic wide-bodied passenger jet aircraft has 28 quantifiers describing allowable variations in size and structure. CONE-DESCRIPTOR: STARBOARD-STABILIZER-CONE Node: PORT-STABILIZER NAME t OBJECT CONE-DESCRIPTORr PORT-STABILIZER-CONE Subpart tree of generic Figure 3. passenger jet. 23 C. Representing Classes It should be clear that to model a subclass of wide-bodied passenger jet aircraft we need only provide a different (more restrictive) set of constraints for the quantifiers used in ehe general model. To model a specific type of aircraft we could force the constraints to be completely specific (e.g. (= FUSELAGE-LENGTH 52.8)) ). Thus we will not need to distinguish between specialization of the general model to a subclass, or an individual. Given that subclasses use different sets of constraints, the problem arises of how to represent multiple subclasses simultaneously. We introduce a new type of node to the representation: a restriction node. These are the embodiment of specialization. A restriction node has a set of constraints associated with it. If a set of values can be found for all the quantifiers mentioned in the constraints such that all the constraints are simultaneously satisifed, then we say the restriction node is satisifiable. A partial order can be defined on restriction nodes by saying that one restriction node is more restrictive than another if its set of sets of satisfying values is a subsee of that of rhe second node. Different views of the generic model. Figure 4.- For the example of the generic wide-bodied passenger jet aircraft the constraints are associated with some restriction node, GENERIC- JET-AIRCRAFT say. To represent the class of 747s a more reskictive node can be included; e.g.: Node: BOEING-747 NAME: RESTRICTION SUPREMA: (GENERIC-JET-AIRCRAFT) TYPE: MODEL-SPECIALIZATION CONSTRAINTSI <list of constraints> It is constructed by taking the constraints associated with the GENERIC-JET-AIRCRAFT restriction node, and merging in additional constraints to specialize to a BOEING-747. A model is always accessed in Ihe context of a restriction node. Thus when reasoning about the generic class of wide-bodied aircraft, the predictor and planner will access the JET-AIRCRAFT model and base its reasoning on the constraints given by the GENERIC-JET-AIRCRAFT restriction node. When reasoning about Boeing 747s it will base its reasoning about the JET-AIRCRAFT model on the constraints given by the BOEING-747 restriction node. Figure 4 conveys the flavor of viewing the JET-AIRCRAFT through different restriction nodes to see different models. (And in fact, the drawings of the two types of aircraft were produced by ACRONYM from the indicated restriction nodes.) In modeling subclasses, restriction nodes typically form a tree rather than a graph. D. Representing Spatial Relations Affixments are coordinate transforms between local coordinate systems of objects. They are comprised of a rotation and a translation. Sometimes affixments vary over an object class. For instance the in generic wide-bodied passenger jet aircraft model the position along the fuselage at which the wings will be attached will vary with particular types of aircraft. Articulated objects are modeled by variable affixments. Variable affixments can also be useful for modeling spatial relationships between two objects - for instance an aircraft is on a runway. We represent a vector as a triple (a,b,c) where a, b and c are scalars. We represent a rotation as a pair <v,m> where v is a unit vector, and m a scalar magnitude. An affixment will be written as a pair (r,t) where r is a rotation and t a translation vector. We will use some special vectors also: 4 Q and f. We use for the composieion of rotations, and QP for the application of a rotation to a vector. In ACRONYM we use the quantifier mechanism to represent affixments which describe a class of coordinate transforms. This gives symbolic representations for rotations and translations. Consider the problem of representing ehe fact that an aircraft is somewhere on a runway. Suppose the runway has its x axis along its length, the y axis perpendicular at one end, and the positive z direction vertically upward. Suppose that the coordinate system for the aircraft has its x axis running along the spine of the fuselage and has its t axis skyward for the standard orientation of an aircraft. Then to represent the aircraft on the runway we could affix it with the affixment: (<f, ORI>, (JET-RUNWAY-X, JET-RUNWAY-Y, 011 where ORI, JET-RUNWAY-X and JET-RUNWAY-Y are quantifiers with the following constraints: (I 0 JET-RUNWAY-X) (ls JET-RUNWAY-X RUNWAY-LENGTH) (5 0 JET-RUNWAY-Y) (< JET-RUNWAY-Y RUNWAY-WIDTH) Notice that OR1 is unconstrained. The aircraft is constrained to be on the runway, in the normal orientation for an aircraft (e.g. not upside down), but it does not constrain the 24 direction in which the aircraft is pointed. If we wished to constrain the aircraft to approximately line up in the direction of the runway we could include a constraint on the quantifier ORI, allowing for some small uncertainty. In general, constraints on a single quantifier may derive from different coordinate systems. III PREDICTION AND MATCHING The left most rotation corresponds to a rotation in the image plane and can be ignored when predicting image shape - i.e. shape is invariant with respect to rotation in the image plane. The right most rotation expression is applied directly to the cylindrical tool. But it is a rotation about the x axis, which is the linear axis of a cylinder in our representation 181 and the appearance of a cylinder is invariant with respect to a rotation about its linear axis. Thus for shape prediction we need only consider: <y^,TILT> A. Using Constraints Prediction consists of examining the model in order to find features which are invariant over the range of variations in the model Es], or over a sufficient range to allow a small number of cases. A mixture of reasoning directly about the model, and reasoning about the constraints is needed to find invariants. If ‘TILT is sufficiently constrained (as in this example) it may be possible to predict the shape directly. The prediction takes the form of expected image features, their relations, and what constraints local matches to such features produce on the three dimensional model. See section III-C for an example. But note here that the prediction is a conjunction of expected features. B. Adding Constraints Electric Screwdriver and Holder Figure 5. If TILT in the above example is not sufficentiy constrained there may be more than one qualitatively different shape possible for the cylinder (e.g. an end view of a cylinder is quite different from a side view). If so it is necessary to make a disjunction of predictions. Note however, that ail views need not be explicitly expanded - they can still share much structure. Consider the electric screwdriver holder and electric screwdriver in figure 5. This is a display of an ACRONYM model of a tool for the Stanford hand-eye table. The position and orientation (about the vertical axis) are not known. Neither are the exact camera pan and tilt known. Under these conditions the expression for the orientation of the screwdriver tool relative to the camera, as obtained directly from the model, is: cx^, TI LT>YK<Q, (- PAN) .<Z?, 3n/Z>r~<y^, 3n/2>I a<f,ORI >* I <(j, 3n/2>r#<y^, n/2><& n/2>1 I I I Each prediction is associated with a new, more restrictive restriction node. It is obtained by adding a new constraint which restricts the model sufficiently to necessitate only a single prediction. Figure 5 gives an indication of the structure of a local prediction, with two different cases considered. Not indicated in that diagram are arcs between the feature predictions which specify relations which should hold between instances of those features in the image. R..LrlCLlm u&m corm PredlCLlon where I is the identity rotation, PAN and TILT are quantifiers associated with the camera orientation, and OR1 is the unconstrained orientation of the screwdriver holder about the vertical axis. We have implemented a set of rules in ACRONYM which simplify such expressions to a canonical form, using identities for re-ordering products of rotations. The details can be found in 173. The canonical form aids in the detection of invariants. E.g. the above expression is transformed to the equivalent expression: <f,3n/Z>r<~,TILT><~, (+ PAN (- ORi))> W/l/ L oca I Graph Strut ture [or feeture Predlctlms cone appearance predfct ton. Figure 5 Consider the problem of predicting the appearance of the cylinder in the image. We outline below the chain of reasoning intended for ACRONYM. (The previous predictor and planner rule set carried out a slightly less general but still powerful line of reasoning. For instance from the knowledge that the aircraft is on the runway ACRONYM deduces its image in an aerial photograph is determined up to a rotation about the vertical plus a translation. The rules necessary for a class of computations including the simple example below will be implemented over summer 1980.) C. Geclerating Restrictions During Matching In this section we give an example of an image prediction which generates restriction nodes during matching. We predict the appearance of a ribbon generated by the fuselage of figure I in an aerial image. A prediction says what a match of a local feature must imply about both the observed feature, and the aspect of the object whose image generated that feature. Checking this implication (deciding whether the new restriction node is satisfiable) provides an acceptance test 25 for a hypothesized match. For a projective imaging system the observed distance m between two points distance 1 apart on the ground is given (approximately) by: Cl m E -- h where c is a constant dependent on the focal distance of the camera and h is the height of the camera above ground. Thus if the camera is at height HEIGHT (a quantifier) and Ml and M2 are the measurements obtained for length and width from a hypothesized match of a ribbon in the observation graph, then the following constraints must hold if the match is correct: 1. (- Ml (QUOTIENT (TIMES CAM-CONST FUSELAGE-LENGTH) HEIGHT) 1 2. (= H2 KtUOTIENT (TIMES 2 CAM-CONST FUSELAGE-RADIUS) HEIGHT) 1 In general Ml and M2 will not be given exactly by the observation graph, rather an interval estimate will be supplied. Thus they can be represented by quantifiers with constraints on their upper and lower bounds. If CAM-CONST is known in advance its numeric value can be substituted into the constraints generated by the match. If it too is a quantifier, then it is just one more constrained unknown in the system. At the time of hypothesizing a match, a new restriction node is generated by addingconstraints 1 and 2 to the const aints of the restriction node associated with the prediction (see figure 5). If the new node can be shown to be unsatisfiable then the match is rejected. The following is not meant to indicate a proposed reasoning chain for ACRONYM. Rather it is illustrative of how constraints can imply that a hypothesized match is incorrect. Suppose for some hypothesized match, where CAM-CONST is known to be 100.0, the observed Ml lies between 4.0 and 5.0. Then given the constraints on the fuselage size of section II-B, the height of the camera must be between 800.0 and 1750.0 due to constraint 1 above. If this is inconsistent with a priori constraints on HEIGHT the match can be rejected. In fact a priori constraints on HEIGHT may also put further restrictions on the possible range for FUSELAGE-LECNTH. Similarly measurement M2 and constrairit 2 above will lead to restrictions on HEIGHT. If these restrictions are inconsistent with the 800.0 to 1750.0 bounds already obtained the match should be rejected. D. From Local to Global After each phase of local matching, the matcher combines local matches into more global interpretations. This involves finding consistent subgraphs of matches. Previously consistency has only concerned the existence of arcs describing relations between matched ribbons. With the introduction of constraints on quantifiers during the ribbon matching process, these too must be checked for consistency. Constraints on a quantifier at different HYPOTHESIS-MATCH restriction nodes may actually refer to different quantities in the scene. For instance each potential match for an aircraft may have constraints on FUSELAGE-LENGTH and on HEIGHT. When combining the matches for aircraft to produce an interpretation of the image, there is no reason to require that the constraints on FUSELAGE-LENGTH at these different nodes be mutually consistent. Different instances of wide-bodied passenger jet aircraft will be different lengths. However ail the constraints on HEIGHT should be mutually consistent, as there is only one HEIGHT of the camera. Sometimes when constraints on quantifiers actually correspond to different quantities in the world, it may be that these quantities should have the same value. For instance the ENGINE-LENGTH for the port and starboard engines correspond to physical measurements of different objects in the world. However since aircraft are symmetric the constraints given by the matches on possible values of ENGINE-LENGTH for each engine should be consistent. Thus when clumping the local matches for an aircraft the ENGINE-LENGTH constraints from each submatch should be checked for consistency. If they are not consistent the particular set of local matches should be rejected as inconsistent. A slot is provided in object units to represent which quantifiers matched at a lower level should be held consistent for interpretation of an object. This is the QUANTIFIERS slot as shown in figure 2. As the matcher is combining local matches it looks up the subpart tree. Any quantifier mentioned in a QUANTIFIERS slot of any ancestor of the object has its constraints copied into the restriction node for the new more global node. As each constraint is introduced it is checked for consistency. This process is not quite straightforward. Sometimes a constraint on a quantifier involves another quantifier which is not being brought into the new match. Such is the case of FUSELAGE-LENGTH and HEIGHT in the example of the previous section when a global interpretation is being made involving many aircraft. Each aircraft provides a constraint on HEIGHT, but each is in terms of the instance of FUSELAGE-LENGTH of the individual aircraft. One solution is to generate a new unique identifier for the quantifier which is not to be constrained by new constraints imposed. Its role is to ensure the continued satisifiabiiity of the local match in light of new global constraints on other quantifiers involved in that match. Other solutions exist, which may result in constraint inconsistencies being missed in return for much simplified constraint analysis. IV REMARKS We have described a single part of ACRONYM and have ignored many important issues involved in the construction of a vision system based on the representations given. In particular we have not discussed the analytic power necessary to decide whether constraint sets are satisfiable. We believe that quite weak analytic methods can lead to powerful interpretive capabilities even though they fail to detect large classes of inconsistencies. Nor have we described in detail the methods to carry out the necessary geometric reasoning. These have been discussed in E7J which includes explicit rules for symbolic geometric reasoning in states of uncertain knowledge. We have provided a representational scheme which facilitates the computaeionai processes necessary for interpretation. The scheme uses restriction graphs to provide specializations of models, to enable case analysis in reasoning about incompletely specified situations, and to manage multiple hypothesized instantiations of modeled objects in a single image. 26 ACKNOWLEDGEMENTS This research was sponsored by ARPA contract MDA-903-76-C-0206 and NSF contract DAR-78 15914. Support was also provided by the ALCOA foundation. REFERENCES 111 Arnold, R. David and Thomas 0. Binford, “Geometric Constraints in Stereo Vision,” Proc. SPIE Meeting, San Diego, July 1980. C21 Baker, H. Harlyn, “Edge Based Stereo Correlation,” Proc. ARPA Image Understanding Workshop, Baltimore, Apr. 1980, 168- 175. 131 Binford, Thomas O., “Visual Invited paper at IEEE Systems .Con.erence, Miami, Dec. 1971. Perception by Computer,” Science and Cybernetics 141 Einford, Thomas O., Proc. NSF Grantees Conference, Cornell Univ., Sep. 1979. 151 Bobrow, Daniel G. and Terry Winograd, “An Overview of KRL, a Knowledge Representation Language,” Cognitive Science 1, 1977, 3-46. 161 Brooks, Rodney A., “Goal-Directed Edge Linking and Ribbon Finding,” PYOC. ARPA tmage Understanding Worhshop, Palo Alto, Apr. 1979, 72-78. 171 Brooks, Rodney A. and Thomas 0. Binford, “Representing and Reasoning About Partiatiy Specified Scenes,” Proc. ARPA Image Understanding WorRshop, Baltimore, Apr. 1980,95-103. I81 Brooks, Rodney A., Russell Greiner and Thomas 0. Binford, “The ACRONYM Model-Based VisionSystem,” Proc. of IJCAI-79, Tokyo, Aug. 1979, 105-l 13. 191 Lowe, David, “Solving for the Paramters of Object Models from Image Descriptions,” Proc. ARPA Image Understanding Worksliop, Baltimore, Apr. 1980, 121-127. Cl01 Soroka, Barry I., “Debugging Manipulator Programs with a Simulator,” to be presented at CAD/CAM8, Anaheim, Nov. 1980. 27	1980	86
85	Sticks, Plates, and Blobs: A Three-Dimensional Object Representation for Scene Analysis Linda G. Shapiro John D. Moriarty Prasanna G. Mulgaonkar Robert M. Haralick Virginia Polytechnic Institute and State University Department of Computer Science ABSTRACT In this paper , we describe a relational model- ing technique which categorizes three-dimensional objects at a gross level. These models may then be used to classify and recognize two dimensional views of the object, in a scene analysis system. I. Introduction The recognition of three-dimensional objects from two-dimensional views is an important and still largely unsolved problem in scene analysis. This problem would be difficult even if the two- dimensional data were perfect, but the data can be noisy, distorted, occluded, shadowed and poorly segmented, making recognition much harder. Since the data is so rough it seems reasonable that very rough models of three-dimensional objects should be used in the process of trying to classify such data. In this paper we describe a relational model and discuss its use in a scene analysis sys- tem. There have been many approaches to modeling three-dimensional objects. For a comprehensive collection see the proceedings of the Workshop on Representation of Three-Dimensional Objects [13]. Also see Voelcker and Requicha [ll] and Brown [4] for mechanical design; York et.al. [14] for curved surface modeling using the Coons surface patch 151; Horn [6] and Waltz [12] for study of light and shadows; Badler et.al. [2] for study of human body modeling; and Pgin and Binford [1] and Neva- tia and Binford [7] for the generalized cylinder approach. The models we suggest are related to the generalized cylinder models, but are rougher descriptions that specify less detail about three-dimensional shape than do generalized cylin- ders. II Sticks, iptions Plates, and Blobs in Relational Dt?S- A relational description of an object consists of a set of parts of the object, the attributes of the parts, and a set of relations that describe This research was supported by the National Sci- ence Foundation under grants MCS-7923827 and MCS-7919741. how the parts fit together. Our models have three kinds of three-dimensional Parts: and blobs. Sticks are long, sticks, Plates. thin&-g onlv one significant dimension. Plates are flat- ish-wide parts consisting of two nearly flat sur- faces connected by a thin edge between them. Plates have two significant dimensions. Blobs are neither thin nor flat; they have three significant dimensions. All three kinds of parts are "near convex"; that is a stick cannot bend very much, the surfaces of a plate cannot fold too much, and a blob can be bumpy, but cannot have large concav- ities. Figure 1 shows several examples of sticks, plates, and blobs. sticks Figure 1 illust rates several examples sticks , plates and blobs. each of In describing an object, we must list the parts, their types (stick, plate, or blob), and their relative sizes; and we must specify how the parts fit together. For any two primitive parts that connect, we specify the type of connection and up to three angle constraints. The type of connection can be end-end, end-interior, end-cen- ter, end-edge, interior-center, or center-center where "end" refers to an end of a stick, "inte- rior" refers to the interior of a stick or surface of a plate or blob, "edge" refers to the edge of a plate, and "center" refers to the center of mass of any part. 28 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. For each type of pairwise connection, there are one, two, or three angles that, when specified as single values, completely describe the connec- tion. For example, for a stick and a plate in the end-edge type connection, two angles are required: the angle between the stick and its projection on the plane of the plate and the angle between that projection and the line from the connection point to the center of mass of the plate. Requiring exact angles is not in the spirit of our rough models. Instead we will specify permis- sible ranges for each required angle. In our relational model, binary connections are described in the CONNECTS/SUPPORTS relation which contains lo-tuples of the form (Partl, Part2, SUPPORTS, HOW, VLl, VHl, VL2, VH2, VL3, VH3) where Part1 connects to Part2, SUPPORTS is true if Part1 sup- ports Part2, HOW gives the connection type, VLi gives the low-value in the permissible range of angle i and VHi gives the high value in the per- missible range of angle i, i = 1, 2, 3. The CONNECTS/SUPPORTS relation is not suffi- cient to describe a three-dimensional object. One shortcoming is its failure to place any global constraints on the resulting object. We can make the model more powerful merely by considering tri- ples of parts (sl,s2,s3) where sl and s3 both touch s2 and describing the spatial relationship between sl and s3 with respect to s2, Such a des- ,ription appears in the TRIPLE CONSTRAINT relation and has two components: 1) a boolean which is true if sl and s3 meet s2 on the same end (or sur- face) and 2) a contraint on the angle subtended by the center of mass of sl and s3 at the center of mass of s2. The angle constraint is also in the form of a range. Our current relational description for an object consists of ten relations. The A/V rela- tion or attribute-value table contains global pro- perties of the object. Our A/V relations cur- rently contain the following attributes: 1) number of base supports, 2) type of topmost part, 3) num- ber of sticks, 4) number of plates, 5) number of blobs, 6) number of upright parts, 7) number of horizontal parts, 8) number of slanted parts. The A/V relation is a simple numeric vector, including none of the structural information in the other relations. It will be used as a screening rela- tion in matching; if two objects have very differ- ent A/V relations, there is no point in comparing the structure-describing relations. We are also using the A/V relations as feature vectors to input to a clustering algorithm. The resulting clusters represent groups of objects which are similar. Matching can then be performed on clus- ter centroids instead of on the entire database of models, Other relations include SIMPLE PARTS, PARALLEL PAIRS, PERPENDICULAR PAIRS, LENGTH CONST- RAINT, BINARY ANGLE CONSTRAINT, AREA CONSTRAINT, VOLUME CONSTRAINT, TRIPLE CONSTRAINT and CON- NECTS/SUPPORTS. III. Matching Relational matching of two-dimensional objects to two-dimensional models is a well-defined opera- tion. See Barrow, Ambler, and Burstall [3] for a discussion of exact relational matching, Shapiro [8] for relational shape matching, and Shapiro and Haralick [lo] for inexact matching. Our problem in scene analysis is to match two-dimensional per- spective projections of objects (as found in an image) to the three-dimensional models stored in the database. Our approach to this problem is to analyze a single two-dimensional view of an object, produce a two-dimensional structural shape description, use the two-dimensional description to infer as much as possible about the correspond- ing three-dimensional description, and then use inexact matching techniques in trying to match incomplete and possibly erroneous three-dimen- sional object descriptions to our stored three-di- mensional relational models. We decompose a two-dimensional view into sim- ple parts by a graph-theoretic clustering scheme as described in [9]. To match a two-dimensional object description to a three-dimensional model is to find a mapping from the tm-dimensional simple parts of the object to the sticks, plates and blobs of the model so that the relationships among the two-dimensional parts are not inconsistent with the relationships among the three-dimensional parts. For example, a binary CONNECTS relation can be constructed for the two-dimensional parts. For a pair (pl,p2) of three-dimensional model parts where (pl,p2,,,,,,,,,) is an element of the CONNECTS/SUPPORTS relation and a mapping h from three-dimensional model parts to two-dimen- sional object parts, if (h(pl),h(p2)) is not an element of the two-dimensional CONNECTS relation, then an error has occured. If a mapping accumu- lates too many errors from various n-tuples of various relations not being satisfied, that map- ping cannot be considered a match. As an example, suppose the three-dimensional model of a simple chair contains two plates (the back B and seat S) and four sticks (legs Ll, L2, L3, L4). The relation obtained from just the first two columns of the CCNNECTS supports rela- tion is {(S,B), (B,S) , U-J,S), (S,Ll) , (L2,S) r (S,L2), (L3,S) I (S,L3) (L4,S), (S,L4)]. Now con- sider the two-dimensional decomposition of Figure 2. We can construct the hypothetical connection relation C = {(sl,s2), (s2,sl), (s3,s2), (s2,s3), (s3,sl), (sl,s3) I (s4,s2), (s2,s4), (s4,sl), (sl,s4), (s5,s2), (s2,s5)]. Then the mapping f defined by {(S,s2), (Btsl) , (Ll,s3), (L2,s4), (L3,s5), (L4,s4)) accumulates no error while the mapping g defined by {(S,sl),(B,s2), (Ll,s3), (L2,s4), (L3,s5), (L4,s4)] accumulates error since (L3,S) is in the model, but (f(L3),f(S)) = (s5,sl) is not in C. Not all of the three-dimensional relations can be directly constructed from two-dimensional data. (If they could, the entire scene analysis problem would be much easier.) For example, only an esti- mate of whether one part supports another can be computed. Relations like PARRALLEL PAIRS and LENGTH CONSTRAINT can also be estimated. Rela- tions involving angles are probably the most dif- 29 m Figure 2 illustrates the decomposition of a two- dimensional chair by graph-theoretic clustering. ficult, since a perspective projection will change the angles between parts. Such .information should be left out of initial matching attempts and used later to try to validate a given match or to "hoose between several possible matches. The pre- cise definition of an inexact match from a two-di- mensional description to a three-dimensional des- cription is the subject of our current research. IV. Sumnary of Current and Future Research We have described a relational model for three-dimensional objects, in which the parts of an object are very grossly described as sticks, plates, or blobs. We are building a database of three-dimensional object models. The objects in the database are being clusterd into groups, using graph-theoretic clustering algorithm. Instead of comparing a two-dimensional view to every object in the database, it will be compared ini- tially only to a centroid objects in each group. Only in those groups where the unknown object is most highly related to the centroid will any full relational matching take place. Relational matching will be a form of the inexact matching we described in [lo]. The gen- eral method will be to obtain estimates of the three-dimensional relations from the two-dimen- sional shape and match these estimates against the three-dimensional models. Deriving the algorithms and heuristics for the matching is one of our most challenging tasks. References 2. Badler, NaI.I J. O'Rourke, and H. Toltzis, "A Spherical Representation of A Human Body For Visualizing Movement", Proceedings of - the IEEE, Oct. 79. 3. Barr==, A.P. Ambler, and R.M. Burstall, "Some Techniques for Recognizing Structure in Pictures", nition, S, Frontiers of Pattern Recog- Watanabe (ed),Academic Press, New York, 1972, pp. l-29. 4. Brown, C.M., A.A.G. Requicha and H.B. Voelcker, "Geometric Modeling Systems for Mechanical Design and Manufacturing", Proc. ACM 1978, Washington D.C., Dee 4:6, m -- l7l” . 5. Coons, S.A., "Surfaces for Computer-Aided Design of Space Forms," M.I.T. Project MAC, MAC-TR-41, June 1967 (AD 663504). 6. Horn, B., "Obtaining Shape From Shading Infor- mation", in The Psychology of Computer 7 Vision, (P.H. Winston, ed.), McGraw-Hill, New York, 1975, pp. 115-155. 7. Nevatia, R.K. and T.0, Binford, "Structured Descriptions of Complex Objects", Proc. Third International Joint Conference on Artificial Intelligence, Stanford, 1973.- 8. Shapiro, L.G., "A Structural Model of Shape", IEEE Transactions on Pattern Analysis and Machine Intelligence, -1-2, No. 2, -980. 9. Shapiro, L.G. and R.M. Haralick, "Decomposi- tion of Two-Dimensional Shapes by Graph- Theoretic Clustering," IEEE Transactions on Pattern Analysis and Machine Intelli- genZ&XX. PAMI- pp- lo-?Z&-ZZi.m 10. Shapiro, L.G. and R.M. Haralick, "Structural Descriptions and Inexact Matching", Tech- nical Report CS79011-R, Department of Com- puter Science, Virginia Polytechnic Insti- tute and State University, Blacksburg, VA 24061, Nov. 1979. 11. Voelcker, H.B. and A.A.G. Requicha, "Geometric Modeling of Mechanical Parts and Pro- cesses", Computer, Vol. 10, No. 12, Dec. 1977, pp.48-57. 12. Waltz, D., "Understandinq Line Drawinqs of Scenes with Shadows", in The Psychology of Computer Vision, (P.H. Winston, ed.x McGraw-Hi-w York, 1975. 13. Workshop on Representation of Three-Dimen- sional Objects, (R. W-Y, Director), University of Pennsylvania, May l-2, 1979. 14. York, B., A. Hanson, and E.M. Riseman, "Repre- sentation of Three-Dimensional Objects with Coons Patches and Cubic B-Splines", Department of Computer and Information Science, University of Massachusetts, Amherst, MA 01003. 1. Agin, G.J. and T.O. Binford, "Computer Des- criptions of Curved Objects", IEEE Tran- sactins on Computers, Vol. 25x03, -- April 1976. 30	1980	87
86	CONSTRAINT-BASED INFERENCE FROM IMAGE MOTION Daryl T. Lawton Computer and Information Science Department University of Massachusetts Amherst, Massachusetts 01003 ABSTRACT We deal with the inference of environmental information (position and velocity) from a sequence of images formed during relative motion of an observer and the environment. A simple method is used to transform relations between environmental points into equations expressed in terms of constants determined from the images and unknown depth values. This is used to develop equations for environmental inference from several cases of rigid body motion, some having direct solutions. Also considered are the problems of non-unique solutions and the necessity of decomposing the inferred motion into natural components. Inference from optic flow is based upon the analysis of the relative motions of points in images formed over time. Here we deal with environmental inferences from optic flow for several cases of rigid body motion and consider extentions to linked systems of rigid bodies. Since locality of processing is very important, we attempt to determine the smallest number of points necessary to infer environmental structure for different types of motion. I INTRODUCTION The processing of motion information from a sequence of images is of fundamental importance. It allows the inference of environmental information at a low level, using local, parallel computations across successive images. Our concern is with processing a particular type of image motion, termed optic flow, to yield environmental information. Optic flow [ll is the set of velocity vectors formed on an imaging surface by the moving projections of environmental points. It is important to note that there are several types of image transformations, caused by environmental motion, which are not optic flow. For example, image lightness changes due to motion relative to light sources, the motion of features produced by surface occlusion, moving shadows, and a host of transduction effects. The occurrance of these different types of image transformations requires explicit recognition so the appropriate inference technique can be applied for each. ---------------- This work was supported by NIH Grant NO. ROI NS14971-02 COM and ONR Grant No. NO001 4-75-C-0459. II CAMERA MODEL AND METHOD ---- The camera model is based upon a 3-D Cartesian coordinate system whose orgin is the focal point (refer to figure 1 throughout this section). The image plane (or retina) is positioned in the positive direction along, and perpendicular to, the Z-axis. The retinal coordinate axes are A and B. They are aligned with, and parallel to, the X and Y axes respectively. For simplicity and without loss of generality, the focal length is set to 1. A point indexed by the number i in the environment at time m is denoted by Pmi. The time index will generally correspond to a frame number from a sequence of images. The projection of an environmental point Pmi onto the retina is determined by the intersection of the retinal surface with the line containing the focal point and Pmi. The position of this intersection in the 3-D coordinate system is represented by the position vector Imi. In this paper, any subscripted I, A, or B, is a constant determined directly from an image. The significant relations concerning Pmi and Imi are I > ( X mi9 Yrni, Zmi J 2 > 3 > = mi 4 ) = i In the method used here, Equation 4 is used to transform expressed relations between environmental points into a set of equations in terms of image position vectors and unknown Z values. Solving these equations yields a set of Z values which provide a consistent interpretatiOn for the positions, over time, of the corresponding set of environmental points under the assumed relations. From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Retina Fig. 1 III INFERENCE FROM RIGID BODY MOTION ---- A. Arbitrary Motion of Rigid Bodies ---- The constraint equations developed for this case reflect the preservation of distances between pairs of points on a rigid body during motion. For two points i and j on a rigid body at times m and n, the preservation of distance yields 5 > - = - i i which expands into the image-based equation 6 > +2Zniznj(1nio Lj) - - 0 To determine a solution, we find the minimum number of points and frames for which the number of independent constraints (in the form of equation 6) generated equals or exceeds the number of unknown Z values. It is then necessary to solve the resulting set of simultaneous equations. Note that each such constraint is a second degree polynomial in 4 unknowns. We begin with the number of unknown Z values. For N (N>2) points in K (K>l) frames there are (NK)-1 unknown Z values. The minus 1 term reflects the degree of freedom due to the loss of absolute scale information. Thus, one of the Z-values can be set to an arbitrary value. The number of rigidity constraints generated by a set of N (N>2) points in K (K>l) frames is the product of 3*(N-2) and (K-1). The first term is the minimum number of unique distances which must be specified between pairs of points, in a body of N points, to assure its rigidity. Thus 4 points require 6 pairwise distances (all that are possible). For configurations of more than 4 points, its is necessary to specify the distance of each additional point to only 3 other points to assure rigidity. The second term is the number of interframe intervals. Each distance specified must be maintained over each interframe interval. The number of constraints is greater or equal to the number of unknowns when Thus minimal solutions (But not unique! necessarily see below) can be found when (N=5,K=2,number of constraint equationszg) or (N=4,K=3,number of constraint equations=l2), in agreement with [2]. The rigidity equations can be simplified by adding restrictions on allowable motions of environmental points. In the following sections we investigate two such restrictions. B. Motion Parallel to the XZ Plane -P-P Here the Y component of an environmental point is assumed to remain constant over time. Otherwise its motion is unrestricted. This corresponds to an observer moving along an arbitrary path in a plane, maintaining his retina at an Orientation perpendicular to the plane, with the motion of objects also so restricted. For Point i at times m and n this is expressed as 9 > Ymi = Zmi Bmi= Zni Bni “Yni 10) Z Bmi ni=Zmi- Bni This allows a substitution, for points i and j, which simplifies the rigidity constraint to 32 - - where the bracketed expressions are constants determinable from an image. This case has a direct solution using 2 points in 2 frames. To see this, consider points 1 and 2 at times 1 and 2. This yields a system of 4 unknowns: 211,212,221,222. The substitution allowed by equation 10 reduces it to a system of 2 unknowns, Zll and 212. Zll can then be set to an arbitrary value 9 reflecting scale independence. 212 is then determined from a constraint of the form of equation 11 rel$ing 212 and the evaluated variable Zll . This is a quadratic equation of 212. C. Translations The constraint expressing the translation of points i and j on a rigid body at times m and n is 12 > I= j j 13 > - i- ii where the operation is vector sub traction. This reflects the preservation of 1 eng th and orientation under translation. Setting Zmi to a constant value C, to reflect scale independence, in equation 13 yields 3 simultaneous 1 inear equations in 3 unknowns CAmi = ZmjAmj + ZniAni - ZnjAnj CBd =ZmjBmj+ZniBni-ZnjBnj C = zlTlj +z,i -Znj Thus environmental inference from translation requires 2 points in 2 frames. A potential implication of this case is for interpreting arbitrary, and not necessarily rigid body, environmental motion. If the resolution of detail and the rate of image formation relative to environmental motion are both very high, then, in general, the motion of nearby points in images can be locally approximated as the result of translational motion in the environment. D. Solving the Constraints A- The rigidity constraints are easily differentiable and can be solved using conventional optimization methods (taking care to avoid the solution where all the Z-values equal zero>. There are, however, in the case of arbitrary rigid body motion , generally many solutions. Here we consider ways of dealing with this. One way utilizes feedforward. It is crucial to note that the rigidity equations needn’t be solved anew at each point in time. If the environmental structure has been determined at time t and an image is then formed at time t+l, half the unknowns in the system of constraint equations disappear. This greatly simplifies finding a solution. Additionally, the solution process can be further simplified by extrapolation of infered motion, if enough frames have been processed . But how can the positions of the environmental points be determined initially? I> . Prior knowledge of the environment could supply the initial estimates of the relative positions. 2). There may be a small number (perhaps less than 50) of generic patterns of image motion (which may be termed flow fields), each associated with a particular class of environmental motion. For example, translational motion is characterized by straight motion paths which radiate from or converge to a single point on the retina. Other flow fields we have analyzed also have such distinguishing characteristics. These characteristics would be used to recognize particular types of image motion, associated with particular types of environmental motion, to initialize and constrain the more detailed solution process based upon solving the constraints for arbitrary rigid body motion. 3). The observer could constrain his own motion for one sampling period to a case of motion for which environmental strut ture can be unambiguously determined. For example, by stabilizing the retina of a moving observer with respect to rotations relative to a stationary environment , all image motion could be interpreted as the result of translation. Another possibility is to use more than the minimum required number of points in the inference process to supply additional constraints. 33 IV APPROACHES TO OTHER CASES OF MOTION ----- A. Sub-Minimal Rigid Configurations A sub-minimal configuration is one consisting of I,2 or 3 points over an arbitrary number of frames or 4 points in 2 frames. Human subjects can get an impression of 3-D rigid motion from displays of such configurations [4], even though there are not sufficient generated constraints, by the above analysis, for a solution. How is this possible? Other assumptions must be used for the inference. A potential one reflects an assumption of smoothness in the 3-D motion. This can be had by minimizing an approximation of the acceleration of a given point. For a point i at times t, t+l, t+2 this can be expressed as 15) Ipti- pt+l,i I - Ipt+l,i- pt+Z,i Perhaps a sum of such expressions, formed using the substitution of equation 4, should be minimized for a set of points over several time periods along with the satisfaction of expressible rigidity constraints. B. Johansson Human Dot Figures --- Experiments initiated by Johansson have shown the ability of subjects to infer the structure and changing spatial disposition of humans performing various tasks with only joint positions displayed over time [53. Here we consider how such inference could be performed. First, it is necessary to determine which points are rigidly connected in the environment. Work by Raschid C63 as shown that this is possible on the basis of image properties only, using relative rates of motion between images. That is, without any inference of environmental structure. Given the determination of rigid linkages, it is necessary to find the relative spatial disposition of the limbs. An approach is to infer environmental position, for each limb, using the rigidity constraints and optimizing the smooth motion measure discussed above. If the figure is recognized as being human, several object specific constraints can also be used. These would involve such things as allowable angles of articulation between limbs, their relative lengths, and body ACKNOWLEDGEMENTS I greatly appreciate Ed Riseman, Al Hanson, and Michael Arbib for their support and the intellectual environment they have created. I would also like to acknowledge the last minute (micro-nano-second!) heroics of Tom Vaughan, Earl Billingsley, Maria de LaVega, Steve Epstein, and Gwyn Mitchell. REFERENCES Cl1 c21 [31 c41 151 C61 Gibson, J.J. The Perception of the Visual -- World. BostoG-- Houghton Mifflin and Co. 1950. Ullman, S. The Interpretation of Visual Motion. Cambridge, Massachusetts:‘ The MIT Press. 1979. Rogers, D.F. and Adams, J.A. Mathematical Elements for Computer Graphics. McGraw-Hill Book Company. 1976. Johansson, G., and Jansson, G. "Percieved Rotary Motion from Changes in a Straight Line." Perception and Psychophysics, 1968, Vol. 4 (3). Johansson, G, "Visual Perception of Biological Motion and a Model for its Analysis." Perception and Psychophysics 14:2 (1973) 201-211. - Rashid, R.F. "Towards a System for the Interpretation of Moving Light Displays", Technical Report 53, Department of Computer Science, University of Rochester, Rochester, New York 146727, May 1979. symmetries. 34	1980	88
87	STATIC ANALYSIS OF MOVING JOINTED OBJECTS Jon A. Webb Department of Computer Science University of Texas at Austin ABSTRACT The problem of interpreting images of moving jointed objects is considered. Assuming the existence of a connectedness model, an algorithn is presented for calculating the rigid part lengths and motion of the jointed objects just from the positions of the joints and some depth information. The algorithn is proved correct. I INTRODUCTION Vision research has only recently begun considering the three-dimensional motion of jointed objects, but progress has been relatively rapid. This paper presents a method for using a very general model to discover the motion and structure of a jointed object. The method is proved correct under reasonable conditions, which are stated precisely. These condit,ions are found to be satified in most normal observation of normal jointed object movement. Jointed objects are important because they include most of the significant moving objects in this world, e.g. rigid objects, hunans, and animals. The method to be described allows the recovery of a wealth of information by a single monocular observer of a moving jointed object. This information could aid recognition from a distance. This paper, like most other research in three-dimensional motion (C1-41) * adopts the feature point model. In this model only the positions of points rigidly attached to the object are recorded. This method makes the mathematical analysis more direct. Moreover, psychological research has shown that humans can readily interpret movies of people where only the joints can be seen 15-71. It is therefore reasonable to try to construct a program that could interpret such images. II THE MODEL -- A. Introduction This paper assumes the existence of a ---------------- * This research supported by the Air Force Office of Scientific Research under grant nLpnber AFOSR 77-3190. connectedness model. This model could be constructed by the methods of [31 or by other methods under development. The jointed object model for a jointed object consists of three parts: joints, rigid parts, and feature points. The - - feature points are fixed on the rigid parts, which are connected by the joints. In this paper, it will be assumed that the jointed object forms a tree (i.e., that is has no cycles) and that the feature points coincide with the joints. The rigid parts are not allowed to bend or stretch. The lengths of the rigid parts are unknown, but are calculated by the algorithm through observation of the jointed object. A connectedness model for a humanoid figure is shown in figure 1. Feature points are indicated by letters. m ?” /i ‘e d a f I. j r’“\ k 1 i Figure 1. B. Description Input The analysis proposed in this paper applies equally well whether the central projection or parallel projection model of vision is used, but central projection will be assuned as it most accurately describes the way caneras work. The camera will be assuned to be at the or igin, with the focal plane at (O,O,f>. Figure 2 shows this model. Figure 2. The correspondence between model and image feature points must be established. The correspondence problem for moving objects has been considered in c2-41. These correspondence algorithms are based on nearest neighbor, and work 35 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. well ([31 reports 98% accuracy) for fraes with small time intervals between them. The algorithm to be described requires a z coordinate for some feature point in every frame. This point will be called the reference point. For simplicity, it will be assuned that the reference point is the same in every frame. The z coordinate of the reference point can be obtained by several means, including the support assumption (used in Cl1 for this purpose and proposed for psychological reasons in [91> but no method is entirely satisfactory. This will be discussed briefly in section IV. III THE ALGORITHM A. Introduction The algorithm treats the model the root being the reference point. this tree for the humanoid model. as a tree with Figure 3 shows The starting _ . point of a rigid part is its joint nearest the reference point (in this tree); its ending point is the joint farthest from the reference point. A first rigid part is said to be above a second if it lies on a path frcm the second to the reference point. Similarly, the second is said to be below the first. a d f h k i 1 Figure 3. The algorithn works by calculating the lengths and the positions of the ending points of the topmost rigid parts (these ending points are m, c, e, and a in figure 4). Next, rigid part lengths and ending point positions immediately below these rigid parts are calculated. The process continues until the positions of all the joints and the lengths of all the rigid parts have been calculated. The calculation of the lengths of rigid parts is done using known lower bounds on their lengths. These lower bounds are obtained from previous frames. (In the first frame a lower bound of zero is used). If the lower bound is too small to account for the observed positions of the joints, the smallest rigid part length that will work is calculated and a new lower bound is established. B. Formal Statement of the Algorithm -- For each frame, do the following for each rigid part in the tree, going from top to bottom: 1. Let the position of the starting point of this rigid part be (x,y,z), the observed coordinates of the ending point be (u,v>, and the lower bound on the rigid part length be r. If the rigid part length is exactly rr then the ending point lies on a sphere of radius r with center at (x,y,z>. At the sane time, the ending point lies on a line through the origin and (u,v,f), where f is the focal length. This situation is shown in figure 4. The coordinates of the ending point under these assumptions can easily be calculated using the quadratic formula. Figure 4. This method gives two values for the position of the end of the rigid part. These two values represent two reflections of the rigid part that could account for the observed position of the ending point. For the algorithm to work, and calculate the correct rigid part lengths, the correct reflection must be chosen. It is assumed that the correct reflection is always chosen by some process. While deciding which of the reflections is correct might be a hard problem (see section IV), once the correct reflection is chosen it can be tracked fairly easily since the two reflections normally differ greatly in the z coordinate, and in the angle they make at the starting point. 2. If the quadratic formula yields no value for the position of the end of the rigid part this means that the rigid part length must be longer than r. Calculate a new lower bound on the rigid part length by the formula (1) r = SQRT[(x-p~~)~+(y-pv)~+(z-pf)~l where (2) p= (ux+vy+fz) u2+v2+f2 The coordinates of the ending point are (Pu*Pv,Pf) l The situation giving rise to this formula is shown in figure 5. Figure 5. Whenever a rigid part length is changed, the previously calculated lower bounds on rigid part lengths below the changed rigid part become invalid, so they must be set to zero. This action introduces an order dependence into the algorithm; for the algorithm to work correctly, the proper view of a rigid part must be seen after the proper views of rigid parts above it are seen. This restriction will be discussed in greater detail later. 36 C. Experimental Results these restrictions can be removed. An experiment was run using three hand-drawn hunanoid figures and the algorithm given above. The figures were drawn with specific rigid part lengths in mind. The rigid part lengths were recovered by the algorithm to within an average relative error of about lo-15%. D. Proof of the Algorithm --- It will now be shown that the algorithm will eventually calculate the correct rigid part lengths and three-dimensional joint positions. In order to show this, these assunptions are necessary: 1. The correc known. t reflec tions of the joints must be 2. Each rigid part must be seen at some time in a position that satisfies figure 6. That is, the angle between the origin, the endpoint, and the starting point of the rigid part must be a right angle. 3. If rigid part A is above rigid part B, condition 2 must be satisfied for B after it is satisfied for A. Theorem. Under the above conditions, the givenwthm will correctly calculate the length and endpoint position for every rigid part. Proof. Let R be a rigid part. The proof will be by induction by the nLanber of rigid parts above R. If there are no rigid parts above R then R is attached to the reference point. As soon as condition (2) is satisfied for R formula (1) will correctly calculate R's length and R's endpoint will be correct. If there are any rigid parts above R then their correct lengths and endpoint positions will eventually be found. Once this has happened, conditions (3) guarantees that condition (2) will be satisfied for R, at which time formula (1) will be used to correctly calculate R's length. This completes the proof. IV EXTENSIONS TO THE ALGORITHM -- There are several restrictions placed on the data available to the system that are undesirable in the sense that hunans cannot make them in their observation of jointed objects. The most serious restrictions are the necessity of a connectedness model for the jointed object, needing a z-coordinate for the reference point in every frame, the necessity of knowing the correct reflections of the rigid parts, and the order dependence in rigid part views. These restrictions are necessary because the analysis of the moving object is only static, and does not take into account invariants in the object's motion. Dynamic analysis of the moving object is under active investigation and is yielding quite encouraging results that suggest that most, and perhaps all, of V SUMMARY A mathematical approach to the problem of jointed object observation has been presented. Given a connectedness model of the jointed object to be observed, the actual three-dimensional motion and rigid part lengths of the jointed object can be discovered by observation of the jointed object. This is done by constantly making minimizing assumptionqabout the object. Further research must take into account the actual motion of the object in a more sophisticated my. In order to overcome the deficiencies of the currently proposed method it is necessary to have a more complete understanding of how objects can be expected to move. ACKNOWLEDGEMENTS Fruitful discussions with J. K. Agtw~l. Larry Davis, Worthy Martin, and John Roach are gratefully acknowledged. REFERENCES 1. 2. 3. 4. 5. 6. 7. 8. 9. Roberts, "Machine perception L G, of three-dimensional solids.11 in Optical and electro-optical information processing, Jx Tippett, et al., Eds., 159-197. 1965. Ullman, S, The interpretation of visual motion. The MIT PreK Cambridge, MA. 19r Rashid. R F. "Lights: A study in motion.11 In Proc of the ARPA image understanding workshop, ---- Los Angeles, CA. 57-68. November 1979. Roach, J W, Determining the three-dimensional motion and model of objez from a sequence of --- -- images. Ph.D. dissertation, University of Texas at Austin, Department of Computer Science. May 1980. Johansson, G, "Visual perception of biological motion and a model for its analysis." Perception and Psychophysics, 2, (21, 201-211. 1973. - Kozlowski 9 L T and J E Cutting, "Recogni zing the sex of a wal ker from dynam ic point-l ight displays .I1 Perception and (61, 575-580. 1977. - Psychophysics, 21, Johan sson, G, "Spatio-temporal differenti ation and integration in visual mot ion percept ion." Psychological Research, 38, 379-383. 1976. Johansson, G and G Jansson 9 "Perceived motion from changes in a straight rotary line." Perception and Psychophysics, 2, (3). 165-170. 1968. - Gibson, J J, The perception of the visual world. Houghtofiifflin Co., E!ozn,950. 37	1980	89
88	A TECWIQUE FOR ESTABLISHING COMPLETENESS RESULTS IN THEOREM PROVING WITH EQUALITY Gerald E. Peterson Department of Mathematical Sciences University of Missouri at St. Louis St. Louis, MO 63121 ABSTRACT This is a summary of the methods and results of a longer paper of the same name which will appear elsewhere. The main result is that an automatic theorem proving system consisting of resolution, paramodul- ation, factoring, equality reversal, simplification and subsumption removal is complete in first-order logic with equality. When restricted to sets of equality units, the resulting system is very much like the Knuth-Bendix procedure. The completeness of resolution and paramodulation without the func- tionally reflexive axioms is a corollary. The methods used are based upon the familiar ideas of reduction and semantic trees, and should be help- ful in showing that other theorem proving systems with equality are complete. I INTRODUCTION A. Paramodulation Attempts to incorporate equality into automa- tic theorem provers began about 1969 when Robinson and Wos [6] introduced paramodulation and proved that if the functionally reflexive axioms were added to the set of clauses, then resolution and paramodulation constituted a complete set of in- ference rules. In 1975 Brand [l] showed that re- solution and paramodulation are complete even with out the functionally reflexive axioms,, Unfortunat ly, the usefulness of these results is limited because unrestricted paramodulation is a weak in- ference rule which rapidly produces mountains of irrelevant clauses. B. The Knuth-Bendix Procedure e- In 1970 Knuth and Bendix [2], working indepen- dently of Robinson and Wos, created a very effec- tive procedure for deriving useful consequences from equality units. Their process used paramodula- tion, but since it also used simplification and subsumption removal, most of the derived equalities were discarded and the search space remained small. The main defects of this procedure are that each equality must be construed as a reduction, so the commutative law is excluded, and the process works only on equality units, so most mathematical theo- ries, including field theory, are excluded. C. The Goal -- Since resolution and paramodulation constitute a complete set of inference rules, their use will provide a proof of any valid theorem if given suf- ficient (usually very large) time and space. On the other hand, the Knuth-Bendix process is effec- tive (usually small time and space) on a small class of theorems. We need to combine these two approaches and produce, if possible, an effective, complete prover. Some progress toward this goal has been re- ported. For example, the commutative law can be incorporated into the Knuth-Bendix procedure by using associative-commutative unification [4,5]. Also, restricted completeness results (i.e. the set of clauses must have a certain form) have been ob- tained for systems which appear to be more effec- tive than resolution and paramodulation [3]. D. Contributions of this Paper -- An impediment to progress toward the goal has been the lack of an easily used technique for ob- taining completeness results. We show here how the use of semantic trees can be generalized to provide completeness proofs for systems involving equality. We use this technique to obtain unrestricted com- pleteness results for a system which is thought to be fairly effective. The verification of effective- ness will require experiments which have not yet been performed. A. Semantic Trees II METHODS AND RESULTS One approach to obtaining completeness theorems is the use of semantic trees. To obtain a semantic tree T(S) for a set S of clauses, we first order the atoms of the Herbrand base I3, 87 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. say 13 = 1~1,~~'". 1. Then we build the binary tree T by giving each node at level k-l two sons labelled Bk and sBk, respectively. There will then be a one- to-one correspondence between the branches of T and the Herbrand interpretations. If the set S is unsatisfiable, then it will be falsified by every branch of T and as we move down a branch b of T we will come to a node n b at which it first becomes clear that b does not satisfy S. The node nb is called a failure node of T. The -- portion of T(S) from the root and extending up to and including the failure nodes is called the closed semantic tree for S, 'c(s). An inference node ofs a node whose children are both failure nodes. Every failure node nb has an associated clause Cb in S which caused the failure. That is, there is a ground instance CbO of C b such that if L is a literal of CbB, then QL occurs on b at or above nb with one such QL occurring at nbO It can be shown that the two clauses associated with the children of an inference node will resolve to produce a new clause C which causes failure at or above the inference node and therefore r(S U C) is smaller than -r(s). By performing a sequence of such resolutions we can eventually get the closed semantic tree to shrink to a single node and this will imply that the empty clause has been inferred. B. Incorporating Equality Problems arise when we attempt to use this pro- cess to obtain completeness results for systems in- volving equality. If S is E-unsatisfiable, then S is falsified by every E-interpretation but not nec- essarily by every interpretation. Thus it will only be on branches which are E-interpretations that failure nodes will exist in the usual sense. The other branches must be handled in some other manner. 1. Altering Interpretations The approach we use is to alter an arbi- trary interpretation I in a way such that the re- sulting interpretation I* is an E-interpretation. If I is itself an E-interpretation, then no altera- tion is needed, I* = I. The alternation is made as follows. First order 8 in a way such that each equality atom occurs before any atom which contains either side of the equality as a subterm. (Other restrictions are also needed on this order.) For an arbitrary interpre- tation I, define a partial order +(I) on 8 such that A-+B means essentially that B has been obtained from A by replacing a subterm s of A by a term t and I(s=t) = T. Now define I(A) as I(A) if A is irre- ducible anfd as I(B) if A-tB. 2. Substitutions For ground substitutions 0, 0' (0 = l$l+-tl,...,vk+tk)), we write 843' if 8' is identical to 8 except that one term t. of 8 has been replaced J by t! and t.+t!. J J J We say that 8 is irreducible if every term ti of 8 is irreducible. Suppose CBl and (Xl2 are ground instances of a clause C. If 81+e2 then I(cel) = I(Cf3,). 3. Failure Nodes Let Ib be the interpretation associated with a branch b of T(S). Then 1;; will be an E-in- terpretation and will, therefore, be falsified by some clause in S. That is, there will be a ground instance CB of a clause C in S such that I;l(Ce) is false and 8 is irreducible (Ib). (If 8 were re- ducible we could, by the previous paragraph, reduce it to an irreducible 8' such that I;(Cel) = F.) Every literal L of C8 will be falsified by 1; and there will exist a failure node nb such that %L occurs on b at or above n b with one such SL occur- ring at n b' These failure nodes can be split into two categories as follows. An R failure node, nb, -- is one such that the associated clause C is irre- ducible (Ib) (thus Ib(C) = F) and a P failure node --- is any failure node which is not an R failure node. 4. Inference Nodes The two categories of failure nodes lead to two categories of inference nodes. A resolution inference node is a node with two R failure node children ams essentially the same thing as an inference node in a semantic tree for a set without equality. A paramodulation inference node is a P failure node nb such that every equalityode ances- tor of nb has a brother which is an R failure node. 5. Summary of the Completeness Proof It is easy to show that if S has no E- model, then r(s), the closed semantic tree for S, has either a resolution or paramodulation inference node. If 'c(s) has a resolution inference node, then there will be a resolvent C of two clauses of S such that T(S U C) is smaller than T(S). If -c(s) has a paramodulation inference node n b' then there is a clause C28 such that I;(C,e) = F and C2B isreducible (Ib), say C28+E. Now C28 reduces to E using some equality s=t such that I,(s=t) = T. Since s=t occurs in the ordering 88 REFERENCES of 8 before the atom in C28 to which it applies, a node labelled s=t occurs on b above n b' This node has a brother which is an R failure node and hence there is a clause Cl0 such that s=t is a literal of Cl0 and if L is any other literal of Cl@, then Ib(L) = F. It follows that Cl6 and C28 have a para- modulant C'. This ground paramodulation can be lifted to the general level since 8 is irreducible and therefore s must start somewhere in C2. (The lifting lemma for paramodulation holds only in this case.) Thus there is a clause C which is obtained by paramodulating Cl into C2 and which has a ground instance C1 which is more reduced than C28. This greater reduction can be the basis for an ordering of the closed semantic trees involved and in the sense of this order, the tree for S U C will be smaller than the tree for S. C. Deletion of Unnecessary Clauses - 1. Subsumption clauses 2. Completeness is not lost if subsumed are deleted from S as the proof proceeds. Simplification If Cl = (s=t), C2 contains an instance so of s as a subterm, and s0e > toe for all ground stitutions 8, then the clause C = C,[to] is a sub- simplification of C2 using Cl. If a clause C has been simplified, then C may be deleted. (Our proof of this fails when the atom simplified is an equality of a certain form, but there are other reasons for believing it is still valid in this case.) D. The Final Result --- A complete system for first-order logic with equality may consist of resolution, paramodulation, factoring, equality reversal, simplification, and subsumption removal with the following restrictions. 1. Simplification and subsumption removal are given priority since they do not increase the size of s. 2. No paramodulation into variables. 3. All paramodulat ions replace s by t where for at least one ground substitution 8, se > te. 111 [21 [31 I141 151 161 D. Brand, "Proving Theorems with the Modifi- cation Method." SIAM J. Comput. 4 (1975) -- 412-430. D. E. Knuth Problems in and P. B. Universal Bendix, "Simple Word Algebras." in Leech. J. (ed.), Computational Problems in Abstract Algebras, Pergammon Press, 1970, 263-297. D. S. Lankford and A. M. Ballantyne, "The Refutation Completeness of Blocked Permutative Narrowing and Resolution." presented at the fourth workshop on automated deduction, February 1-3, 1979. D. S. Lankford and A. M. Ballantyne, "Decision Procedures for Simple Equational Theories with Commutative-associative Axioms: Com- plete Sets of Commutative-associative Reduc- tions." Technical Report, Mathematics Department, University of Texas at Austin, August 1977. G. E. Peterson and M. E. Stickel, "Complete Sets of Reductions for Some Equational Theories." to appear in JACM. G. A. Robinson and L. Wo s, "Paramodulation and Theorem Proving in F irst Order Theories with Equality." Machine Intelligence 4, American Elseviermork, 1969, 135-150. 4. If s > t then no reversal of the equality (s=t) will be necessary, and if Cl is obtained from C by reversing (t=s) then C may be deleted.	1980	9
89	BOOTSTRAP STEREO Marsha Jo Hannah Lockheed Palo Alto Research Laboratory Department 52-53, Building 204 3251 Hanover Street, Palo Alto, CA 94304 ABSTRACT Lockheed has been working on techniques for navigation of an autonomous aerial vehicle using passively sensed images. One technique which shows promise is bootstrap stereo, in which the vehicle's position is determined from the per- ceived locations of known ground control points, then two known vehicle camera positions are used to locate corresponding image points on the ground, creating new control points. This paper describes the components of bootstrap stereo. I INTRODUCTION Before the advent of sophisticated navigation aids such as radio beacons, barnstorming pilots relied primarily on visual navigation. A pilot would lookart the window of his airplane, see landmarks below him, and know where he was. He would watch the ground passing beneath him and estimate how fast and in what direction he was moving. Today, there exist applications for which a computer implementation of this simple, visually oriented form of navigation would be useful. One scenario hypothesizes a small, unmanned vehicle which must fly accurately from its launch point to its target under possibly hostile circum- stances. I I 1 2 - L TERRAIN PROFILE KNOWN FEATURES -DETERMINE VEHICLE LOCATION ---DETERMINE LOCATION OF UNKNOWN TERRAIN FEATURES TIME POSITION OF CRAFT DETERMINE DERIVED FROM POSITION OF I : I 0 0, b I - 1 1 0, b c, d t 2 I c. d I e.f Figure 1 Navigation Using Bootstrap Stereo. Cur overall approach to the problem involves providing the vehicle with a Navigation Expert having approximately the sophistication of an early barnstorming pilot. This expert will navi- gate partly by its simple instruments (altimeter, airspeed indicator, and attitude gyros), but mostly by what it sees of the terrain below it. This paper covers one aspect of the Navigation Expert, a technique which we call bootstrap stereo. II THE BOOTSTRAP STEREO CONCEPT Given a set of ground control points with known real-world positions, and given the 1ocatioIs of the projections of these points onto the image plane, it is possible to determine the position and orientation of the camera which collected the image. Conversely, given the positions and or- ientations of two cameras and the locations of corresponding point-pairs in the two image planes, the real-world locations of the viewed ground points can be determined [l]. Combining these two techniques iteratively produces the basis for bootstrap stereo. Figure 1 shows an Autonomous Aerial Vehicle (AAV) which has obtained images at three points in its trajectory. The bootstrap stereo process begins with a set of landmark points, simplified here to two points a and b, whose real-world coordinates are known. From these, the camera position and orientation are determined for the image frame taken at Time 0. Standard image- matching correlation techniques [23 are then used to locate these same points in the second, overlapping frame taken at Time 1. This permits the second camera position and orientation to be determined. Because the aircraft will soon be out of sight of the known landmarks, new landmark points must be established whenever possible. For this purpose, "interesting points" -- points with a high likelihood of being matched [3] -- are se- lected in the first image and matched inthe second image. Successfully matched points have their red- world locations calculated from the camera posi- tion and orientation data, then join the landmarks list. In Figure 1, landmarks c and d are located in this manner at Time 1; these new points are later used to position the aircraft at Time 2. Similarly, at Time 2, new landmarks e and f join the list; old landmarks a and b, which are 38 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. no longer in the field of view, are dropped the landmarks list. Once initialized from a set of known land- marks, bootstrap stereo has four components -- camera calibration, new landmark selection, point matching, and control point positioning. Because camera calibration and control point positioning have been well covered in the photogrammetric and imaging literatures (e.g., [1], [4], [S], [63), we will discuss only landmark selection and point matching in the following sections. III NEW LANDMARK SELECTION Because the aircraft rapidly moves beyond the known landmarks, new landmark points must constant- ly be established. For this purpose, "interesting points" -- points with a high likelihood of being matched [3] -- are selected in the old image of each pair, then matched with their corresponding points in the new image and located on the terrain. Matching is done on the basis of the normal- ized cross-correlation between small windows of data (typically 11 x 11) around the two points in question. Matching has trouble in areas that con- tain little information or whose only information results from a strong linear edge, therefore such areas make poor candidate landmarks. To avoid mismatches from attempting to use such aras, various measures on the information in the window have been used, including the simple statistical variance of the image intensities over the window [2] and the minimum of the directed variances over the window [3]. We have combined these into another interest measure which we call edged variance, which appears to perform better than either of its components [7]. We have defined our interesting points to be those which are local peaks in our interest measure, with a lower bound established to reject undesirable areas. Figure 2 includes some examples of the application of this interest measure. IV POINT MATCHING The actual matching of points in an image pair is done by maximizing normalized cross- correlation over small windows surrounding the points. Given an approximation to the displace- ment which describes the match, a simple spiral- ing grid search is a fairly efficient way to refine the precise match [2J. To provide that initial approximation, we have employed a form of reduction matching [3]. We first create a hierarchy of N-ary reduc- tion images. Each NxN square of pixels in an image is averaged to form a single pixel at the next level. This reduction process is re- peated at each level, stopping when the image becomes approximately the size of the correlation windows being used. Matching then begins at the smallest images, with the center point of the first image being matched via a spiral search. There- after, each matched point spawns four points around itself, offset by half a window radius along the diagonals of the window. These are mapped down to the next level of images, carrying their parent's displacement (suitably magnified) as their suggested match approximation. These matches are refined by a spiraling search before spawning new points. This process continues until the largest images are reached, effectively setting up a grid of matched points. In our implementation of bootstrap stereo, reduction matching is used to determine approx- imate registration of the images and to initial- ize the second-order match prediction polynomials. Matching of old landmarks and of interesting points to create new landmarks uses these polynomials to predict an approximate match, which is then re- fined by a local search. Autocorrelation threshold- ing is used to test the reliability of the match, then points are located more closely than the image grid permits by parabolic interpolation of the X- and Y-slices of the correlation values. V ANEXAMPLE In Figure 2, we present an example of the control-point handling portion of bootstrap stereo. The original data set, a sequence of 3 images from a video tape taken over the Night Vision Laboratory terrain model, is shown in Figure 2a. Figure 2b shows the interesting points in the first image, indicated by + overlays. If these were the control points from a landmark processor, we would use them to locate the first camera. These landmark points are next matched with their corresponding points in the second image; Figure 2c shows the successful matches overlaid on the first and second images. From the image plane positions of these points, the position and orientation of the second camera are determined. Next, the areas of the second image which were not covered by matches are blocked out and interesting points are found in the uncovered areas, as seen in Figure 2d. The old landmark points and the interesting points are then matched in the third image, as shown in Figure 2e. The old control points from the second image are used to calibrate the third camera; the camera calibrations are then used to locate the matched interesting points on the ground, forming new control points. These two steps are then repeated for subsequent pairs of images in longer sequences. VI CONCLUSICNS When an autonomous aerial vehicle must navigate without using external signals or rad- iating energy, a visual navigator is an enticing possibility. We have proposed a Navigation Expert capable of emulating the behavior of an early barn- storming pilot in using terrain imagery. One tool 39 such a Navigation Expert could use is bootstrap stereo. This is a technique by which the vehicle's position is determined from the perceived positions of known landmarks, then uses two known camera positions to locate real-world points which serve as new landmarks. The components of bootstrap stereo are well established in the photogrammetry and image process- ing literature. We have combined these, with improvement, into a workable system. We are work- ing on an error simulation, to determine how the errors propagate and accumulate. VII REFERENCES Cl3 C23 Thompson, M. M., Manual of Photogrammetry, American Society of Photogrammetry, Falls Church, Virginia, 1944. Hannah, M. J., Computer Matching of Areas in Stereo Image=, PhD,Thesis, AIM#239, Computer Science Department, Stanford University, California, 1974. Moravec, H. P., "Visual Mapping by a Robot Rover", Proceedings of the 6th IJCAI, Tokyo, Japan, 1979. Duda, R. 0. and P. E. Hart, Pattern Classi- fication and Scene Analysis, John Wiley and Sons, New York, New York, 1973. Fischler, M. A. and R. C. Bolles, "Random Sampling Consensus", Proceedings: Image Understanding Workshoe, College Park, Mary- land, April 30, 1980. Gennery, D. B., "A Stereo Vision System for an Autonomous Vehicle", Proceedings of the 5th IJCAI, Cambridge, Massachusetts, 1977. Hannah, M. J., "Bootstrap Stereo", Proceedings: Image Understanding Workshop, College Park, Maryland, April 30, 1980. Figure 2 An Example of the Control-Point Handling for Bootstrap Stereo a> The original sequence of 3 images. b) The interesting points in Image 1. c) The matched points between Images 1 and 2. d) e) The areas of Image 2 covered by matches, The control points in Image 2 matched to with interesting points found Image 3. in the uncovered areas. 40	1980	90
90	LOCATING PARTIALLY VISIBLE OBJECTS: THE LOCAL FEATURE FOCUS METHOD SRI Internati onal, Menlo Park, California 94 025 ABSTRACT Robert C. process is robust, because it bases its decisions on grou s of mutual1 is rela P- ively fast, % consistent features, and it ecause it concentrates on key features that are automatically selected on the basis of a detailed analysis of CAD type of models of the objects. @I. INTRODUCTION There are several tasks that involve locating partial1 relative y easy tasks, such as 4 visible objects. The 9 range from ocating a single two-dimensional object, to the extremely difficult task of locating and identifying thrF;-t;TEnsional objects jumbled together in a bin. we describe a technique to locate and identi y . gap-3 overlapping two-dimensional objects on the basis of two-dimensional models. Sequential 1,2,3] approaches have and parallel [4,51roblem een taken to solve this F ' In the sequential approach, one feature a ter another is located and as much information as possible is derived from the position and orientation of each feature. This approach is fast because it locates the minimum number of features; however, if the ob'ects are complicated, determining the or a er of the features to be located may be difvficult. Development of the location strategy becomes even more difficult when mistakes are taken into account. In the parallel approach, all the features in an image are located, and then large grou s of r-;,g;ed to recogn'ze relaxation 5,7 t 7 8 :j",;tgisto;;$- niques can be used to determine the eature'groups. This ap roach is robust because it bases its decisions on a E 1 the available information, and the location strategy is straightforward because all the features are used. For even moderately complex objects, however, the quantit approac h of data to be processed makes use of this impractical on current computers. Described here is a method called the Local Feature FOCUS (LFF), that combines the advanta es of the sequential and parallel approaches, whi 3 e avoidin ifI some of their disadvantages. This is achieve by careful analysis of the object models and selection of the best features. Belles II. LOCAL FEATURE FOCUS METHOD The basic F rinciple of the LFF method is to locate one rela ively reliable feature and use it to partially define a coordinate system within which a f roup of other key features is located. Enough o the secondary features are located to uniquely identify the focus feature and determine the position and orientation of the object of which it is a part. Robustness is achieved by using a s arallel matching scheme to make the final ecisions, and speed is achieved by carefully selecting information-rich features. The idea of concentrating on one feature is not new; it has been use P urpose vision programs he ability to generate their secondary features automatically from objet models. This automatic feature selection, when perfected, will significantly reduce the need for peo !c le wil to program recognition procedures and thus make possible quick and inexpensive application of the LFF method to new objects. OBJECT MODELS FOCUS FEATURES AND THEIR SECONDARY FEATURES I # , I OBJECT IMAGE OF EXECUTION-TIME OBJECTS- PROCESSING IDENTITIES -AND POSITIONS Figure 1 THE TOP-LEVEL BLOCK DIAGRAM As Figure 1 shows, the analysis of ob'ect,, models is performed once during training c ime and the results of the analysis are used repeatedly during nexecution time, making this a preach particularly attractive when large t num ers of objects are to be processed. In the rest of this paper, we concentrate on the training-time analysis. III. ANALYSIS The goal of the analysis is to examine a model of an object (or objects) such as the one in Figure 2, and generate a fist of focus features and 41 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Figure 2 AN OBJECT MODEL Figure 3 AN EXAMPLE TO BE PROCESSED their associated secondary features. Given this information and a picture such as the one in Figure 3, the execution-time system tries to locate occurrences of the objects. In the current implementation of the system, ob'ects are modeled as structures of regions, each o I which is bounded TK b a'sequence of line segments and arcs of circles. e execution-time system uses a maximal-cli ue ?! P ra h-matching method to locate the groups o B ea ures objects. that correspond to occurrences of the Therefore, the analysis is tailored to produce the information required by the maximal- clique matching system. In particular, the description of each secondary feature includes the feature-type, its distance from the focus feature, a,E;t;r;ist of the possible identities for the The analysis to produce this information is perfirmed in five steps: (1) (2) (3) Location of interesting features Grouping of similar features Rotational symmetry analysis of each object (4) Selection of secondary features (5) Ranking of focus features. The purpose of the first step is to generate the set of all features of the objects that could be located at execution time. Typical features include holes, corners, protrusions, and intrusions. For the model in Fi ure 2, the set of features contains all 14 interna !? holes. In the second step, the set of features is P artitioned into subsets of "similar" features. eatures are defined to be similar if they are like1 For to be indistinguishable at execution time. t e model in Figure 2, feature detectors can a distinguish at most,,three "small holes," and large ty es !? ho es. $f holes: "slots," Therefore, the set of interesting features is partitioned into three subsets, each defining a possible focus feature. In the third step, T$ s mmetry anal I sis a complete rotational e rotationa of each object is performed [12]. symmetry is used to determine the number of structurally different occurrences of each feature. Because the model in Figure 2 is twofold rotationally symmetric, the features occur in pairs, the members of which are indistinguishable on the basis of the relative positions of other features of the object. Instead of four types of small holes, there are only two, one on the axis between between the slots and one off that axis. Figure 4 SECONDARY FEATURES FOR SMALL HOLES The fourth step in the analysis is the most corn licated. P The oal is to select secondary fea ures for each ? ecus feature. The secondary features must distinguish between the structurally different occurrences of the focus feature and determine the position and orientation of the object. In Figure 2, for example, given an occurrence of a small hole, what nearby features could be used to determine whether it is one of the holes on the axis or off of it? There are two slots close to the small hole on the axis and on1 one near the off-axis occurrence. In addition, g t e slots are at different distances from the holes. Let Dl be the distance between the on-axis small hole and its slots and let D2 be the distance from the off-axis small hole to the nearest slot. Figure 4 shows circles of radii Dl and D2 centered on the two different types of small holes. Tabulated below are the feature occurrences that are sufficient to determine the type of the small hole and compute the position and orientation of the object. ON-AXIS SMALL HOLE -- Two slots at Dl No slots at D2 OFF-AXIS SMALL HOLE -- No slots at Dl One slot at D2 The analysis in step 4 locates secondary features in two substeps. First it performs a rotational symmetry analysis centered on each structural1 different occurrence of a focus feature. T% is analysis builds a descri R tion of the object in terms of of features t at are similar and equidis ?! roups ant from the focus feature. Figure 5 shows the groups of features produced by the current system when focusing on one of the small holes. In the second substep, the anal sis iteratively selects groups of features from t ese K descriptions to be included in the set of secondary features associated with tne focus feature. Groups 42 REFERENCES are selected for their contribution to identifying an occurrence of the focus feature or determining the position and orientation of the object. Figure 5 FEATURE GROUPS ABOUT A SMALL HOLE The fifth and final step in the training-time analysis is the rankin 8 oal is to determine t e a of the focus features. The order in which the focus eatures should be checked at execution time. The current system simply ranks them according to the number of secondary features required at execution time. IV. DISCUSSION The LFF method is a simple combination of the se uential and parallel approaches. 9. It offers the re lability of a arallel approach and most of the speed of a sequen 7. la1 approach. The s eed is achieved by usin to define a coor 5 the location of the !f ecus feature inate system within which the other features are located. Quickly establishing a coordinate system significantly reduces the time required to find secondary features. The utilit reliabilitv of 4 of the LFF method depends on the ocatinn focus features and the number of structurally-different occurrences of these features in the objects. Fortunately, most industrial features. !i arts have good candidates for focus. time so they he problem is to find them at tra;;ing can be used at execution time. fact, the more information gathered at training IX-X-,, the more efficient the system at execution Also, as the training-time analysis is made more'automatic correspondingly less time is required of a human programmer. The current implementation of the training- time analysis forms the basis for a coS@n;;ly automatic feature selection system. extensions are ! ossible. could select ex ra features to guarantee tha For example the s si;z % execution-time system would function proper1 even if a prespecified number of mistakes were ma % the feature detectors. e by The system could use the orientation of a focus feature, if it exists, to determine the orientation of the feature-centered coordinate system. The system could also select two or.more groups of features at one time, which is necessary for some more difficult tasks such as distinguishing an object from its mirror image. Finally, the system could incorporate the cost and reliability of locating a feature in the evaluation of the feature. In conclusion, the LFF method is a combination of the sequential and parallel approaches that B rovides s eed and reliability for many two- imensiona P location tasks. The automatic selection of features makes it particularly attractive for industries such as the aircraft industry that have hundreds of thousands of different parts and cannot afford a special-purpose program for each one. hl r4 L-31 t41 II51 b1 [71 b1 191 r.103 Cl11 IAl S. Tsuji and A. Nakamura, "Recognition,,of an Ob'ect in a Stack of Industrial Parts, IJSAI-75 Tbilisi, Proc. (August i 975). Georgia, USSR, pp. 811-818 S. W. Holland, "A Programmable Compute: Vision System Based on Spatial Relationshi General Motors Research Pub. s, GMR-20 8 (February 1976). 1; W. A. Perkins, "A Model-Based Vision System for Industrial Parts," IEEE Transactions on Computers, Vol. C-27, pTp';;f26-143 (Yebruam . A. P. Ambler et al., "A Verzatile Computer- Controlled Assembly System Proc. IJCAI-73, Stanford, California, pp. 29m(August 1973). S. W. Zucker and R. A. Hummel. "Toward a Low- Level Description of Dot Clusters: Labeling Edge, Interior, and Noise Points," Computer Graphics and Image Proc ssing Vol. y, No. 5, R. C. Bolles, "Rgbust Feature Matching Through Maximal Cliques, Proc. SPIE's Technical Symposium on Imagi~plications Ior AUtOmatea Xidustrial inspection an?f-Kssembly, Wasnlngton, D.C. (April 19'19) . S. T. Barnard and W. B. Thomnson. llDi.SDaratV Analysis of Images )( To appear in IEEE- " Transactions h n Pattern Analysis amachine lntellieence ulv 1Ytw). -- R. 0. Duda and P. E. Hart, "Use of the Hough Transform To Detect Lines and Curves in Pictures," CACM, Vol. 15, No. 1, pp. 11-15 (January 1977)T S. Tsuji and F. Matsumoto, "Detection of Ellipses by a Modified Hou h Transformation," IEEE Transactions on Compu ers, ? -Pp. '/'('I-'(81 (Au~St 1 Y'(U) Vol. C-27, No. . J. T. Olsztyn_and L. Rossol, "An A gl,;;;:on o~~~o;~;x~ Vision to a Simulated Ap T IJCPR-73, Washington, D.C. Oct6bee3y D. F.,,McGhie, "Programmable Part Presenta- ~~~~iligence Research A plied to Industrial in the SRI Ninth Report on Machine Automation, pp. 39-44 &gust 1979). R. C. Bolles, "Symmetry Analysis of Two- Dimensional Patterns for Computer Vision," Proc. IJCAI-79, Tokyo, Japan, pp. 70-72 mst 43	1980	91
91	INTERFERENCE DETECTION AND COLLISION AVOIDANCE AMONG THREE DIMENSIONAL OBJECTS* N. Ahuja, R. T. Chien, R. Yen, and N. Bridwell Coordinated Science Laboratory University of Illinois at Urbana-Champaign Urbana, Illinois 61801 ABSTRACT Two methods for detecting intersections among three dimensional objects are described. The first method involves detecting overlap among the projections of the objects on a given set of planes. The second method uses a three dimension- al octree representation of the objects. Intersec- tions are detected by traversing the trees for the obstacles and the moving objects. Application of the methods to collision avoidance is discussed. I INTRODUCTION An important problem in the robotics manipula- tion of its environment is that of following colli- sion free trajectories when objects have to be moved. A representation or model of the spatial configuration of objects is necessary to obtain safe solutions, i.e. those which avoid collisions. In addition, a capability?0 plan an efficient trajectory, say, one following the shortest avail- able path between the starting and the goal con- figurations, is desirable. In either case, a procedure is required to decide if an arbitrary configuration involves occupancy of a given region of space by multiple objects. The problem of interference detection usually involves static objects whereas collision detec- tion refers to a situation where at least one ob- ject is in motion. Collision detection may be viewed as a sequence of intersection checks among appropriately defined static objects, and at appropriate time intervals. Thus the basic prob- lem appears to be that of detecting intersection among a set of objects. Given a configuration of objects, it is not hard to develop an algorithm to check any inter- sections among the objects. However, in a reason- ably complex environment and for reasonable speeds of object manipulation, e.g. translation, rotation, etc., the availability of a limited computational power may demand efficient procedures. At the heart of the design of such procedures lies the need for a good representation of the objects. Algorithms must be developed that take advantage of the properties of chosen representation to efficiently track the dynamic environment. solid in terms of a set of relatively simpler, planar patches used to approximate its surface. For the case of convex objects, Comba [21 obtains a pseudo-characteristic functioninterms of expres- sions for the planar patches. The function as- sumes nonpositive values in regions that approxi- mate the parts of the space where all the objects intersect. Maruyama 151 compares the minimal boxes containing the objects. Boyse 111 considers three different types of objects: solids, surfac- es and containers. Intersections are detected by checking if an edge of one object intersects a face of another. Collisions are detected by checking interference among the obstacles and the surfaces traced by the edges and faces of the mov- ing objects. Udupa cl11 uses two different representations for the manipulator and the environment. The ob- stacles are modelled by polyhedra (not necessarily convex) in Cartesian space. The manipulator links of the Scheinman arm are approximated by the mini- mum bounding cylinders. Abstractions are intro- duced by replacing the links by their axes and enlarging the obstacles proportionately, by retain- ing only the major link, etc. Lozano-Perez and Wesley [41 also grow the obstacles and shrink the moving parts such that collisions in the modified representation occur if and only if they occur in the original space. It is clear that the representation of the objects plays a major role in determining the feasibility and performance of any intersection or collision detection method using that representa- tion. This paper discusses two methods of repre- sentation with the aim of making the resulting interference checking operations more efficient. Section II examines a method based upon a set of two dimensional projections of the objects. It uses a conservative criterion to detect the occur- rence of interference. Section III discusses a representation of the objects by regions of a fix- ed shape, but varying sizes determined by a recur- sive decomposition of the space into successively smaller regions. Section IV outlines the use of the representations described in Sections II and III to collision avoidance. Section V presents some concluding remarks. In the past, mos,t of the related work has used polyhedral approximation of objects [1,2,4, 5,111. This facilitates a treatment of a complex From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. II PLANAR PROJECTIONS A planar projection of a three dimensional configuration of objects will always show an over- lap between the projections of any objects that intersect. However the reverse is not necessarily true, i.e., overlapping projections do not neces- sarily correspond to intersecting objects. Such false alarms may sometimes be decreased by consid- ering projections on a collection of planes. A configuration then is adjudged to be safe if each pair of objects has nonoverlapping projections on at lease one plane. However, for any given number and choice of planes, spatial arrangements of non- interfering objects may be devised whose project- ions overlap. An increase in the number of planes may only decrease, but not eliminate, the probabil- ity of erroneous decisions. The error may also be lower for objects with simpler shapes, e.g., convex objects. A. Basic Method ~- We will make the usual and reasonable assump- tion that objects may be represented by polyhedra. Each polyhedron is uniquely described by the coor- dinates of, and the adjacency relationships among its vertices. The projection of a polyhedron on any plane is determined by the projections of the vertices, using the original adjacency relation- ships. The set of vectors OV! for various i defines the vertices of the projecsion. To determine the shape of the polygonal projection, we must deter- mine its border edges. This is done by obtaining the convex hull of the projections of the vertices. For convex polyhedra, the edges of the convex hull will be a subset of the set of straight lines Vl V! where V. and V. are adjacent vertices. On the J other hind, noaconvex polyhedra may give rise to nonconvexpolygonalprojections. The convex hull of the corresponding projected vertices will include edges V! V! where Vi and Vj are not adjacent. The actual or er of the projection can be found by i: a replacing all such edges VI Vi by a sequence of edges V!- V! , V! VI-, ..t ,JV!- Jl J? J2 J3 VI-, where k > 2, j, = 1, j, = J, and Vj Jk-l Jk and V. r Jr+l are adja- cent, 1 5 r I k-l (fig. 1). Obtaining the convex hull of N projected vertices takes O(N log N) time ClOl. This also determines the overall complexity of obtaining a polygonal projection, since the com- putations required to obtain the vertices Vi are O(N). Thus, assuming that the fraction of edges of a polygonal projection that form concave cavities is low, the entire set of operations takes O(N log N) time. Given a set of M planes, projections of all the objects are obtained on each of the planes. The polygons corresponding to a given pair of ob- jects to be examined for intersection are then checked for overlap in the different planes. If a plane is found in which the two objects do not overlap, noninterference between them is guaran- teed. Otherwise the objects may or may not inter- sect. Shamos LlOl gives an O(m+n) algorithm to find the intersection of a convex m-gon with a convex n-gon. Shamos [lOI also describes a simpler algo- rithm suggested by Hoey. We use the latter. We are not interested in obtaining the exact region of intersection, but only in knowing if two given polygons intersect. For an m-gon and an n-gon, this needs only O(m+n> computations compared to the O(m log m) and O(n log n) computations required to obtain the individual polygons. When the projection of an object is nonconvex, we must extract a description in terms of convex polygons in order to make use of the above algo- rithm. One obvious way is to use the convex hull of each nonconvex m-gon. However this introduces an additional source of false alarm since the cav- ities in an object, appearing as concavities in its projection, are treated as solids. An alternative way is to decompose a noncon- vex polygon into a set of convex polygons (fig. 2). Each of these convex polygons must then be con- sidered in each of the intersection tests where , 45 Figure 1 A nonconvex polyhedron and its nonconvex poly- gonal projection. To obtain the actual pro- jection from its convex hull, the dotted line must be replaced by the remaining two sides of the triangle. A /e WY : ..A.*.. . . . . . . / Figure 2 Decomposition of a nonconvex polygonal projection into con- vex polygons. the parent polygon is involved. Pavlidis [71, and Feng and Pavlidis [31 give algorithms for con- vex decomposition. Schachter [91 constructs a partial Delaunay triangulation of an m-gon to obtain an O(rm> algorithm, where r is the number of concave vertices. B. Improving Efficiency The complexity of intersection detection in- creases linearly with the number of projection planes used, since each plane must be examined for overlap until a zero overlap is found, or the planes are exhausted. To avoid applying the intersection algorithm to pairs of polygons that are far apart, coarse tests of intersection on the envelopes of the polygons may be used. For exam- ple, rectangular envelopes may be obtained by identifying the points having minimum and maximum x or y coordinate values. Similarly, a circular envelope may be obtained in terms of the diameter of the polygon. An overlap between two envelopes can be detected in constant time. However, ob- taining either of the two envelopes itself takes O(m) time for an m-gon, and hence such a coarse test may be of marginal utility since the inter- section algorithm for polygons is already linear in the total number of vertices. Suppose we are considering a pair of objects that do not intersect. Also suppose that their projections on at least one of the planes do not overlap. Then we would like to discover such a plane at the earliest, It may be useful to order the planes for examination according to some mea- sure of the likelihood that a given plane demon- strates the desired relationships. An example of such a measure is the total black (polygonal) area. It takes O(m) time to compute the area of an m-gon. Thus the planes may be ordered for examination by performing a computation on each plane which is linear in the number of vertices in the plane. The choice of the appropriate planes depends upon the given configuration. In general, a mini- mum of three planes would appear to be desirable to work with three dimensional objects. The com- putation of projections is trivial when the three orthogonal planes are used. The three projections are obtained by successively replacing one of its three coordinates by zero. Other planes with convenient orientations may also be used. III THREE DIMENSIONAL REPRESENTATIONS Extensions of the methods for representing two dimensional images C8lmay be used for the rep- resentation of three dimensional objects. Thus MAT (medial axis transform), generalized cylinders and recursive subdivision of space may all be used among others. In this paper, we will be concerned with the third of these methods. A. Octrees Just as the plane is recursively divided into squares in the quadtree representation of the images [12, 131, the three dimensional space may be subdivided into octantsL143. We start with the entire space as one block. If a block is com- pletely contained within the object whose repre- sentation is sought, it is left alone. Otherwise, it is divided into eight octants (fig. 3a) each of which is treated similarly. The splitting contin- ues until all the blocks are either completely within or completely outside the object (fig. 3a), or a block of minimum allowed size is reached, reflecting the limits on resulution. A method of subdivision of space was also employed by Maruyama c61. However, he used rectanguloids. A rectangu- lar block was divided into two, along one of the three axes as determined by the region of inter- section of the block with the object. The recursive subdivision allows a tree des- cription of the occupancy of the space (fig. 3b). Each block corresponds to a node in the tree. Let us label a node black or white if it corresponds to a block which is completely contained within the (black) object or the (white) free space, respectively. Otherwise the node is labelled gray. The gray nodes have children unless they are of the minimum allowed size, in which case they are relabelled black. The free space covered by such nodes is thus treated as part of the object in order to be conservative in detecting interference. The final tree has only black or white leaves. Figure 3 (a) An object and its representation by recursive subdivision of space into octants. (b) Octree for the object in (a). north-east, The north-west, south-west and the south-east oc- tants in the upper layer correspond to the children having labels 1, 2, 3 and 4, respect- ively. The nodes corresponding to the octants in the lower layer are labelled 5, 6, 7 and 8. Dark (white) circles indicate black(white) leaves. The children nodes are arranged in increasing order of label values from left to right. 46 B. Interference Detection Suppose we are given the octree representa- tions of a pair of objects. Clearly, the objects intersect if there exists at least one pair of corresponding nodes in the two trees such that one of them is black, and the other is black or gray. Thus the existence of interference can be deter- mined by traversing the two trees in parallel. Let A and B denote a pair of corresponding nodes at any time during the traversal. If either A or B is white, we do not traverse their children, and continue the traversal of the remaining nodes. The depth of the traversal along each path down from the root is determined by the shallower of the two paths terminating in a white leaf. The time required is proportional to the number of nodes in the subset of the nodes traversed. To detect interference among n objects, n> 2, n trees must be traversed. The traversal below a certain node in a tree stops only if either the node is white, or the corresponding nodes in the remaining n-l trees are all white. The time re- quired depends upon the actual number of nodes visited. IV COLLISION AVOIDANCE The use of interference detection in robotics manipulators lies in planning trajectories of mov- ing objects in a given environment. The occupancy of any part of the space by more than one object can be used as indicative of a collision. To avoid a collision its imminence must be foreseen. Therefore, intersections should be detected in a modified space. Suppose that it requires at least a distance d to brake the speed of a moving object or to change its course to avoid a collision. The value of d may depend upon different factors in- cluding the speed. Then we must detect any obsta- cles within a distance d along the path of a mov- ing object. For the obstacles off the path of the object, the required safe distance can be easily computed in terms of the relative location of the obstacles. Similarly, for two moving objects the required safe distance is 2d if they are approach- ing each other head on, and larger otherwise, given by the sizes of the objects, their speeds and directions of motion. To use intersection detection for collision avoidance, the moving objects are grown Cl,111 by a thickness d. Any intersection among the modi- fied set of objects produces a timely warning of a possible collision. The representations of the (static) obstacles are computed only once. Thus, we have a fixed set of projections of the obstacles on the planes being used, and a single octree representing the space occupied by the obstacles. Each moving ob- ject is represented separately. Every time an interference check has to be made, the representa- tions of the moving objects are obtained. Thus, in case of the planar projections, the current state of each of the moving objects is projected on each of the planes, which already have the appropriate projections of the obstacles. The usual procedure to check intersection is then car- ried out. In case of the octree representation, a new tree is generated for each of the moving ob- jects. A parallel traversal of these trees and the obstacle tree detects any interference present. The above checks for intersection are applied to a snapshot of the potentially continuously varying spatial configuration of objects. Hence care must be taken to ensure that no collisions will occur between any two successive executions of the algorithm. This requires that not only should we detect any existing intersections, but also note if some objects are too close, and might collide before the instant when the algorithm is applied again. The safe distance D between any two objects is proportional to the speeds and relative loca- tions of the objects involved. The interval T between two successive applications of the algo- rithm is controlled such that the objects do not come closer than D during the time interval T. This is accomplished by growing the objects further by a distance D. Then the problem of collision avoidance only involves periodic intersection de- tection among the modified objects. V CONCLUDING REMARKS We have discussed two approaches to detecting interference and collision among three dimensional objects. The first method involves detecting over- laps among projections of the objects on a given set of planes. It may thus be viewed as a two and a half dimensional approach. The criterion used for detecting interference is a conservative one, since for any given set of planes there exist many spatial configurations of noninterfering objects whose projections overlap on each of the planes. The second method uses a three dimensional octree representation of the objects, and does not suffer an information loss from the reduction in dimen- sionality like the first method. Here, the inter- ference is detected by a parallel traversal of the octrees for the obstacles and for each of the moving objects. Computer controlled manipulation often invol- ves the use of a manipulator such as the Scheinman arm. Its degrees of freedom are given by the boom length and the joint angles. These parameters thus define a natural representation of the arm. In particular, the updating of the representation as the arm moves becomes trivial. However, for the other objects in the environment, the above representation is not very useful. In addition, the origin of the cylindrical coordinate system above moves with the manipulator. Therefore, the representation of the whole environment must be modified for each new position of the manipulator as noted by Udupa [ill. The use of the Cartesian space for the representation of all objects re- quires regeneration of the octree for the moving objects only. Use of hardware such as proximity sensors may significantly improve the efficiency of the proce- dures. For example, a "cluttered zone" signal from a certain sensor may channel the available computational power to a detailed investigation of the desired region of space, leaving the relative- ly safe areas unexamined. This essentially amounts to being a hardware implementation of comparing coarse envelopes of objects in parallel. The computation of projections of the poly- hedra in Section II may also be replaced by cameras in the appropriate planes. This will provide real time projection generation. The projections are then treated as a collection of two dimensional binary images. The projections of the moving ob- jects are tracked and their positions checked against the projections of the obstacles. We have not addressed the problem of finding good trajectories for moving an object from a giv- en source position to a given goal position. The representations discussed here, in conjunction with the planning strategies discussed in [4,11], may be used to develop the desired trajectories. Ex- periments employing the methods described in this paper, and two Scheinman arms are being carried out currently. ACKNOWLEDGEMENTS This work was supported in part by the United States Department of Transportation and Federal Aviation Administration under Contract DOT FA79- WA-4360 and the Joint Services Electronics Program (U. S. Army, U. S. Navy and TJ. S. Air Force) under Contract N00014-79-C-0424. Cl1 c21 c31 c41 c51 [61 REFERENCES J. W. Boyse, "Interference Detection among Solids and Surfaces," Comm. ACM 22, January 1979, pp. 3-9. P. G. Comba, "A Procedure for Detecting Inter- sections of Three-Dimensional Objects," Jnl. ACM, 15, July 1968, pp. 354-366. H. Feng and T. Pavlidis, "Decomposition of Polygons into Simpler Components," IEEE Trans. Comp. C-14, June 1975, pp. 636-650. T. Lozano-Perez and M. A. Wesley, "An Algo- rithm for Planning Collision-free Paths among Polyhedral Obstacles," Comm. ACM 22, October 1979, pp. 560-570. K. Maruyama, "A Procedure to Determine Inter- sections between Polyhedral Objects," Int. Jnl. Comp. Inf. Sci. 1, 3, 1972, pp. 255-266. K. Maruyama, "A Procedure for Detecting Inter- sections and its Application," University of Illinois Computer Science Technical Report No. 449, May 1971. c71 [81 c91 Cl01 Cl11 Cl21 Cl31 Cl41 T. Pavlidis, "Analysis Pattern Recognition 1, of Set Patterns," 1968, pp. 165-178. A. Rosenfeld and A. C. Kak, Digital Picture Processing, Academic Press, New York, 1976. B. J. Schachter, "Decomposition of Polygons into Convex Sets," IEEE Trans. Comp. C-27 November 1978, pp. 1078-1082. , M. I. Shamos, Computational Geometry, Springer-Verlag, New York, 1977. S. Udupa, "Collision Detection and Avoidance in Computer Controlled Manipulators," Proc. 5th Int. Joint Conf. Art. Intel., Cambridge, Massachusetts, 1977, pp. 737-748. G. M. Hunter and K. Steiglitz, "Onerations on Images Using Quadtrees,"-IEEE-Trans. Pattern Analysis Mach. Int. 1, April 1979, pp. 145- 153. A. Klinger and C. R. Dyer, "Experiments in Picture Representation using Regular Decom- position," Computer Graphics and Image Pro- cess% 5, 1975, pp. 68-105. C. L. Jackins and S. L. Tanimoto, "Ott-trees and their Use in Representing Three- dimensional Objects," University of Washing- ton Computer Science Technical Report, January 1979. 48	1980	92
92	AUTOMATED INSPECTION USING GRAY-SCALE STATISTICS Stephen T. Barnard SRI International, Menlo Park, California ABSTRACT A method for using gray-scale statistics for the inspection of assemblies is described. A test image of an assembly under inspection is registered with a model image of a nondefective assembly and the two images are corn ared on the basis of two statistical tests: a X 3 test of the two marginal gray-level distributions and the correlation coefficient of the joint distribution. These tests are made in local subareas that correspond to important structure, such as parts and subassemblies. The subareas are compiled in an off-line training phase. TheX* measure is most sensitive to missi?ng or damaged parts, whereas the correlation coefficient is most sensitive to mispositioned parts. It is also possible to detect overall lighting changes and misregistration with these measures. Two examples are presented that show how the tests detect two types of defects. I INTRODUCTION Binary machine-vision techniques have received a great deal of attention for industrial inspection [1,2,3,41. High-contrast lighting and thresholding may be used to obtain an accurate silhouette that can be processed at video rates to yield useful features, such as area, perimeter, centroid, and higher moments. In addition, structural information is available in the geometric relationships between the local features of the outline (holes, corners, and so on). This kind of information is sometimes sufficient for some industrial automation (IA) tasks, such as part identification and acquisition. Other tasks, however, are not so easily approached. Although many simple parts can be adequately represented by a mere outline, most assemblies cannot because they are typically composites of several overlapping parts or subassemblies. Binary techniques will not be effective in such cases because thresholding will not, in general, separate the important components. Thorough binary inspection of even simple parts may not be feasible if one wishes to find defects in surface finish or other types of defects not limited to the outline. Gray-scale techniques have lately received more attention [5,6,7]. Compared to binary methods, there is a great variety of ways to use gray-scale information. This paper describes an approach for exploiting gray-scale information for inspection in a very basic form. Statistical tests of marginal and joint intensity distributions are used to compare test assemblies with an ideal model assembly. These tests are very efficiently computed and they are sensitive to defects and discrepancies that are not easily handled in the binary domain. II REPRESENTATION We use a representation that directly specifies the expected appearance of the assembly. Important structures of the assembly are represented as subareas that are tagged for special consideration. Using this representation, an image of a test assembly is more-or-less directly compared to a model with a minimum of preprocessing. It is necessary to do some sort of photometric and geometric normalization to the test image to bring it into correspondence with the model. At the lowest level, an assembly is represented by a gray-scale image. At the highest level, an assembly is represented by a list of named subareas, each of which corresponds to a particularly meaningful segment. Attached to each of these subareas are properties that specify the important characteristics; these characteristics identify the corresponding segment as "good" or "defective." Ideally, these subareas could have arbitrary shapes and sizes, but, for now, think of them as rectangular windows. Each has a specific location and size in the normal reference frame. The inspection system begins with a representation of an ideal assembly, called the model. This may be constructed interactively in a training phase. Given a low-level representation of a test case (i.e., an image), the system proceeds to build a high-level representation by comparing segments of the test image to segments of the model. The first step is to bring the test image into geometric registration with the model image. This is not strictly necessary. We could directly compare regions in the normal reference frame 49 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. (i.e., in the model image) with regions in the translated and shifted reference frame. Nevertheless, geometric registration simplifies further processing by establishing a one-to-one correspondence between model pixels and test pixels. We have assumed that the positional variation of the test assemblies is restricted to translation and rotation. We must therefore determine three parameters--Ax, Ay, and 6~ . There are several ways to approach this problem. Binary techniques may be adequate to determine the normalization parameters to subpixel accuracy with a system such as the SRI vision module [3]. In the gray-scale domain, one may search for maximal cross- correlation, although this will probably be very time-consuming without special hardware. A potentially more efficient method is to find distinguishing local features in the test image and then match them to their counterparts in the model b1- Once the translation and rotation has been determined it is easy to register the test image using a linear interpolation [g]. III STATISTICAL COMPARISON Two statistical measures that are useful for comparing model and test subareas are theX* test and the correlationPcoefficient. They are both extremely efficient and simple to implement, and they are sufficiently different to distinguish two broad classes of defects. A. The X2 Test - -- TheX2 test measures the difference between two frequency distributions. Let hm(k) be the frequency distribution of gray-scale intensities in a model window. Let ht(k) be the frequency distribution of a test window. We can consider h to be a hypothetical ideal distribution. The X 2m test gives a measure of how far h+, deviates from the hypothetical distribution h,. The significance of the test depends on the number of samples. x2 z-y& ((h,(k)-h+) I2 --------------- . k ht(k) This yields a measure of difference, but, to be consistent with what follows, we want a measure of similarity. Let f = e XL/C where c is some positive constant. 7 is a measure of the similarity of two distributions (in the X2 sense). If the distributions are identical, then 7 will be unity; if they are very different, t will be close to zero. B. The Correlation Coefficient Let hmt be the joint frequency distribution of the model and test windows. That is, hmt(u,v) is the frequency with which a model pixel has gray value u and its corresponding test pixel has gray value v. Let ml be the mean of h, and m2 be the mean of ht. Let 0, be the standard deviation f12 be the standard deviation of ht. of h,, and The central moments of the joint distribution h mt are p(i,j> = 1 c ( (Xm(k)-ml)i (Xt(k)-m2)j) --- n k where X,(k) and Xt(k) are gray values of the kth pixel in the model image and and the test image, respectively. The correlation coefficient, p , is cc OJ) P ------- 7 O2 p is in the interval [-1,1]. If it has the value +l or -1 the total "mass" in the joint frequency distribution must lie on a straight line. This will be the case when the test image and the model image are identical, and P will be +l. In general, if there is a linear functional dependence between the test and model windows, P will be +l (or, in the extremely unlikely case that one window is a llnegativetl of the other, P will be -1). If the windows are independent distributions, however, P will be 0. We can reasonably expect that intermediate values will measure the degree of dependence between the two windows. c. r=p 7 is not sensitive to the location of pixels. It simply measures the degree of similarity between two marginal distributions. P , on the other hand, measures the degree to which the model pixels agree with their corresponding test pixels; therefore, it is sensitive to location. This implies that T is a good test for missing and severely damaged parts (for they are likely to change the distribution of the test pixels compared to the distribution of the model pixels), while P is a good test for the proper location of parts (for misalignment will change the joint distribution). A systematic change in lighting can also be detected. T would be small because the lighting change would change the intensity distributions in the test windows, but P would be large because the test and model distributions would still be well- correlated. If this observation were made for only one window, it would not be very meaningful. However, if we make the reasonable assumption that most of the windows locate regions that are not defective, this leads to the observation that a systematic pattern of small 7 and large P indicates a lighting change. Small detectable mi sregistration errors are also Small misregistration would produce large 7 because the margi .nal d istributions of the test windows would not be much different from the IV EXPERIMENTS model windows. On the other hand, P would be smaller than if there were poor registration because the windows would not correlate as well. The same result for a single window would be caused by a misplaced part, but, again using the assumption that most of the windows locate non- defective regions, a systematic pattern of large r and small P over many windows would indicate an overall misregistration. These relationships are summarized in Table 1. Table 1 Defect Pattern vs. 7 and P * OK Missing Misplaced Lighting Registration ++++~+***~+*~~~~~~*~~~~~~~~~ * r + LARGE SMALL LARGE SMALL LARGE * * (SYSTEMATIC) * P" LARGE SMALL SMALL LARGE SMALL * * )E D. A Two-Stage System - The gray-scale statistics discussed above provide a basis for an integrated minimal system for inspection that is composed of two modules--a training module that is u&d off-line and that allows a human operator to build a high-level description of the assembly, and an inspection module that matches images of test assemblies to the model built in the training phase. In the training phase the operator works with an image of an ideal assembly, identifying the important parts that are likely to be defective and describing the modes of defects that may be relevant. For example, the location of a particular part may not be precisely fixed, but rather permitted to range over a rather large area. In this case the operator might indicate that 7 (the location insensitive measure) is a relevant test, but not P. In another case there may exist areas that have extremely variable appearance, perhaps because of individual part identification markings, and these areas might be excluded from testing altogether. In the normal case, however, a part will be fixed in one location and orientation, perhaps with some tolerance, and the operator will merely specify allowable limits for 7 and P. The on-line inspection phase is entirely automatic. The knowledge about the assembly collected in the training phase is applied to specific test assemblies and a judgment is made as to how well what is known fits what is seen. The 7 and P measures are computed for each area and are used to derive probability estimates of the various types of defects. We have tried the f and p tests on two assemblies. Figure 1 and Table 2 show the results for a water pump. The upper left portion of Figure 1 is the "model" image of a nondefective pump in a normal position and orientation. Several windows representing important parts of the assembly are also shown. The upper right portion of Figure 1 is a "test" image of a defective assembly. dark pulley in the center is missing. The round, In the lower left part of Figure 1 the test image has been registered with the model. The lower right image is the difference between the model image and the registered test image, and has been included to _ indicate how close the registration is. Table 2 shows the t windows. Note that 7 and P are both very small and p values for the various for the (missing) pulley compared to the other (nondefective) parts, just as predicted. 7 is also small for the window representing the total assembly because this includes the defective pulley. Figure 1. Table 2 Pump Statistics * T (c=800) * o * +++++JHC~**************~* Total 8 0492 Y .801 Pulley + .236 8 .354 * <= Defect Link * .981 * .824 * spout * -919 * -904 * Clip1 * .862 9 0879 * Clip2 + 0978 * .780 * Clip3 * ,949 * .898 * ++~~++*~+~~~~~~~~~~ 51 Figure 2 and Table 3 show the results for a hardcopy computer terminal. In this case the write head is not positioned in the correct place. Note that the window for the "head" includes the entire area where it might be located. As predicted, the 7 value is high, while the P value is small. In practice this might not be considered a defect because it may be permissible for the head to be located anywhere along the track. If this were the case, the window could be tagged to indicate that P is not a relevant test. Figure 2. Terminal Table 3 Terminal Statistics Jc 7 (c=800) * p * +++*~++++***8~~~~*~~~~~~ Total * 0746 * .g10 * Platen * 0674 * .868 * Head * .890 * 0458 * <= Defect Keys * 0740 * 0923 * ~~8~~~~~~*~~8~~~~*~~~~~ REFERENCES 1. M. Ejiri, T. Uno, M. Mese, and S. Ikeda, "A Process for Detecting Defects in Complicated Patterns,n Computer Graphics and Image Processing, Vol. 2, pp. 326-339 (1973). 4. 5. 6. 7. 8. 9. G. J. Gleason and G. J. Agin, "A Modular Vision System for Sensor Controlled Manipulation and Inspection," A.I. Center Technical Note 178, SRI International, Menlo Park, California (March 1979). G. J. Gleason, A. E. Brain, and D. McGhie, "An Optical Method for Measuring Fastener Edge Distances," A.I. Center Technical Note 190, SRI International, Menlo Park, California (July 1979). M. L. Baird, "SIGHT-I: A Computer Vision System for Automated IC Chip Manufacture," IEEE Trans. on Syst., Man, Cybern., Vol. SMC- 8, Nomebruary 197'c W. Perkins, "A Model Based Vision System for Industrial Parts," IEEE Trans. Comput., Vol. C-27, pp. 126-143 (1978). R. J. Woodham, "Reflectance Map Techniques for Analyzing Surface Defects in Metal Castings," Ph.D. Dissertation, Massachusetts Institute of Technology, Cambridge, Massachusetts (September 1977). S. T. Barnard, "The Image Correspondence Problem," Ph.D. Dissertation, University of Minnesota, Minneapolis, Minnesota (December 1979). A. Rosenfeld and A. Kak, Digital Picture Processing (New York: Acamrn76). 2. W .C. Lin and C. Chan, "Feasibility Study of Automatic Assembly and Inspection of Light Bulb Filaments," Proc. 1445 (October 19?r IEEE, Vo1.63, pp. 1437- 52	1980	93
93	HUMAN MOVEMENT UNDERSTANDING: A VARIETY OF PERSPECTIVES Norman I. Badler Joseph O'Rourke Stephen Platt Mary Ann Morris Department of Computer and Information Science Moore School D2 University of Pennsylvania Philadelphia, PA 19104 ABSTRACT Our laboratory is examining human movement from a variety of perspectives: synthesis of anima- ted movements and analysis of moving images. Both gestural (limb and hand) and facial movements are being explored. Key Words and Phrases: Human movement, move- ment representation, motion analysis, computer vision, facial expression, facial analysis, con- straint network. I HUMAN MOVEMENT SYNTHESIS Our laboratory is examining human movement representation from a variety of perspectives, in- cluding synthesis of three-dimensional animated movements and analysis of moving images. These broad areas are further refined into gestural (limb and hand) and facial movements since very different sorts of actions take place in jointed skeletal movement and "rubber-sheet" facial distortions. Our human body model r5](Fig. 1) and hand model [2](Fig. 2) are based on spherical decompositions of three-dimensional objects. Our goals have been to develop representations for human movements which permit both analysis and synthesis, are "com- plete" in terms of the range of human movements, and yet provide a "high-level" interface suitable for specifying or describing movement. Many of these issues are addressed in a recent survey by Badler and Smoliar [6], so we shall emphasize only the more recent work in human motion analysis and facial expression synthesis. II HUMAN MOVEMENT ANALYSIS There have been a rather small number of at- tempts to analyze complex movement presented in time-varying images. Rashid [13] uses moving light displays to track body joints and infer three- dimensional connections. Tsotsos [14] describes non-articulated shape changes. Badler [l] attempts conceptual descriptions of rigid jointed movements. Recently O'Rourke [9,10] describes a computer sys- tem which accepts a sequence of two-dimensional images of a solid, three-dimensional body [5] per- forming some motion sequence (Fig. 3). The output of the system is a description of the motion as coordinate-versus-time trackings of all body joints and as movement instructions suitable for controll- ing the simulation of a human body model [3,4,15]. The simulation includes a detailed model of a human body which incorporates the structural relationships between the body parts and the physical limitations to relative movement between the parts, This model is used to guide the image analysis through a pre- diction/verification control cycle. Predictions are made at a high level on the basis of previous analysis and the properties of the human model. The low level image analysis then verifies the pre- diction and the model is adjusted according to any discrepancies found or new knowledge acquired. This cycle is repeated for each input frame. The information extracted from the image is in- tegrated into the current position of the model through a constraint network operating on real- valued coordinate regions in three-dimensional space. Possible positions of body features are represented as unions of orthogonally-oriented rectangular boxes. Relationships among body parts-for example, distance constraints imposted by the body skeleton- are enforced in the network. As new joint positio- nal information is extracted from the image it is added to the network and its geometrical conse- quences immediately propagated throughout the net- work, Only head, hands, and feet are located in the image space , yet all remaining body joints may be tracked by the geometric inference process in the network. Figure 4 shows the constraint boxes for each joint of the body given the moving images of Fig. 3. III FACIAL ANIMATION We are also investigating the representation and simulation of the actions performable on the human face, The face presents an interesting prob- lem for simulation, as it is composed of mostly in- dependent sections of non-rigid masses of skin and muscle, such as the areas around the eyes, mouth, and cheeks, This type of action is basically dif- ferent from gross body movement in that a facial action will affect the visible results of other actions performed in the same area of the face. 53 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Our internal representation of the face is based on FACS, the Facial Action Coding System [7]. The system categorizes basic actions performable and recognizable (by a human notator) on the face. It is also easily translated into a state-descrip- tion of the face, in terms of muscle contractions. A complete recognition and simulation system for the face would consist of a camera, computer processing to obtain an idealized internal representation of the action, and a simulation of the action perform- ed on a graphic output device. Once the camera image is obtained, analysis is performed to produce the AU (FACS Action Unit) state of the face. This analysis is relatively simple, as it consists of identifying the presence/absence of "features" such as wrinkles and bulges on the face. (Note that this analysis does not require "recognition" of a particular face, just good comparison between suc- cessive images of the same face.) The current tech- nique under investigation uses an improved method of "rubber-sheet" matching [8]. Each AU effects only a small set of muscles; their union gives the muscle-status of the face. The specified muscle contractions are then simulated on the face. The face is represented by a network of points and interconnecting arcs (Figs. 5 and 6) 1121. It also has a higher level of organization which par- titions the network into the skin surface and spe- cific muscles. (It is this muscle organization which distinguishes our work from that of Parke [XL].) The skin is a "warped" two-dimensional surface of points and arcs. The points represent the basic surface, while the arcs contain informa- tion specific to their locale, such as the elasti- city ("stretchiness") of the skin between the arc's points. Stretching the skin (by contracting a mus- cle) causes first local motion, followed by propa- gation of the skin distortion. Muscles are also points, connected to but beneath the skin. They are distinguished by being the initiation of the distortion of the skin surface. An AU is thus mere- ly a set of muscles, with appropriate magnitudes of initial force of contraction. IV FUTURES Our research into human movement understanding has the joint goals of achieving conceptual des- criptions of human activities and producing effect- ive animations of three-dimensional human movement from formal specification or notation systems 161. One application of this effort is in the synthesis and analysis of American Sign Language. A proto- type synthesizer for ASL is being designed to faci- litate experimentation with human dynamics and movement variations which have linguistic import for facial and manual communication. ACKNOWLEDGEMENTS The support of NSF Grants MCS76-19464, MCS78- 07466, and O'Rourke's IBM Fellowship are gratefully acknowledged. lx1 I: 23 13J PJ [5J 171 b31 El ho3 IKl l-12J 1133 t-143 REFERENCES Badler, N.I. Temporal scene analysis: Concep- tual descriptions of object movements. Tech. Rep. 80, Dept. of Computer Science, Univ. of Toronto, Feb. 1975. Badler, N.1, and J. O'Rourke. Representation of articulable, quasi-rigid, three-dimensional objects, Proc. Workshop on Representation of Three-Dimensional Objects, Philadelphia, PA, May 1979. Badler, N.I., J. O'Rourke, and B. Kaufman. Special problems in human movement simulation, Computer Graphics Summer 1980. Badler, N.I., J, O'Rourke, S.W. Smoliar, and L, Weber, The simulation of human movement by computer, Tech. Rep. Dept. of Computer and Information Science, Univ. of Pennsylvania, July 1978. Badler, N.I., J. O'Rourke, and H. Toltzis. A spherical representation of a human body for visualizing movement. IEEE Proceedings 67: 10 (1979) pp. 1397-1403. Badler, N.I. and Smoliar, S.W. Digital repre- sentations of human movement. Computing Sur- veys 11: 1 (1979) pp. 19-38. Ekman, P. and W. Friesen. Facial ActionCoding System. Consulting Psychologists Press,Palo Alto, CA. 1978. Fischler,M,A. and R.A. Erschlager. The repre- sentation and matching of pictoral structures. IEEE Tr. on Computers C-22: 1 (1973). O'Rourke, J. Image analysis of human motion. Ph.D. Diss. Dept. of Computer and Information Science, Univ. of Pennsylvania, 1980. O'Rourke, 3, and N.I. Badler. Human motion analysis using constraint propagation. To appear IEEE-PAM1 November 1980. Parke, F.I. Animation of faces. Proc. ACM Annual Conf. 1972, pp. 451-457. Platt, S. Animating facial expressions. MSE Thesis, Dept. of Computer and Information Science, Univ. of Pennsylvania, 1980. Rashid, R. Lights: A system for the inter- pretation of moving light displays. To appear IEEE-PAMI, Nov. 1980. Tsotsos, J. A framework for visual motion analysis. Ph.D. Diss. Dept. of Computer Science, Univ. of Toronto, 1980. [15] Weber, L., S.W. Smoliar, and N.I. Badler. An architecture for the simulation of human move- ment, Proc. ACM Annual Conf. 1978, pp. 737- 745, 54 Simulated movement Figure 4. Constraint boxes for mo vements in Fig. 3. showing some of th butline and muscles 55 Figure 7. Upper porti muscles; t has been pu right brow.	1980	94
94	AN OPTIMIZATION APPROACH FOR USING CONTEXTUAL INFORMATION IN COMPUTER VISION Olivier D. Faugeras Image Processing Institute University of Southern California Los Angeles, California 90007, U.S.A. ABSTRACT Local parallel processes are a very efficient way of using contextual information in a very large class of problems commonly encountered in Computer Vision. An approach to the design and analysis of such processes based on the minimiza- tion of a global criterion by local computation is presented. INTRODUCTION The problem of assigning names or labels to a set of units/objects is central to the fields of Pattern Recognition, Scene Analysis and Artificial Intelligence. Of course, not all possible names are possible for every unit and constraints exist that limit the number of valid assignments. These constraints may be thought of as contextual information that is brought to bear on the particu- lar problem, or more boldly as a world model to help us decide whether any particular assignment of names to units makes sense or not. Depending upon the type of world model that we are using, the problem can be attacked by discrete methods (search and discrete relaxation) or continuous methods (continuous relaxation). In the first case our contextual information consists of a description of consistent/compatible labels for some pairs, or more generally n-tuples of units. In the second case the description includes a numerical measure of their compatibility that may or may not be stated in a probabilistic framework. Initial estimates of likelihoods of name assignments can be obtained from measurements performed on the data to be analyzed. Usually, because of noise, these initial estimates are ambiguous and inconsistent with the world model. Continuous relaxation (also sometimes called probabilistic relaxation or stochastic labeling) is thus concerned with the design and study of algorithms that will update the original estimates in such a way that ambiguity is decreased and consistency (in terms of the world model) is increased. This work was supported in part by the Defense ARPA contract F-33615-76-C-1203. The author is with the Image Processing Institute and Department of Electrical Engineering, USC, Los Angeles, California 90007. More precisely, let us denote by "Lithe finite set of N units and by ztthe finite set of M possible labels. In the discrete case, the world model consists of an n-ary relation RC("L(x,qn. The fact that the n-typle {(ul,~,),...,(un,~n)} belongs to R means that it is valid to assign name ai to unit Ui for i=l,...,n. In the continuous case, the world model consists of a function c of WQP into a closed interval [a,b] of the real line: c: (QxZ~)~ + [a,bl In most applications [a,bl=[O,l] or [-l,l] and n=2. The numbers c(ul,~l,...,un,Rn) measure the compa- tibility of assigning label Ri to unit ui for i=l ,**, n. Good compatibility is reflected by large values of c, incompatibility by small values. We will present in this paper two ways of measuring the inadequacy of a given labeling of units with respect to a world model and show that these measures can be minimized using only local cooperative computation. We will compare this approach with the original probabilistic relaxation scheme proposed by Rosenfeld, Hummel and Zucker [3] and a matching scheme proposed by Ullman [6]. To conclude the section, we will discuss the possibility of using Decentralization and Decomposition techniques to alleviate the curse of dimensionality and show how the Optimization approach can be extended very easily to the analysis of multilevel, possibly hierarchical, systems. We will not discuss in this paper any specific application. For an early application to scene analysis and discussion of some of the issues addressed in this paper, see [2]. For recent surveys, see [l] and [4]. For an applica- tion to graph matching, see [18]. I. Basic Optimization Based Probabilistic Relaxation Scheme We assume that attached to every unit ui are measures of certainty pi(a), for kin gthat can be thought of loosely as probabilities c Pi(R) = ' (1) R in ;;e 56 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. The wor_ld model is embedded in a function c mapping m .aL into [O,l], Again, c(ul,R,u2,m) measures the compatibility of calling unit ul,~ and unit u2,m. This function also allows us to define a topological structure on the set of units by assigning to every unit Ui and label R in ZT!Z. a set Vi(R) of related units uj for which these exists at least one label m in 2 such that c(ui,R,uj,m) is defined. A compatibility vector di is then computed for every unit ui that measures the compatibility in each label R in Xwith the current labeling at related units in Vi(a). The simplest way of defining Qi(a) is [l]: c Qi(a) = +I uj in v (a) Qij(R) (2) i where /Vi(g) 1 is the number of units in the set Vi(a) and Qij(") is given by: Qij (‘) = c C(Ui,R,U j ,m)Pj (ml (3) m in 2 Loosely speaking, Qi(a) will be large if for many units U in Vi(a), the compatible labels (that is the labgls m such that c(ui,R,uj,m) is close to 1) have high probabilisties, and low otherwise. In some cases the compatibility coefficients may be given a probabilistic interpretation, that is c(ui,R,uj,m) is the conditional probability pij(R/m) that unit ui is labeled R given that unit uj is labeled m. The next step in designing the Relaxation scheme is to specify a way of combining the two sources of information thatwe can use, i.e. the initial probabilities and the contextual informa- tion, to update the label probabilities. This updating should result in a less ambiguous and more compatible overall labeling in a sense that will remain vague until later on. Rosenfeld et al. [3] proposed the following iterative algorithm: for every unit ui and every label R in di set p(n+l) (a) = i (4) The denominator of the right hand side is simply a normalizing factor to ensure that numbers p("+l)(~) still add up to one. Intuitively, the libels R whose compatibility Qi(a) is larger than others will see their probability increase whereas the labels with smaller compatibility will see their probability decrease. One criticism with this approach is that it does not take explicitly into account measures of the two most important characteristics of a labeling of units, namely its consistency and its ambiguity. Faugeras and Berthod [5,7,81 have proposed several such measures and turned the labeling task into a well defined optimization problem which can be solved by local computations in a network of processors. We saw before that we can asso$iate with every unit Ui a probakility vector pi and a compatibility vector Qi whose components are given by equation (2). In general, the vectors di are not probability vectors in that their components do not sum to 1. This can be easily changed by normalizationand we can define: qi(a> = Qi (‘) c Q; (4 m in i I The vectors G, are now probability measure of coi&sistency for unit u; can be defined as the vector norm1 (5) vectors and a (local measure) (6) where llIlcan be any norm (in practice the Euclidean normb Similarly a local measure of ambiguity can be defined as Hi = R in ~pi(")(l-pi(")) = '-ll~ill~ (7) where ~~~~2 is th e Euclidean norm. Combining the two measures yie Ids a local criterion Ji = clCi + BHi (8) where c1 and 8 weight the relative importance we attribute to ambiguity versus consistency. A global measure can then be defined over the whole set of units by averaging the local measures. Using the arithmetic average for example, we define J= Ji (9) all units ui The labeling problem can then be stated as follows: given an initial labeing {pi '('))of the set of units Q, find the local minimum of the function J closest to the initial conditions, subject to the constraints that the vectors p. are probability vectors. More precisely, &is implies that c pi(R)=1 and pi(~)>0 for all units - R in 2 U. 1 (gal Aside from the fact that we are now confronted to a well-defined mathematical problem which can be tackled using Optimization techniques, we are also sure that some weighted measure of inconsistency and ambiguity is going to decrease. 57 As pointed out in [8], one minor drawback with the definition (6) of consistency is that its n$nimiz%tion implicitly implies the minimization of qi and pi and therefore the maximization of the entropy term H. (equation (7)). Thus there is an inherent problhm with the definition (8) in the sense thai consistency and ambiguity tend opposite directions. One very simple way resolving that contradiction is to define measure of both ambiguity and consistency J! = 1 -Gi l ;;i where 0 denotes the vector inner product, definition of a global criterion proceeds before: J' = c J! all units u. 1 1 and the labeling problem can be stated as to go in of a local as (10) The now as (111 (&I, replacing J with J'. This is similar to the minimal mapping theory developed by Ullman [6] for motion correspondence. Given two image frames with elements ui in the first one (our units) and element k (our names) in the second one, he studied the problem of selecting the most plausible correspondence between the two frames. Defining the cost qi(k) of pairing element ui with element k and the variables pi(k) equal to 1 if Ui is paired with k and 0 otherwisephe rephrased the motion correspondence probl em as a linear programming problem by defining the cost function (LPI 513, = c c all element u. all element k Pi(k)qi(k) 1 (12) which is precisely equation (11). The important difference between criteria J' and JA is that the costs qi(k) in (12) are not functions of the variables pi(k) whereas in (11) they are. In particular,minimizing J' is not an LP problem, in general. Nevertheless, the parallel between the two approaches is interesting and confirms that what we called the compatibility coefficients qi(k) defined in Eq. (5) are also a measure of the satisfaction/profit implied in assigning name k to unit u.. 1 II. Computational Requirements: Locality, Parallelism, Convergence As described in [7,9], imagine that we attach to the set Q of units and the sets Vi= U R in & ‘i CR) a simple network, that isapair <G,R> where G is a connected graph and R a set of processors, one for each node in the graph. There is a one to one correspondence between the units and the nodes of the graph on one hand, and the nodes of the graph and the processors on the other hand. This in turn implies that there is a one to one correspon- dence between the neighbors of the ith processor ri, i.e., the processors that are connected by arcs of G to r., and units in V.. 1 1 As shown in [7,8], the minimization of criteria J or J' can be achieved by using only local computation. More precisely, denoting by &7 (a function of all the vectors $i) either criterion J or J', we can attach to every unit ui a local gradient vector &-k ='Fi(cj) a Si (13) where Fi is a function of the vectors sj of units u. in the set Vi of neighbors previously defined. Eiplicit formula for the functions F. can be found in [4,5,7,8]. The iterative scheme is then defined as $n+ll +(n> i = Pi + p P. n l (14) where p is a positive stepsize and Pi a linear project?on operator de$e;T$ped by the constraints imposed on the vector pi [5,7], (for example that it is a probability vector). The main point is that both functions Fi and operator Pi can be computed by processor ri by communicating only with neighboring processors (local computation) while guaranteeing that the cost function 8 will decrease globally. It was stated before that a large amount of parallelism can be introduced in the process of minimizing criteria J and J'. This is achieved by attaching to every unit ui a processor ri connected only to processors r. attached to units u. related t0 Uim The global csiterion can then be Minimized by having processors ri perform simple operations mostly in parallel while a simple sequential communication process allows them to work toward the final goal in a coordinated fashion. If nonetheless our supply of processors is limited, we may want to split our original problem into several pieces and assign sequentially our pool of processors to the different pieces. The net result has of course to be the minimization of the original global criterion and some coordination must therefore take place. Solutions to this problem can be found in the so-called Decomposition and Decentralization techniques which have been developed to solve similar problems in Economics, Numerical Analysis, Systems Theory and Optical Control [12,13,14,15]. Decomposition techniques proceed from an algorithm standpoint: we are confronted with a problem of large dimensionality and try to substitute for it a sequence of problems of smaller dimensionalities. Decentralization techniques take a different viewpoint: we are confronted with a global problem and have at our disposal P decision centers. The question is whether it is possible to solve the global problem while letting the decision centers solve only local problems. The structure of criteria J and J' as sums of local measures allows us to develop both types of techniques [121. The key idea is to partition the set of units. For detailed algorithms, see [16]. III. Extension to Hierarchical Systems, Conclusions The optimization approach presented in Section I can be extended to the case where several labeling problems are present and embedded in a pyramid or cone structure with, for example, L levels. The different levels can be the same picture at different spatial resolutions as in [17] or represent different states of abstraction. For example the lowest level could be the edge element level, then the link level [lo], then the level dealing with elementary shapes like straight lines, ellipses, cubits, etc... These different levels form a multilevel system, each level having to solve a stochastic labeling problem. command vector for level i, that is Gi Let Vi be the is a NiMi dimensional vector, if there are Ni units and Mi possible classes, obtained in concatenating the probability vectors $j, j=l,...,$i. At leyel i we have to minimize a criterion J;(v,.$,.....v,). The fact that criterion Ji depends&up& the c&&and vectors at other levels accounts for the inter- action between the levels. A natural, but not always rational, way of solving this multilevel problem is to assume that every level i (i=l,... ,L) considers as given the command vectors of the other levels and minimizes its own criterion. The result is a nongcooperative equilibrium [12l or Nash point (ul,...,uL) verifying: Ji(;l,...,zi l,;i,:i+l,...,:L) < J.(;l,...& 1, -1 + -f vi>"i+19"'> L :I for all i and c.. This notion can certainly be criticized beca?ise by cooperating each of the L levels can, in general, improve its situation compared with the non-cooperative case. In other words, (61 t$e following situation is possible: if ,..+.,uL) set (Ui,.. i-y a Nash point, there exists another .,uL) of command vectors such that Ji(Zi ,...,$) < Ji(;l ,...,:L, for all i. This introduces the notion of Pareto points which, intuitively, are optimal in the sense that it is impossible to find another set of L command vectors that will decrease all criteria. It is possible to show that under very general conditions [12], Pareto points can be obtained by minimizing only one criterion! In other words if ?i=(?il,...,zL) is a Pareto point, then there exists L positive number Al ,...,AL such that 3 is a minimum of criterion J$ ,...,-;,) = L XiJi(;l,...,GL) i=l the ‘i’s can therefore be interpreted as weigh fat tors the different levels have agreed upon. .ting Another interesting possibility is to assume a hierarchical structure within the L levels, level 1 being the lowest and level L the highest. We then have a cascade of optimization problems similar to what happens in the price decentraliza- tion technique men$ioned i+n section II, where level 1 considers v2,... ,vL as given and computes + u1 = mJn Jl(Gl,c2,... ,:,I 3 This defines zl as a function of !2,... ,; . Then levzl $ solve5 th$ probl$m of minimizing kri$erion .JJ~ul(v2,...,vL),v2,...,vL) with respect to v2, . . . Even though the theory of hierarchical multi- level systems is still in its infancy it has been recognized for some time now [ll] that it carries the possibility of solving many difficult problems in Economics, Physiology, Biology [13,14,15], Numerical Analysis and Systems Theory [12], Optimal Control. It is clear that this theory is relevant to Image Analysis. In conclusion, we think that probabilistic relaxation techniques will play a growing role in the near future as building blocks of more and more complex vision systems. The need to quantify the behavior of these relaxation processes will become more and more pressing as the complexity of the tasks at hand rapidly increases and the global optimization framework offers a solid basis for this analysis. REFERENCES El1 PI [31 [41 [51 [61 [71 L.S. Davis and A. Rosenfeld,%ooperating processes for low-level vision: a survey," TR-123, Department of Computer Sciences, University of Texas, Austin, January 1980. H.G. Barrow and J.M. Tenenbaum, "MSYS: A System for Reasoning About Scenes," Tech. Note 121, AIC-SRI Int., Menlo Park, Ca., 1976. A. Rosenfeld, R.A. Hummel and S.W. Zucker, "Scene Labeling by Relaxation Operations," IEEE Trans. on Syst., Man, and Cybern., SMC-6, No. 6, pp. 420-453, June 1976. 0-D. Faugeras, "An Overview of Probabilistic Relaxation Theory and Applications," Proceedings of the NATO ASI, Bonas, France, June-July 1980, D. Reidel Publishing Co. O.D. Faugeras and M. Berthod, "Scene Labeling: An Optimization Approach," Proc. of 1979 PRIP Conference, pp. 318-326. S. Ullman, The Interpretation of Visual Motion, MIT Press, 1979. O.D. Faugeras and M. Berthod,"Improving Consistency and Reducing Ambiguity in Stochastic Labeling: An Optimization Approach," to appear in the IEEE Trans. on Pattern Analysis and Machine Intelligence, 1980. 59 [8 1 DOI Ull WI D31 1141 WI 1161 El71 W3 1 M. Berthod and O.D. Faugeras, "Using Context in the global recognition of a set of objects: an optimization approach," 8th World Computer Congress (IFIP 80). S. Ullman, "Relaxation and Constrained Optimization by Local Processes," Computer Graphics and Image Processing, 10, pp. 115- 125, 1979. S.W. Zucker and J.L. Mohammed, "A Hierarchi- cal System for Line Labeling and Grouping," Proc. of the 1978 IEEE Computer Society Conference on Pattern Recognition and Image Processing, pp. 410-415, Chicago, 1978. M.D. Mesarovic, D. Macho and Y. Takahara, Theory of Hierarchical Multilevel Systems, Academic Press, 1970. J.L. Lions and G.I. Marchuk, Sur Les Methodes Numeriques En Sciences Physiques Et Econo- Collect miques, ion Methodes Mathematiques de L'Informatique, Dunod, 1976. Goldstein,"Levels and Ontogeny,"Am. Scientist 50, 1, 1962. M. Polanyi, "Life's Irreducible Structures,11 Science 160, 3884, 1969. D.F. Bradley, "Multilevel Systems and Biology - View of a Submolecular Biologist,11 in Systems Theory and Biology (M.D. Messarovic, ed.), Springer 168. O.D. Faugeras, "Decomposition and Decentrali- zation Techniques in Relaxation Labeling," to appear in Computer Graphics and Image Processing, 1980. A.R. Hanson and E.M. Riseman, of Natural Scenes, in A. Hanson and Segmentation E. Riseman, eds., Computer Vision Systems, Academic Press, NY, 1978, 129-163. O.D. Faugeras and K. Price, "Semantic Description of Aerial Images Using Stochastic Labeling," submitted to the 5th ICPR. 60	1980	95
95	SWIRL: AN OBJECT-ORIENTED AIR BATTLE SIMULATOR Philip Klahr, David McArthur and Sanjai Narain;: The Rand Corporation 1700 Main Street Santa Monica, California 90406 ABSTRACT ROSS, an object-oriented language that has evolved over the last two years as part of the knowledge-based simulation research at Rand [1,3,6,7,8,91. We describe a program called SWIRL designed for simulating military air battles between offensive and defensive forces. SWIRL is written in an object-oriented language (ROSS) where the knowledge base consists of a set of objects and their In the following sections we discuss the goal of SWIRL, outline the main objects in the air-battle domain, and note how those objects and their behaviors map onto the ROSS objects and ROSS behaviors that constitute the SWIRL program. We discuss some of the problems encountered in designing an object-oriented simulation, and present our solutions to these problems. associated behaviors. problems we encountered We discuss some of the and in designing SWIRL present our approaches to them. I INTRODUCTION Object-oriented programming languages such as SMALLTALK [2], PLASMA [4], and DIRECTOR [5], as well as ROSS [S], enforce a 'message-passing' style of programming. A program in these languages II THE GOAL OF SWIRL - ---~ The goal of SWIRL is to provide a prototype of a design tool for military strategists in the domain of air battles. SWIRL embeds knowledge about offensive and defensive battle strategies and tactics. SWIRL accepts from the user a simulation environment representing offensive and defensive forces, and uses the specifications in its knowledge base to produce a simulation of an air battle. SWIRL also enables the user to observe, by means of a graphical interface, the progress of the air battle in time. Finally, SWIRL provides some limited user aids. Chief among these are an interactive browsing and documentation facility, written in ROSS, for reading and understanding SWIRL code, and an interactive history recording facility for analyzing simulation runs. This, coupled with ROSS's ability to easily modify simulation objects and their behaviors, encourages the user to explore a wide variety of alternatives in the space of offensive and defensive strategies and to discover increasingly effective options in that space. consists of a set of objects called actors that interact with one another via the transmission of messages. set of message Each actor has a set of attributes and a templates. Associated with each message template is a behavior that is invoked when the actor receives a message that matches that template. A behavior is itself a set of message transmissions to other actors. Computation is the selective invocation of actor behaviors via message pasSing. This style of computation is especially suited simulation in domains that may be thought of as to consisting of autonomous interacting components. In such domains one can discern a natural mapping of their constituent components onto actors and of their Indeed, interactions experts in many onto message transmissions. may find the domains object-oriented metaphor a natural one around which to organize and express their addition, object-oriented simulations can achieve knowledge In high intelligibility, modifiability and credibility III SWIRL'S DOMAIN - ~~ [1,6,91. However, while these languages provide a potentially powerful simulation environment, they can easily be misused, since good programming style in object-oriented languages is not as well-defined as in more standard procedural languages. In this paper we describe a program called SWIRL, designed for simulations in the domain of air battles, and use SWIRL to demonstrate effective simulation programming techniques in an object-oriented language. SWIRL is written in 5: Views expressed in this paper are the authors' own and are not necessarily shared by Rand or its research sponsors. In our air-battle domain, penetrators enter an airspace with a pre-planned route and bombing mission. The goal of the defensive forces is to eliminate those penetrators. Below we list the objects that comprise this domain and briefly outline their behaviors. 1. Penetrators. These are the primary offensive objects. They are assumed to enter the defensive air space with a mission plan and route. 2. GCIs. "Ground control intercept" radars detect incoming penetrators and guide fighters to 331 From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. intercept penetrators. 3. AWACS. These are airborne radars that also detect and guide. 4. SAMs. Surface-to-air missile installations have radar capabilities and fire missiles at invading penetrators. 5. Missiles. These are objects fired by SAMs. 6. Filter Centers. They serve to integrate and interpret radar reports; they send their conclusions to command centers. 7. Fighter Bases. Bases are alerted by filter centers and send fighters out to intercept penetrators when requested to by command centers. 8. Fighters. Fighters receive messages from their base about their target penetrator. They are guided to the penetrator by a radar that is tracking the penetrator. 9. Command Centers. These represent the top level in the command-and-control hierarchy. Command centers receive processed input about penetrators from filter centers and make decisions about which resource (fighter base) should be allocated to deal with a penetrator. 10. Target. Targets are the objects penetrators intend to bomb. Figure 1 shows an example snapshot of an air-battle simulation. A complete description of the SWIRL domain can be found in [7]. IV THE DESIGN OF SWIRL - -~- In this section we outline how the above flow of command and control among the different kinds of objects is modeled in ROSS. The first step in modeling in an object-oriented language such as ROSS is to decide upon the generic actors and their behaviors. A generic object or actor in ROSS represents an object type or class and includes the attributes and behaviors of all instances of that class. For example, the generic object FIGHTER represents each of the individual fighters that may be present in any simulation environment. Second, one may need to design a set of auxiliary actors to take care of modeling any important phenomena that are unaccounted for by the generic objects. A The Basic Objects 2 -~ We begin by defining one ROSS generic object for each of the kinds of real-world objects mentioned in the previous section. We call these objects basic objects. Each of these has several different attributes representing the structural knowledge associated with that type of object. For example, to express our structural knowledge of penetrators in ROSS, we create a generic object Figure 1. Graphical Snapshot of SWIRL Simulation. 332 called PENETRATOR and define its the following ROSS command: attributes using times. (The entire hierarchical SWIRL is given in [7].) organization for (ask MOVING-OBJECT create generic PENETRATOR with Each object type in the class hierarchy can be position 1 a position' construed as a description or view of the objects max-speed 'a maximum speed' below it. One object (AWACS) happens to inherit speed 'current speed' its behaviors along more than one branch of the bombs 'current number of bombs' hierarchy (via RADAR and MOVING-OBJECT). Such status 'a status' 'multiple views' Or 'multiple-inheritance' is flight-plan 'a flight plan') possible in ROSS but not in most other object-oriented programming environments. where phrases in single-quotes represent variables. C. Modeling Non-Intentional Events To capture the behaviors of each kind of - real-world object, we begin by asking what The basic objects and their behaviors have a different kinds of input messages each of these clear correspondence to real-world objects and real-world objects could receive. For example, a their responses to the deliberate actions of fighter can receive a message (a) from its fighter others. These actions comprise most of the base telling it to chase a penetrator under significant events that we wish to simulate. guidance from a radar, (b) from that radar telling However, there are several important kinds of it to vector to a projected intercept point with events that represent side effects of deliberate the penetrator, or (c) an 'in-range' message actions (e.g., a penetrator appearing as a blip on informing it that the penetrator is in its radar a radar screen is a side effect of the penetrator range. Each of these messages then becomes the flying its course and entering a radar range). basis for a fighter behavior written in ROSS. To Such events are important since they may trigger determine the structure of each of these behaviors other actions (e.g., a radar detecting a penetrator we ask what messages the object transmits in and notifying a filter center). However, these response to each of its inputs. For example, in non-intentional events do not correspond to response to a 'chase' message from its fighter real-world message transmissions (e-s., base, a fighter will send a message to itself to penetrator does not notify a radar that it ha: take off, and then send a message to the specified entered the radar's range). An important issue in radar requesting guidance to the penetrator. The the development of SWIRL has been how to capture following ROSS command captures this behavior: these non-intentional events in an object-oriented framework (i.e., via message transmissions). (ask FIGHTER when receiving (chase 'penetrator guided by >gci) One method of capturing non-intentional events (-you unplan all (land)) could be to refine the grain of simulation. The (-you set your status to scrambled) grain of a simulation is determined by the kind of (if (-you are on the ground) real-world object one chooses to represent as a then (-you take off)) ROSS object. A division of the air-battle domain (-requiring (-your guide-time) tell -the gci into objects like penetrators and radars is guide -yourself to -the penetrator)). relatively coarse grained; a finer grain is possible. In particular, one could choose to (The '-'s signal abbreviations. The ROSS create objects that represent small parts of the abbreviations package t81 enables the user to introduce English-like airspace through which penetrators fly. Then, as expressions into his programs and to tailor the expressions to his own penetrators move they would send messages to those preference. This approach towards readability is sectors that they were entering or leaving (just as 'send particularly flexible, since the user is objects moving through real space impact or not restricted to system-defined messages' to that space). Sectors could be any English associated with radars whose interface.) ranges they define, and act as intermediary objects to notify radars B. Organizing Objects Into a Hierarchy when penetrators enter their ranges. Essentially - - - this solution proposes modeling the situation at an almost 'molecule-strikes-molecule' level of detail The behaviors of basic objects often have many that are revealed in the process of since, by adopting this level, one can achieve a commonalities strict mechanical cause-and-effect chain that is defining their 'behaviors. For example, GCIs, AWACS and SAMs all share the ability to detect, and their simple to model via message transmissions between real objects. detection behaviors are identical. We can take advantage of ROSS's inheritance hierarchy (see [8]) However, although this method solves one to reorganize object behaviors in a way that both it causes two others that make it emphasizes these conceptual similarities and modeling problem, intractable. First, the method entails a eliminates redundant code. For example, for prohibitive amount of computation. Second, in most objects that have the behaviors of detection in cases, the extra detail would make abstract generic object modeling very common, we define a more called RADAR to store these common behaviors. We awkward and unnatural (at least for our purposes in building and using SWIRL). The natural level of then place it above GCI, AWACS and SAM in the decomposition is that of 'coarse objects' such as hierarchy, so that they automatically inherit the penetrator and fighter. To the extent we stray behavior for detection whenever necessary. Hence from this, we make the simulation writer's job more we avoid writing these behaviors separately three difficult since he can no longer conceive of the 333 task in the way that comes simplest or most y PERFORMANCE FIGURES The ROSS interpreter is written in MACLISP and runs under the TOPS-20 operating system on a DEC20 (KLlO). The space requirement for this interpreter is about 112K 36-bit words. SWIRL currently contains the basic and auxiliary objects mentioned above, along with approximately 175 behaviors. Compiled SWIRL code uses about 48K words. A typical SWIRL simulation environment contains well over 100 objects and the file defining these objects uses about 3K words. Total CPU usage for the simulation of an air-battle about three hours long is about 95 seconds. This includes the time needed to drive the graphics interface. VI CONCLUSIONS - We have found the object-oriented environment afforded by ROSS to be a rich and powerful medium in which to develop a non-trivial simulation program in the domain of air-battles. The program adheres to the criteria of intelligibility, modifiability and credibility laid out in [1,6,9]. Liberal use of the abbreviations package has led to highly English-like code, almost self-documenting in nature. BY adhering to several stylistic principles for object-oriented programming, such as the Appropriate Knowledge Principle and the Appropriate Decomposition Principle, we have been able to further enhance SWIRL's modifiability. This has enabled us to easily experiment with a wide range of air-battle strategies. Coupled with a graphics interface which allows us to quickly trace a simulation run and test the credibility of its behaviors, SWIRL provides a powerful environment in which to develop and debug military strategies and tactics. naturally to him: In summary, we reject this technique because it violates the following principle that we have found object-oriented simulators: useful in designing THE APPROPRIATE DECOMPOSITION PRINCIPLE: Select a level of object decomposition that is 'natural' and at a level of detail commensurate with the goals and purposes of the model. A second solution for modeling non-intentional events would be to allow the basic objects themselves to transmit the appropriate messages. For example, if we allow a penetrator (with a fixed route) to see the position and ranges of all radars, it could compute when it would enter those ranges and send the appropriate 'in-range' messages to the radars. This solution is computationally tractable. However, it has the important drawback that it allows the penetrator to access pieces of knowledge that, in reality, it cannot access. Penetrators in the real world know their routes but they may not know the location of all enemy radars. Even if they did, they do not send messages to radars telling the radars about themselves: In short, we reiect this technique because it violates another useful principle that can be formulated as follows: THE APPROPRIATE KNOWLEDGE PRINCIPLE: Try to embed in your objects only legitimate knowledge, i.e., knowledge that can be directly accessed by being modeled. the real-world objects that are D L Auxiliary Objects We feel that the above principles should be considered by anyone attempting to develop an object-oriented simulation, as they are critical to insure readable and conceptually clear code. The solution we offer in SWIRL represents one technique that adheres to both principles. After we decompose the domain into a set of basic objects, we create auxiliary objects to handle non-intentional events. Auxiliary objects are full objects in the ROSS sense. However, they do not have real-world correlates. Nevertheless, such objects provide a useful device for handling certain computations that cannot be naturally delegated to real-world objects. We have included two auxiliary objects in SWIRL, the SCHEDULER and the PHYSICIST. The SCHEDULER represents an omniscient, god-like being which, given current information, anticipates non-intentional events in the future and informs the appropriate objects as to their occurrence. The PHYSICIST models non-intentional events involving physical phenomena such as bomb explosions and ecm (electronic counter measures). Although we have now introduced objects which have no real-world correlates, the objects that do have real-world correlates adhere to the above principles. Hence, code for the basic objects remains realistic and transparent. REFERENCES Faught, W. S., Klahr, P. and Martins, G. R. 'An Artificial Intelligence Approach To Large- Scale Simulation." In Proc. 1980 Summer ~ - Computer Simulation Conference, Seattle, 1980, 231-235. Goldberg, A. and Kay, A. "Smalltalk- Instruction Manual.' SSL 76-6, Xerox PARC, Palo Alto, 1976. Goldin, S. E. and Klahr, P. "Learning and Abstraction in Simulation." In Proc. IJCAI-81, Vancouver, 1981, 212-214. Hewitt, C. "Viewing Control Structures as Patterns of Message Passing." Artificial Intelligence 8 (1977), 323-364. Kahn, K. M. "Director Guide." AI Memo 482B, MIT, 1979. Klahr, P. and Faught, W. S. "Knowledge-Based Simulation." Proc. AAAI-80, Palo Alto, 1980, 181-183. Klahr, P., McArthur, D., Narain, S. and Best, E. "SWIRL: Simulating Warfare in the ROSS Language." The Rand Corporation, 1982. McArthur, D. and Klahr, P. "The ROSS Language Manual." N-1854-AF, The Rand Corporation, Santa Monica, 1982. McArthur, D. and Sowizral, H. "An Object- Oriented Language for Constructing Simulations." Proc. IJCAI-81, Vancouver: 1981, 809-814. [II [21 [31 [41 [51 [61 [71 [81 [91 334	1982	1
96	SPEX: A Second-Gcncration Espcrimcnt Iksifg System Yumi Iwasaki and Pctcr Friedland IIcuristic Programming Project, Computer Science Department Stanford University, Stanford, Ca. 94305 Abstract ‘I’hc design of laboratory cxperimcnts is a complex and important scientific task. The MOI,GEN project has been dcvcloping computer jystems for automating Ihc design process in the domain of molecular biology. SPEX is a scLonl-.gcncration syslcm which synlhcsizcs the best ideas of two previous MOI,CEN hierarchical planning systems: stcpwisc rcfincmcnt of skeletal plans ant1 a layered cotlirot structure. It has been tcstcd successfully on several problems in the task domain and promises to serve as a testbcd for future work in explanation, exp;rimeat debugging, and ctnpirical evaluation of different basic design strategies. 1. !nt reduction Expcrimcnt design is the process of choosing an ordered set of laboratory operations to accomplish some given analytical or synthetic goal. This process is one of the fundamental tasks of cxpcrimcntal scientists; it involves large amounts of domain expcrtisc and specialized design heuristics. ‘I’hc design of such expcrimcnt plans has been one of the fundamental research efforts of the MOIGEN project at Stanford. SIJi<X (Skclctal Planner of EXpcrimcnts) is a second-generation experiment design system. It is a synthesis of the best ideas of two previous planning systems, and will serve as a “laboratory” for the empirical testing of design strategies at many levels. SPEX will also be used for MOI-GEN work on experiment verification, optimijlation, and debugging. This paper is a report of the work in progress on SPEX. 1.1. Previous MOLGEN Planning Systems Fricdland developed an experiment planning system using the methodology of hierarchical planning by progressive refmemcnt of skeletal plans [l] [2]. A skeletal plan is a linear scqlrcnce of several abstract steps; actual plans arc gcncratcd by refining each of the abstract steps to use specific tcchniqucs and objects by going down a general-to-specific hierarchy of laboratory operations stored within a knowlcdgc base built by cxpcrimcnt molecular biologists. F’ricdland’s experiment planner chooses a skeletal-plan suitable for the given goal of an cxpcriment and rcfincs each step by choosing the best spccialilation of the laboratory method at each lcvcl of abstraction. Stcfik developed another cxpcrimeht design system [3]. His hierarchical planner first constructs an abstract plan by simple clifl‘crcnce reduction, and then gcncratcs a specific plan from that abstract plan by propagating constraints. It has a multi-layerec! control structure to separate out diffcrcnt lcvcls of decisions to bc made by the planner [4]. The two systems were complcmcntary. Fricdland’s system made cfficicnt USC of large amounts of domain knowledge to produce practical, but not necessarily optimal cxpcrimcnt designs for a large subset ofanalytical tasks in molecular biology. The assumption of near- indcpcndcnce of abstract plan-steps worked well in the great majority of casts. Stcfik’s system took much longer to plan reasonable experiments, but worked bcttcr when plan-steps were highly dcpcndcnt and kept a much richer description of the planning process, this bccausc of the well-designed control structure. Sl+TX was dcvclopcd to synthcsi;lc two fundamental ideas from these planners, namely Fricdtand’s skeletal-plan rcfinctncnt and Stcfik’s multi- layered control structure, in the hope of making further progress in the construction of a design system that would be used by experts. In addition, SPl’X has a grsatly cnhanccd capacity to simulate the changing world st;ltc during an cxpcriment. ‘I‘hc rctnaindcr of this paper describes the laycrcd control structure and the simulation mechanism used by SPEX. I,ikc Friedland’s and Stefik’s systems, the knowledge base ofSPl<X is constructed using the Unit System [5], a frame-based knowledge rcprcscntation systctn dcvelopcd in the MOi,Gl1N project. In SEX, the Unit System is also used to rcprcscnt a tract of the planning process and the changing states of objects in the world. 2. Method 2.1. Layers of Control In order to leave a trace of a planning process, it is necessary to identify the different kinds of operations the planner is expected to perform and to represent the entire process as a scqucncc of such opcrntions and their conscqucnccs. The notion of a multi-layered control structure was introduced and operations at three diffcrcnt tcvcts in lhc planning process were idcntificd and rcprcscntcd within SPBX. The bottom level, called the Iloruain Space by Stcfik, consists of the objects and operators in the task domain, termed lob-s!eps. They are experiment goals, skctctal-plans, and laboratory tcchniqucs. On top of the IJotnain Space exists the De5ig~2 Space, which consists of the vzt-ious planning tasks pcrfortncd by SPEX, for example, the tasks of finding an appropriate skclctal-plan, expanding a skeletal-plan, or refining a technique. Thcsc arc termed plan-sfeps. When such tasks arc cxccutcd, they crcatc or dctete lab-steps in the Domain Space, or create new tasks in the I)csign Space. I:inally, the third layer. the S/rcrfeu Space, consists of scvcral different strntcgics to control the execution of tasks in the Design Space. Diffcrcnt types of decisions are made in the three different spaces. In the Domain Space, decisions are biology-oricntcd. The two major types of decisions are environmental, i.c. whcthcr environmental conditions and structural propcrtics allow a given laboratory technique to be used, and detailed selection, i.e. tl:c process of deciding on the From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. basis of selection heuristics among several tcchniqucs all of which will carry out the specified experimental goal. In the I>csign Space, decisions are more goal-oricntcd. l’hesc decisions rclatc the specific goal of a step in the skeletal plan to the potential utility of a laboratory technique for satisfying that goal. Finally, in the Strategy Space, choices arc made among various design strntcgics, whcthcr to rcfinc in a breadth-first, depth-first, or heuristic manner, for cxamplc. I,ab-steps in the Domain Space arc also rcprcsentcd by units. I.ah- step units initially record the 11orrtain Space results of decisions made in the I>csign Space, but may also record later decisions made in the Domain space. Some of the slots in the prototypical lab-step unit arc shown in l;igurc 3-2. Unit: LAD-STEP 2.2. Modeling World Objects FOLLOWING-STEP: <UNIT> A pointer IO the next iubstep in the plan. ----- Ncithcr previous MOI,GEN design system did a thorough job of monitoring the state of the laboratory cnvironmcnt during the simulated execution of an cxperimcnt plan. nut, the laboratory environment, i.e. the physical conditions and the molecular structures involved in the cxpcrimcnt, undergoes changes in the COUISC of carrying out the experiment. Therefore, for the planning of an experiment design to bc consistently successful, it is csscntial to predict the changes caused by carrying out a step in the plan. SEX simulates those changes, using the part of its knowlcdgc base that contzirls the yotcntial effects of carrying out an laboratory tcchniquc. ‘I’hc prcdictcd states of the world at certain points in the cxpcrimcnt arc used as part of the selection criteria in choosing the appropriate technique for the next step in the plan. ‘I’hcy arc also used to make sure that the preconditions for a chosen technique arc met before the clpplication of the tcchniquc. If any of the preconditions arc not satisfied, a sub-plan to modify the world state is produced and inscrtcd in the original plan. PRECEDING-STEP: <UNIT> A pointer IO the prcccd~ng lab-step in the plan. ----- CREATED-BY: <UNIT> A pointer to 11x plan-step which created thiy lab-step. ----- STATUS : <Sl RING> When a lubstep has just been created, it b given the status ‘jurt-created”, I ater when it is ndcd in cr out, the status is chnngcd to “ruled-In” or “ndcd-out”. 14gurc 3-2: Slots in a lab-step SPEX uses an agenda mechanism to keep track of all the pending tasks. The Strategy Space of SPEX c:msists of simple strntcgics about how to choose the task from the agenda. Currently, a slratcgy is chosen at the beginning of a session by the user and it is used throughout the session. ‘I’here arc three strategies currently available in the system. With Strategy 1. the agenda is used like a qucuc and tasks arc fctchcd in first-in, firsL-out manner. With Strategy 2, the agenda is n~mnally used like a stack and tasks arc fctchcd in first-in, List-out fashion. With 3. Implementation Strategy 3, the plan-steps arc given priorities according to their types. SPEX is implcmcnted in Interlisp, making extensive USC of the IJnit System in order to rcprescnt each operation in diffcrcnt planning spaces. In the Design Space, thcrc arc at present scvcn types of plan- Tasks of a higher priority are executed before any tasks of lower priorities. We are currently cxpcrimenting with various prioritizing schcmcs for the different plan-step types. steps: obtaining the goal of an cxpcrimcnt, choosing a skclctnl plan, expanding a skclctal plan, refining a laboratory tcchniquc, ruling in or out a laboratory tcchniquc, comparing altcrnativc laboratory techniques, and checking the world state to verify that the preconditions for application of a tcchniquc arc satisfied. When a new task needs to hc gencratcd, a plan-step unit is created to represent the task and to 1c;rvc a record of SPEX’s pcrformancc. A plan-step unit has slots containing such information as when the task was created, when it was pcrformcd and what the consequence was. Some of the slots in the prototypical plan-step unit arc shown in Figure 3-l. this plan-step NtXT-PLAN-STEP: <UNIT> 7%u next plan-step J&hed ----- LAST-PLAN-STEP: <UNIT> The previous plan-step fetched 3.1. Simulating the World State At the bcgilllling of ;I planning scss~on, the user is asked to provide a description of the world ol!jccls. ‘l‘his inclirclcs the current physical cnvironmcnt (tcinpcraturc, 1~1-1, etc.) of the cxpcrimcnt and what tic knows about rhc d&led molecular structure of his cxpcrimcntal objects, normally nucleic acid scqucnces. A unit is created to rcprcscnt the initial dcscriplion for each HY~II:! object. When a skclctJ-plan is cxpnndcd to individual steps, units arc crcatcd to rcprcscnt rhc simulated state of each world objjcct bcforc and Acr ,Ipplication of each step. This is done by ulili& (1 simulation information stored in the laboratory-tcchniqucs hierarchy of the knowlcdgc base. l;igurc 3-3 sl~ows some of the slots in the prototypical unit fc)r describing Iluclcic acid structures. ‘I’his unit is called l)N/\-S’l‘liUC’I‘U121~. Its composiiion h,js cvolvcd over scccrnl years of colLrborari\c wwk by scvcral of tlw molecular biologists associalcd with the IL101 .CEN project. It rcprcscnts all of the potential information a scicnlist might wish to supply about a particul‘lr nucleic acid structure. In an actual cxpcrimcnt, a scicnlist would bc able to fill only a few of the olcr 50 slots in the prototypical unit. I:igurc 3-4 shows rlic actual values gi\,cn to the sample slots during the plannin g of an cxperimcnt design by swx. STATUS: <STRING> The .status of this plan-step; either Succeeded’: “postponed, ” or ‘created”. Figure 3-1: Slots in a plan-step 342 Unit: DNA-STRUCTURES ----_ STRANDLDNESS: <STRING> One of: ["HYBRID" "SS" "DS"] ----_ LENGTH: <INTEGFR> [l lOM] MEASUREMCNT UNIT EtASE-PAIRS ----- #-TERMINI: <INTEGER> co 41 ----- #-FORKS: <INTLCER> co 21 ----_ 10POLOGY: <STRING> One of: ["DELIA-FORM" "TtIEIA-I-ORM" "EYC-FORM" "Y-FORM" "LINEAR" "CIRCULAR"] ----- TYI'E: <STRING> One of: ["HYBRID" "RNA" "DNA" ] Figure 3-3: Slots in DNA-STRUCTURE Unit: STRUC-1 STRANF'CDNESS: <STRING> "SS" LENGTH: <INfFGCR> 300 MFASUREMENT UNIT UASE-PAIRS #-TERMINI: <INFEGCR> 2 ----- #-FORKS <INTEGER> 0 ----- TOPOLOGY <STRING> "LINEAR" ----- TYPE: <STRING> "DNA" Figure 3-4: Slots in an instance of DNA-S’I‘RUC’I’URC~ Sl’l+X ha:; been successfully tostctf 01i scv2r::l of 1112 same problems uccd 2s test ca\cs in l*‘ricdlnnd’s thesis [l]. Combining the iclcas 9f skclctnl-plan refinement and multi-la: cl-cd control ~,iruclurc proved uscti~l in keeping the l)omain Space oi iicicncy of I~ri~:d!:~nd’s system, while introducing thL Planning and Stratcu Space flexibility of Stcfik’? syTtcm. ‘l’hc grenlly improved simulillion mcclianism incrcascd the reliability of the decisions made and allowed SPlIX to sl~ggcst, in detail, ways of correcting low-level incompatibilites bctwccn chosen laboratory tccllIliqlles. SI’I<X keeps a clear trace of all the decisions rnaclc during tilt cxpcrimcnt design process. This tract will now bc used for a variety of’ purposes in ongoing MOI~GIIN research. One goal of this rcscarch to add a dctailcd cxplnn&ion capability to SPEX. Iiarlier MOI.GEN cxpclimcnt design bqstcms only provided explanation by listing the domain-specific rules used to make dcc-isions at the Icvcl of laboratory tcchlliyucs. The planning trace and the modular IKILLIK of the diffcrcnt planning spaces will allow this explanation Elcility to bc cxpandcd. ‘l‘hc user will bc able to direct his questioning to domain. design, or strategic motivations. WC envision an initial cxplanntion facility silnilar to that used by ME’CIN [(,I. ‘1’11~ user asks questions using a szt of key words i\S “Why”, “WilCn”, “how”, CIC. WC believe that beside<; promoting the USC of SPEX among expert users for whom full explanations arc a nccersity, the explanation facility will also greatly Lrciiitatc debugging of the knowI~Jge base. A necessary cumplcrncnt to the task of cxpcrimcct &sign is the task of cxpcrimcnt debugging. Initial cspcrimcnt designs prod~~cccl by even the very best of scientists rarely work perfectly the first time. WC arc certain that the same will be bc true of SPEX-produced cxpcrimcnt designs. Given the scqucncc of tcchniqucs employotl in an cxpcrimcnt, a dcbuggcr will compnrc the prcdictcd states of the world bcforc and after each step and the actual world states during the cxpcrimcnt (sL:pplicd by the niolccular biologist ulcr). ‘l’hen, it will point out the steps which might possibly have gone wrong and suggest solutions: altcmstivc techniques or a rcmcdial proccdurc to be licrformcd. If the “buggy” cxpcriment &sign had come from a human scictltist, then the debugging information will enable him to correct his personal “knowledge bitsc.” In a similar manner, wc bclicvt: that the Lomparison of actual laboratory results to the previously predicted results should allow the automatic improvcmcnt of the knowlcdgc base in the cast of a Sl’I<X-gcncratcd cxpcrimcnt design. WC do not illltiCipXC trujor difficulties in building the cxpcrimcnl dcbuggcr gikcn the existing mcchanicms of SPIIX. In summary, SI’l;X rcprcscnts a synthesis of scvcr31 mcthodologics for dcqign, nnmcly skclctnl-plan rcfincmcnt and the multi-layered control structure. It is a framework for a gcn~‘~,al-pul’posc design testing and tlcbugging system, which cm bc easily tailored to do planning in any specific npplication domain. ‘I’hcrc arc no nlolcculi~r-biology specific mcch,lnisms inhcrcnt in SI’IIX; nil of th: domain-spcci~ic kn()\vlcL]gc in ill tllc :lssocinled knowlcdgc b:jsc. SI’liX C;\Il ah) h ~lsCd to test clir[cFctlt l>itSiC &sign ~ir,ltCgiC!: by LhC i~~!~)~ciIlCtltiltioll Oi‘ In~llly fiddilioll;ll stra[c$ics iti its Str,ltcsy Sp:lcC. WC bclic\c tI13t IllC lKSt way to dctcrminc the efficacy of tlic Illittly diffcl cnt potential ~li~dtcgics is cmpiric;J, and Sl+tX will bc useful as a hhratory for thCsC cxperimcntS. This work is p:lrt of the MOI~CiFN project, n joint rc$car.ch effort among Ihc I):partmcnts o!’ Computer Scicncc, Medicine, and iGochcniistry al Stanford Uuivcrsity. '1'1~ mearch has been \upported ~ntlcr NSl: gr#\rlt MC%&16247. Cornl~L1ti~tiollal rcsourccs hn~c IXCII provided by the SUMEX-AIM Nation,\] lGomcdica1 Rcscarch I<csourcc, NI I I grant RR-00785-08, and by the Dcpal tlncnt of Computer Science. WC wish to th,lnk our many cnthusinstic MO1 6 l:N collA)orlltors for their Ast,\ncc in tllis work. WC arc cspeciallq grateful to I<cnL Ihch and I-arty Kcdcs for pro\ iding tilC ~?lOlKtlli~~ biology c\pcrLisc ncccbsary to test Sl’I:X and tu f2ruic l~uch;tlli111 and Mike Gcnzscrcth for advice on the artificial intclligcllcc methotiolo~ics cmpioycd. 343 References 1. Fricdiand, P.E., Ktrowledge- Rased l:icperhen~ Desigtl In Mol~~~~!ur Gerre[ics, Phl1 dissertation, Stanford University, OctobCr 1979. 2. Fricdland, P.E., “Knowlcdgc-1jascd Expcrimcnt Design in Molccularc Genetics ” LIC’AI-754 ‘The International Joint Confcrcnce on hrtif&al Intelligence, 1979, pp. 285-287. 3. Stcfik, M.J., PIunning wilh Conslrairtts, Phi> dissertation, Stanford University, January 1980. 4. Stcfik, M.J., “Planning and MC&Planning,” HPP-mcmo HPP- 80-13, Stanford University Heuristic Programming Project, 1380. 5. Smith, KG., Friedland, P.R., “Unit Package User’s Guide,” HPP-memo :wP-80-28, Stzmford University Heuristic Programming Project, 1980. 6. Scott, AC., Clanccy, W.J., Davis, R., Shortliffc, E.H., “Explanation Capahilitics of I’roduction-Rascd Consultation System,” Anwican fournal of ~orqwilaliot~al Linguislics, 1979, .	1982	10
97	CIRCUMSCRIPTION IMPLIES PREDICATE COMPLETION (SOMETIMES) Raymond Reiter Department of Computer Science Rutgers University New Brunswick, N. J. 08903 ABSTRACT Predicate completion is an approach to closed world reasoning which assumes that the given sufficient conditions on a predicate are also necessary. Circumscription is a formal device characterizing minimal reasoning i.e. reasoning in minimal models, and is realized by an axiom schema. The basic result of this paper is that for first order theories which are Horn in a predicate P, the circumscription of P logically implies P's completion axiom. Predicate completion [Clark 1978, Kowalski 19781 is a device for "closing off" a first order representation. This concept stems from the observation that frequently a world descrip- tion provides sufficient, but not necessary, conditions on one or more of its predicates and hence is an incomplete description of that world. In reasoning about such worlds, one often appeals to a convention of common sense reasoning which sanctions the assumption - the so-called closed world assumption [Reiter 19781 - that the infor- mation given about a certain predicate is all and only the relevant information about that predicate. Clark interprets this assumption formally as the assumption that the sufficient conditions on the predicate, which are explicitly given by the world description, are also neces- sary. The idea is best illustrated by an example, so consider the following simple blocks world description: A and B are distinct blocks. A is on the table. B is on A. (1) These statements translate naturally into the following first order theory with equality, assuming the availability of general knowledge to the effect that blocks cannot be tables: BLOCK (A) BLOCK (B) 0~ (A,TABLE) ON (B,A) (2) A#B A# TABLE B # TABLE Notice that we cannot, from (2), prove that nothing is on B, i.e., (2) /+ (x> -ON(x,B), yet there is a common sense convention about the description (1) which should admit this conclu- sion. This convention holds that, roughly speaking, (1) is a description of all and only the relevant information about this world. To see how Clark understands this convention, consid- er the formulae (x) .x = A V x = B ZJ BLOCK (x) (3) (xy).x = A 6 y = TABLE V x * B d y = A 3 ON(x,y) which are equivalent, respectively, to the facts about the predicate BLOCK, and the predicate ON in (2). These can be read as "if halves", or sufficient conditions, of the predicates BLOCK and ON. Clark identifies the closed world as- sumption with the assumption that these sufficient conditions are also necessary. This assumption can be made explicit by augmenting the representa- tion (2) by the "only if halves" or necessary conditions, of BLOCK and ON: (x). BLOCK(x) 1 x=A V x=B (xy). ON (x,y) 3 x=A & y=TABLE V x=B & y=A Clark refers to these "only if" formulae as the completions of the predicates BLOCK and ON respectively. It now follows that the first order representation (1) under the closed world assumption is (x). BLOCK(x) : x=A V x=B (xy). ON(x,y) = x=A & y=TABLE V x=B & y=A AfB A # TABLE B # TABLE From this theory we can prove that nothing is on B - (x)-ON(x,B) - a fact which was not derivable from the original theory (2). Circumscription [McCarthy 19801 is a dif- ferent approach to the problem of "closing off" a first order representation. McCarthy's intu- itions about the closed world assumption are es- sentially semantic. For him, those statements derivable from a first order theory T under the closed world assumption about a predicate P are just the statements true in all models of T which are minimal with respect to P. Roughly speaking, these are models in which P's extension is minimal. McCarthy forces the consideration of only such models by augmenting T with the fol- lowing axiom schema, called the circumscription of P in T: T(t) & b> .4(x) = P(x)] = (xl. P(x) = @(xl From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. Here, if P is an n-ary predicate, then $ is an n-ary predicate parameter. T(4) is the conjunction of the formulae of T with each occur- rence of P replaced by 9. Reasoning about the theory T under the closed world assumption about P is formally identified with first order deduc- tions from the theory T together with this axiom schema. This enlarged theory, denoted by CLOSUREp(T), is called the closure of T with respect to P. Typically, the way this schema is used is to "guess" a suitable instance of 4, one which permits the derivation of something use- ful. a fact which is not derivable from the original theory T. Notice that in order to make this work, a judicious choice of the predicate parameter 4, namely (5), was required. Notice also that this choice of $ is precisely the antecedent of the "if half" (3) of ON and that, by (6), the "only if half" - the completion of ON - is derivable from the closure of T with respect to ON. For this example, circumscription is at least as powerful as predicate completion. To see how this all works in practice, con- sider the blocks world theory (2), which we shall denote by T. To close T with repect to ON, augment T with the circumscription schema X,Y> = ON(x,y)l ON(x,y) = $(x,Y) (4) Here 4 is a 2-place predicate parameter. Intuitively, this schema says that if 4 is a predicate satisfying the same axioms in T as does ON, and if 4's extension is a subset of ON's, then ON's extension is a subset of 4's, i.e., ON has the minimal extension of all predicates satisfying the same axioms as ON. In fact, this example is an instance of a large class of first order theories for which cir- cumscription implies predicate completion. Let T be a first order theory in clausal form (so that existential quantifiers have been eliminated in favour of Skolem functions, all variables are universally quantified, and each formula of T is a disjunct of literals). If P is a predicate symbol occurring in some clause of T, then T is said to be Horn in P iff every clause of T con- tains at most one positive literal in the pred- icate P. Notice that the definition allows any number of positive literals in the clauses of T so long as their predicates are distinct from P. Any such theory T may be partitioned into two disjoint sets To see how one might reason with the theory CLOSUREON( consider the following choice of the parameter + in the schema (4): Tp: those clauses of T containing exactly one positive literal in P, and T-Tp: those clauses of T containing no positive $(x,y) q x=A & y=TABLE V x=B 6 y=A (5) (but possibly negative) literals in P. Then T(4) is BLOCK(A) & BLOCK(B) & [A=A & TABLE=TABLE v A=B 6r TABLE=A] 6 CB=A 6 A=TABLE v B=B & A=AI & A# B & A# TABLE &B # TABLE so that, for this choice of $, CLOSUREON I-- T($) It is also easy to see that, for this choice of CLOSUREON j- (xy) .$(x,y> = ON(x,y) . Thus, the antecedent of (4) is provable, whence CLOSUREON t (XY> .ON(x,y) 3 $(x,y). 1. * e. Clark (1978) provides a simple effective procedure for transforming a set of clauses of the form Tp into a single, logically equivalent formula of the form (x). A(x) = P(x). The con- verse of this formula, namely (x). P(x) = A(x), Clark calls the completion axiom for the predicate P, and he argues that augmenting T with P's com- pletion axiom is the appropriate formalization of the notion of "closing off" a theory with re- spect to P. Our basic result relates this notion of closure with McCarthy's, as follows: Theorem Let T be a first order theory in clausal form, Horn in the predicate P.Let (x).P(x) 3 A(x) be P's completion axiom. Then CLOSUREp(T) I- (x). P(x) 1 A(x) CLOSURE (T) I- (xy).ON(x,y) ' x=A & y%BLE V x=B & y=A (6) i.e. the only instances of ON are (A,TABLE) and @,A ) l It is now a simple matter to show that nothing is on B, i.e. CLOSUREON t (x).-ON(x,B) i.e. P's completion axiom is derivable by cir- cumscription. Discussion Circumscription and predicate completion are two seemingly different approaches to the formal- ization of certain forms of common sense reason- ing, a problem which has recently become of major 419 concern in Artificial Intelligence (see e.g. [AI 19801). That circumscription subsumes predicate completion for a wide class of first order theories is thus of some theoretical interest. Moreover, circumscription is a new formalism, one whose properties are little understood. Predicate completion on the other hand, has a solid intuitive foundation, namely, assume that the given sufficient conditions on a predicate are also necessary. The fact that predicate com- pletion is at least sometimes implied by cir- cumscription lends support to the hypothesis that circumscription is an appropriate formalization of the notion of closing off a first order represen- tation. Finally, the theorem has computational im- port. Notice that in order to reason with McCarthy's circumscription schema it is first necessary to determine a suitable instance of the predicate parameter Cp. This is the central computational problem with circumscription. With- out a mechanism for determining "good $'s", one cannot feasibly use the circumscription schema in reasoning about the closure of a represen- tation. This problem of determining useful 4's is very like that of determining suitable pre- dicates on which to perform induction in, say, number theory. Number theory provides an induc- tion axiom schema, but no rules for instantiating this schema in order to derive interesting theorems. In this respect, the circumscription schema acts like an induction schema. Now the above theorem provides a useful heuristic for computing with the closure of first order Horn theories. For we know a priori, with- out having to guess a $ at all, at least one non trivial consequence of the circumscription schema, namely the completion axiom. Clearly, one should first try reasoning with this axiom before invok- ing the full power of circumscription by "gues- sing $‘s". REFERENCES AI (1980). Special issue on non-monotonic logic, Artificial Intelligence 13 (1,2), April. Clark, K. L. (1978). Negation as failure, in Logic and Data Bases, H. Gallaire and J. Minker (eds.), Plenum Press, NY, 293-322. Kowalski, R. (1978). Logic for data description, in Logic and Data Bases, H. Gallaire and J. Minker (eds.), Plenum Press, NY, 77-103. McCarthy, J. (1980). Circumscription - a form of non-monotonic reasoning, Artificial Intelligence 13, 27-39. Reiter, R. (1978). On closed world data bases, in Logic and Data Bases, H. Gallaire and J. Minker (eds.), Plenum Press, NY,55-76. 420	1982	100
98	Monitors as Responses to Questions: Determining Competence Eric Mays Department of Computer and Information Science Moore School of Electrical Engineering/D2 University of Pennsylvania Philadelphia, Pa. 19104 ABSTRACT This paper discusses the application of a propositional temporal logic to determining the competence of a monitor offer as an extended response by a question-answering system. Determining monitor competence involves reasoning about the possibility of some future state given a description of the current state and possible transitions. I INTRODUCTION The offer of a monitor as a response becomes possible when the system views the knowledge base (KB) as a dynamic rather than a static entity. That is, in addition to answering a user's question on the basis of the information the system currently contains, a system with a dynamic view can offer to monitor for additional relevant information which it will provide to the user if and when it learns of it. Such additional information could be about some possible future event or some previous event about which the system's knowledge is currently incomplete. In the following question-answer pairs, Q-Al illustrates a monitor for some possible future event. The pair Q-A2 is an example of a monitor for some additional information about a previous event. Responses such as Q-A2 require reasoning that some event, of which knowledge regarding its outcome would enable us to answer the question, has taken place. At some point in the future we will learn of its outcome, when we will then answer. Ql: Al: 42: A2: Has John registered for CSEllO? No, shall I let you know if he does? Did John pass CSEllO? I don't know yet. The semester has ended, but Prof. Tardy has not turned in his grades. Shall I let you know when I find out? In order to offer monitors as extended responses the system must be competent to offer to monitor for only those events which might possibly occur or, if the system has incomplete knowledge * This work i s partially supported by a grant from the Na tional Science foundation, NSF-MCS 81-O 7290. of some event that has occured, only that additional information it may learn of.* This requires some notion of what events are possible or what additional information may be acquired given the current state of the knowledge base. For example, ignorance of the stages through which undergraduates proceed in the university would leave a system attempting to offer monitors unable to discriminate between the following two cases. Ql: Is John a sophomore? Al: No, he's a freshman. Shall I let you know when he becomes a sophomore? Q2: Is Mary a sophomore? A2: No, she's a junior. Shall I let you know when she becomes a sophomore? The remainder of this paper is concerned with determining monitor competence with regard to possible future events. We leave open for now the question of competence for those monitors that require reasoning about incomplete knowledge of some previous event. II REPRESENTATION Temporal logic [3] is a modal logic for reasoning about the relative possibility of some state to some other state, where the relative possibility is with respect to time or sequences of events. (In contrast to, for example, relative possibility with respect to obligation or belief.) Although one might develop a suitable first order theory to deal with the problems discussed here, it seems worthwhile to study this problem within the framework of temporal logic for reasons of conceptual clarity. Restriction to the propositional setting enables us to concentrate on those issues involved with reasoning about possible change. We model the evolution of a KB in a propositional temporal logic. The future fragment is a unified branching temporal logic [l] which * In either case it must be able to identify those future conditions which are relevant to the user's intentions. The discussion here will be limited to determination of monitor competence. for a brief discussion on relevance. See [3] 421 From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. makes it possible to describe properties on some or all futures. By merging the existential operators with the universal operators, a linear temporal logic is formed for the past fragment (i.e. AXp <-> EXp). A. Syntax Formulas are composed from the symbols, - A set P of atomic propositions. - Boolean connectives: v, -. - Temporal operators: AX (every next), EX (some next), AG (every always), EG (some always), AF (every eventually), EF (some eventually), L (immediately past), P (sometime past), H (always past). using the rules, - If p e P, then p is a formula - If p and q are formulas, then (-p), (p v q) are formulas. - If m is a temporal operator and p is a formula, then (m)p is a formula. Parenthesis will occasionally be omitted, and &, ->, <-> used as abbreviations. B. Semantics A structure T is a triple (S, TT, R) where, - S is a set of states. - TT:(S -> 2-P) is an assignment of atomic propositi0t-Z to states. - R C (S x S) is an accessibility relation on s.- Each state is required to have at least one successor and exactly one predecessor, As (Et (sRt) & E!t (tRs)). Define an s-branch b = (..., s(-1), s=s(O), s(l), . ..) such that s(i)Rs(i+l). The satisfaction of a formula p at a node s in a structure T, <T, s> I= p, is defined as follows: (Note: "err denotes set inclusion.) <T, s> I= p <T, s> I= -p <T, s> I= p v <T, s> I= AGp (p is true at <T, s> I= AFp (p is true at <T, s> I= AXp (p is true at <T, s> I= EGp (P is true at <T, s> I= EFp (p is true at <T, s> I= EXp (p is true at <T, s> I= Hp (p is true at every time of the past) iff peTT(s), for p a proposition iff not <T,s>l=p q iff <T,s>l=p or <T,s>]=q iff Ab At ((teb & t>s) -> <T,t>l=pT every time of every future) iff Ab Et (teb & t>s & <T,t>l=p) some time of every future) iff At (sRt -> <T,t>l=p) every immediate future) iff Eb At ((teb & t>s) -> <T,t>l=p) every time of some future) iff Eb Et (teb & tks & <T,t>l=p) some time of some future) iff Et (sRt & <T,t>l=p) some immediate future) iff Ab At ((teb 6 t(s) -> <T,t>l=p) <T, s> I= Pp iff Ab Et (teb & t<s & <T,t>l=p) (p is true at some time of the past) <T, s> I= Lp iff Et (tRs & <T,t>l=p) (p is true at the immediate past) A formula p is valid if for every structure T and every node s in T, <T, s> I= p. C. Axioms In the following axioms Dl-3, Al-4, El-4 and rules Rl-3 form a complete deductive system for the future fragment [l]. Similarly, D4-5, Pl-4, Rl, R2, R4 are complete for the past fragment. The idea of Ul and U2 is that the relationship between the past and the future may be described locally. Using the induction axioms, we can derive the following theorems, which are a more conventional form (as in 131): EF(Hp) -> Hp P(AGp) -> AGp See [1] for a list of many other useful theorems for the future fragment. Dl) AFp <-> -EG-p D2) EFp <-> -AG-p D3) AXp <-> -EX-p D4) Pp <-> -H-p DS) Lp <-> -L-p Al) AG(p -> q) -> (AGp -> AGq) A2) AX(p -> q) -> (AXp -> AXq) A3) AGp -> p & AXp & AX(AGp) A4) AG(p -> AXp) -> (p -> AGp) El) AG(p -> q) -> (EGp -> EGq) E2) EGp -> p & EXp & EX(EGp) E3) AGp -> EGp E4) AG(p -> EXp) -> (p -> EGp) Pl) H(p -> q) -> (Hp -> Hq) P2) L(p -> q> -> (Lp -> Lq) P3) Hp -> p & Lp & L(Hp) P4) H(p -> Lp) -> (p -> Ilp) Ul) L(AXp) -> p U2) p -> WLP) Rl) If p is a tautology, then I- p. R2) If I- p and I- (p -> q), then I- q. R3) If I- p, then I- AGp. R4) If I- p, then I- Up III EXAMPLE Consider as an example representing that portion of a university KB dealing with students passing and registering for courses. Let the propositional variables Q and R mean "student has passed course" and "student is registered for course", respectively. One might have the following non-logical axioms: 1) (AG)(Q -> (AX)Q) - once course it remains so a student has passed a 422 2) (AG)((-Q & -R) -> (EX)R) - if a student has not passed a course and is not registered then it is next possible that s/he is registered 3) (AG)(R -> (EX)Q) - if a student is registered for a course then it is next possible that s/he has passed 4) (AG)(R -> (EX)(-Q & -R)) - if a student is registered for a course then it is next possible that s/he has not passed and is not registered 5) (AG)(Q -> -R) - if a student has passed a course s/he is not registered for it 6) (AG)(R -> -Q> - if a student is registered for a course s/he has not passed it (equivalent to 5) Given the following question, Is John registered for CSEllO?, there are three possibilities depending on the present state of the KB: 1) John is not registered (-R), but he has passed (Q>* If we consider John registering for CSEllO as a possible monitor, it would be ruled out on the basis that it is provable that John cannot register for CSEllO. Specifically, from Q and axioms 1 and 5, it is provable that -(EF)R. It would thereFore be incompetent to offer to monitor for that condition. 2) John is not registered (-R), but he has not passed (-Q). In this case we could offer to monitor for John registering for CSEllO, since (EF)R is derivable from axiom 2. 3) John is registered for (R), hence he has not passed (-Q). One could competently offer to monitor for any of the following: a) John no longer registered for CSEllO; (EF)-R b) John passed CSEllO; (EF)Q c) John registered for CSEllO again; (EF)(-R & (EX)R) This last case is interesting in that it can be viewed as a monitor for -R whose action is to set a monitor for R (whose aciton is to inform the user of R). Also, one may wish to include monitors that are responsible for determining whether or not some monitor that is set can still be satisfied. That is, just because something was possible once does not imply that it will always be possible. The user should probably be informed when a situation s/he may still be expecting (because a monitor was offered) can no longer occur. For example, if the system has offered to inform the user if John registers for CSEllO, then s/he should be informed if John receives advance placement credit and can no longer register. The following set of axioms illustrate the use of the past operators. Note that axiom 1 from the above set may be eliminated, due to the ability of the past operators to "look back". 1) (AG)((-(P)Q & -RI -> (EX)R) 2) (AG)(R -> OWQ) 3) (AG)(R -> (EX)(-Q 6( -R)) 4) (AG)((P)Q -> -RI 5) (AG)(R -> -(P)Q) A more important use of the past operators the ability to express conditions that depend on sequences of events. Thus, expressing the condition that in order for a student to registe for a course s/he must not have registered for i twice before (say, because s/he dropped out or failed), requires a formula of the following for (AG)(-(P)(R & (L)(P)R) & -(P)Q & -R -> (EX)R) is r t m: IV CONCLUSION A simple theorem prover based on the tableau method has been implemented for the propositional branching time temporal logic as described in [l]. Current investigations are aimed towards formulating a quantified temporal logic, as well as the complicated issues involved in increasing efficiency of making deductions. This effort is part of a larger, more general attempt to provide extended responses in question-answering systems (41. ,4 final comment as to general structure of this enterprise. One could conceivably develop a suitable first order theory to deal with the problems discussed here. It seems worthwhile, however, to study this problem within the framework of temporal logic for reasons of conceptual clarity. Restriction to the propositional setting enables us to deal strictly with those issues involved with reasoning about possible change. Also, we may be able to gain some insight into a reasonable decision procedure. ACKNOWLEDGEMENT Scott Weinstein, Aravind Joshi, Bonnie Webber, Sitaram Lanka, and Kathy McCoy have provided valuable discussion and/or comments. REFERENCES [l] M. Ben-Ari, Z. Manna, and A. Pneuli, "The Temporal Logic of Branching Time", Eighth Annual ACM Symposium on Principles of Programming Languages, Williamsburg, Va., January 1981. [2] E. Mays, A. Joshi, and B. Webber, "Taking the Initiative in Natural Language Data Base Interactions: Monitoring as Response", Proceedings of ECAI 82, Orsay, France, July 1982. [3] N. Rescher and A. Urquhart, Temporal Logic, Springer-Verlag, New York, 1971. [4] B. Webber, A. Joshi, E. Mays, and K. McKeown, "Extended Natural Language Data Base Interactions", to appear.	1982	101
99	A SYSTEMATIC APPROACH TO CONTINUOUS GRAPH LABELING WITH APPLICATION TO COMPUTER VISION* M. D. Diamond, N. Narasimhamurthi, and S. Ganapathy Department of Electrical and Computer Engineering University of Michigan Ann Arbor, MI 48109 ABSTRACT The discrete and continuous graph labeling problem are discussed. A basis for the continuous graph labeling problem is presented, in which an explicit connection between the discrete and continuous problems is made. The need for this basis is argued by noting conditions which must be satisfied before solutions can be pursued in a formal manner. Several cooperative solution algo- rithms based on the proposed formulation and results of the application of these algorithms to the problem of extracting line drawings are presented. I XHECIONTINlrnrlrGRAPHLAREXWGJzQRLEM A graph labeling problem is one in which a unique label, A from a set A of possible labels must be assigned to each vertex of a graph G = (V,E). The assignment must be performed given information about the relation- ship between labels on adjacent vertices and incomplete local information about the correct label at each vertex. In a discrete graph labeling problem [ 1,2,3], the local information consists of a subset, & s A, of the label set associated with vertex vi E V, from which the correct label for each vertex must be chosen. The contextual information consists of binary relations Ru s Axh, referred to as constraint relations, assigned to each edge vivj E E. The function of the constraint relations is to make explicit which labels can co-occur on adjacent vertices. The graph, label set, and constraint relations together form a constraint network [2,5]. An (unambi- guous) labeling is a mapping which assigns a unique label h E A to each vertex of the graph. A labeling is con- sistent if none of the constraint relations is violated, that is, if label h is assigned to vertex Vi and label h’ is assigned to vertex vj then the pair (h,X’) is in the con- straint relation Rii for the edge ViVj E E. Given initial labeling information, several search techniques have been developed which can be used to derive consistent labelings. The original backtracking search described by Waltz [ 11 was later implemented in parallel by Rosenfeld et al. [6], resulting in the discrete relaxation operator. At the same time a continuous analogue, the continuous graph labeling problem was * This work was supported in part by the Robotics Research Laboratory, and in part by the Ultrasonics Im- aging Laboratory both in the Department of Electrical and Computer Engineering, University of Michigan. proposed, as well as a continuous relaxation algorithm for its solution, and since then several other relaxation algorithms have been proposed [7,8]. In a continuous graph labeling problem, the initial information consists of strength measures or figures of merit, pi (Aj), given for each label Aj E A on each vertex Vi E I! The strength measures are assumed generated by feature detectors which are making observations in the presence of noise. They usually take onvalues in the range [O,l], a 0 indicating no response, and a 1 indicat- ing a strong response. The contextual information, which is represented in terms of constraint relations for the discrete graph labeling problem, are replace by measures of compatibility, usually taking values in the range [-1,1] or [O,l], which serve to indicate how likely the pairs of labels are to co-occur on adjacent vertices. Several problems have resulted in the extension of the graph labeling problem from the discrete to continu- ous case. In the discrete case the presence or absence of a pair in a constraint relation can be determined with certainty depending on what labelings are to be con- sidered consistent. In the continuous case, however, there is apparently no formal means to assign specific numeric values to the compatibility coefficients, partic- ularly for shades of compatibility between “impossible” and “very likely”, although several heuristic techniques have been proposed [7,9,10]. Furthermore, with respect to a constraint network, the concept of consistency is well defined, The objective the continuous relaxation labeling processes has often been stated to be that of improving consistency, however, the definition for con- sistency has not been given explicitly. This latter issue is circumvented in several of the optimization approaches which have been proposed [11,12,13], where the an objective function, defined in terms of the compa- tibility coefficients and the initial strength measures is given. However, because of the dependence of the objec- tive functions on the compatibility coefficients, and because no real understanding of the role which these coefficients play yet exists, it is often difficult to describe the significance of these approaches in terms of what is being achieved in solving the problem. In an alternate approach to the continuous graph labeling problem [14] an attempt has been made to maintain the characteristics of the original problem while allowing for more systematic approaches toward a 50 From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. solution. It is felt that solutions to the reformulated problem will be more useful because it will be easier to relate the results of the solution algorithm to what is being achieved in the problem domain. In order to develop this approach, we review the characteristics of the solutions to the graph labeling problem which have appeared so far (refer to Fig. 1). The inputs to the process are the initial strength measures ipi'( i=l,s..,W 0 j=l ,...,mj which can be represented by an 7~ x’I)?. dimensional vector: F = (p 10 (AI), p : @2), . . . , p:(&)) E Pm. Since the selection of a particular label at a given vertex is related to the label selections made at other (not necessarily adjacent) vertices, information about the label selection at that vertex is contained in the initial labeling values distributed over the extent of the net- work. The function of the global enhancement process, g, is to accumulate this evidence into the labeling values at the given vertex. The output vector: is used by a process of local maxima selection [ 151, s , to choose a labeling: r; = (A,, 43 . . . I &J, where & is the label assigned to vertex ui. Thus g is a function, g:Rnm + Rnm and s is a function is s:Rnm + C,,(A), where C,(A) is the set of possible label- ings. The hope is that labeling resulting from the pro- cess s (e @)) is an improvement over the labeling result- ing from direct local maxima selections @). If a numerical solution is to be sought for this prob- lem, then a formal definition must be given to the con- cept of an improved labeling. In previous work, particu- larly with respect to computer vision, improvements were rated subjectively, or in the case of an experiment global enhancement process, g Fig. 1: Function of the global enhancement process: X represents an improved labeling with respect to 2. local -+ g+ maxima I-! -x selection, S where the solution was known in advance, by the number of misclassified vertices. In our formulation this issue is resolved by assuming that the problem domain specifies an underlying constraint network, or can be modeled to do so. The objective is then to use the initial information to choose a labeling that is (a) consistent with respect to this constraint network, and (b) which optimizes a prespecified objective function. In this extension from the discrete to the continuous graph labeling problem, the constraint relations remain intact. We are currently investigating optimal solutions to this formulation of the graph labeling problem based on a maaGnu~-sum decision rule, that is, the rule is to choose a consistent labeling such that sum of the initial labeling values is maximal. A solution to this problem could be extended in a straightforward manner to cer- tain well established decision rules such as,is found, for example, in nearest neighbor classification. Though the decision rule serves to make explicit what is meant by an improved labeling, it is defined glo- bally. The problem remains to implement it in terms of a cooperative process. The concept of a cooperative process, although not well defined, can be characterized in terms certain general properties [U1179]. Our research is into algorithms which exhibit certain of these properties, such as locality, and simplicity. In an aptimul solution, the labeling algorithm must, further- more, perform the label selection in accordance with the given decision rule. Other important issues, such as speed of convergence are also being addressed. Two approaches which have some of these properties are demonstrated in the following section. The first is an heuristic approach based on dynamic programming [ 141 which converges very rapidly and with good results, but does not guarantee a consistent labeling. The second approach is based on linear programming. Details on the latter algorithm will be presented at a later date. III -- In this section, we demonstrate the application of the two approaches discussed above to the problem of extracting polygon approximations of the outlines of objects in a scene. The experiments described here are based on the reconstruction of simple closed curves (Fig. 2) when noise has been added to the initial labeling values. The graph used in this experiment is a 16.by 16 ras- ter. Each vertex is represented by a pixel, and is adja- cent to its eight immediate neighbors. The associated label set is shown in Fig. 3. A pair of labels on adjacent pixels are consistent if an outgoing line segment is not broken across a common border or corner, and incon- sistent otherwise. Examples of consistent pairs of labels are given in Fig. 4, and examples of inconsistent pairs of labels are given in Fig. 5. *In terms of decision theory, every consistent label- ing constitutes a class and the input vector p is a point in an 7~x9~~ dimensional feature space. 51 Fig. 2: Initial labeling lzl lzl til la vi III 6l El q El q Ic El q lil III Gl w El q q Fig. 4: Examples of locally consistent label pairs. Uniformly independently distributed noise was added to the labeling values at each pixel resulting in the labeling, by local maxima selection, shown in Fig. 6. The two cooperative algorithms were applied to the ini- tial labeling in attempt to reconstruct the original curves. The first is the dynamic programming approach with data moving along the eight major directions of the raster (two horizontal, two vertical, and four diagonal). The second is the algorithm based on a linear program- ming approach. The performance of these algorithms are presented in Fig. ‘7 and Fig. 6, which show the result- ing labeling (by choosing the label with greatest strength at each pixel) after 2 and 4 iterations. The dynamic pro- gramming approach reaches a fixed point after 2 itera- tions, however, the result is not a consistent labeling. The linear programming algorithm reconstructs the ori- ginal labeling after six iterations. IV IIMXzWB Our interest here has been to restate the continu- ous graph labeling problem in a manner which allows for a systematic approachs to a solution. The formulation which we have presented amounts to the classification of Fig. 3: Label set for line drawing description. Fig. 5: Examples of inconsistent label pairs. 52 1 I I VI IU I I I III nwl I DI I Fig. 6: Initial labeling plus noise. I i i ii L-44 i ii i Fig. 8a: Output of the linear programming algorithm after two iterations. Fig. 7: Output of the dynamic programming algorithm after two iterations. Note: the algorithm has reached a fixed point. Fig. 8b: Output of the linear programming algorithm after four iterations. 53 consistent labelings according to a prespecified decision rule. As with previous approaches, consistency is defined on a local basis to make sense with respect to a [101 particular problem. For example, if the objective is to extract continuous curves as in the experiment described above, consistency is maintained between pairs of labels when the scene events they represent do 1111 not allow for broken lines. The global nature of the deci- sion rule leads to a more intuitive description of what the techniques accomplishes with respect to the original [12 problem. However, as a consequence, the problem of implementing this rule on a local basis arises. Two approaches to the reformulated problem have been demonstrated above. Our present feeling is that a [ 13 linear programming approach should yield an optimal solution to the continuous graph labeling problem based on a maximum-sum decision rule. However, the restric- tion that the algorithm must be implemented in a local manner has led to some theoretical problems, such as resolving cycling under degeneracy which remain to be [l-+1 solved. Our investigation into these problems is continu- ing. Obviously, the value of this approach and any tech- niques which may be derived from it will depend on whether or not real world applications can be modeled in [151 such a manner so that the absolute consistency between pairs of labels is meaningful. We hope to demonstrate this in at least one problem, deriving line drawings from real world scenes, in forthcoming results. PI PI L31 WI bl PI PI ml PI WI REFERENCES D. L. Waltz, “Generating semantic descriptions from drawings of scenes with shadows,” Technical [171 Report A1271, M.I.T., 1972. U. Montanari, “Networks of constraints: funda- mental properties and application to picture pro- cessing,” Information Sciences, vol. 7, pp. 95- 132, 1974. R. M. Haralick and L. G. Shapiro, “The consistent labeling problem: part I,” IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-1, pp. 173-184, 1979. R. M. Haralick and L. G. Shapiro, The consistent labeling problem: part II,” IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-2, pp. 193-203, 1980. A. K. Mackworth, “Consistency in networks of relations,” Artificial Intelligence, vol. 8, pp. 99- 118, 1977. A. Rosenfeld, R. A. Hummel, and S. W. Zucker, “Scene labeling by relaxation operations,” IEEE D-an-s. Syst., Man, Cybern., vol. SMCQ, pp. 420- 433. S. Peleg, “A new probabilistic relaxation scheme ,” IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-2, pp. 362-369, 1980. R. L. Kirby, “A product rule relaxation method,” Comput. CA-aphics Image Processing, vol. 12, pp. 158-189, 1980. S. Peleg and A. Rosenfeld, “Determining compati- bility coefficients for curve enhancement relaxa- tion processes,” IEEE Trans. Syst., Man, Cybem., vol. SMC-8, pp 548-555, 1978. H. Yamamoto, “A method of deriving compatibil- ity coefficients for relaxation operators,” Com- put. Graphics Image Processing, vol. 10, pp. 256- 271, 1978. S. IJllman, “Relaxation and constrained optimiza- tion by local processes,” Comput. Graphics Image Processing, vol. 11, pp. 115-125, 1979. R. A. Hummel, and S. W. Zucker, “On the founda- tions of relaxation labeling processes,” 7+8@7, Dept. of Elect. Eng., McGill University, Montreal, Quebec, Canada. 0. D. P’augeras, and M. Berthod, “Improving con- sistency and reducing ambiguity in stochastic labeling: an optimization approach,” IEEE Trans. Pattern Anal. Machine Intell,, vol. PAMI-3, pp. 412-424, 1981. M. D. Diamond, and S. Ganapathy, “Cooperative solutions to the graph labeling problems,” Proc. PRIP 82 Conference on Patt. Recognition and Image Processings, June, 1982, to appear. S. W. Zucker, Y. G. Leclerc, and J. L. Mohammed, “Continuous relaxation and local maxima selec- tion: conditions for equivalence IEEE Trans. Pat- tern Anal. Machine Intell., vol. PAMI-3, pp. 1:7- 127, 1981. L. S. Davis, and A. Rosenfeld, “Cooperating processes for low-level vision: a survey,” TR-123, Dept. of Computer Science, University of Texas, Austin, 1980. E. C. Freuder, “Synthesizing constraint eupres- sions ,I’ Comm. ACM, vol. 21, pp. 958-966, 1978. 54	1982	102

0

QUESTION ORDERING IN MIXED INITIATIVE PROGRAM SPECIFICATION DIALOGUE Louis Steinberg Department of Computer Science Rutgers University New Brunswick, N. J. 08903 ABSTRACT It would be nice if a computer system could accept a program specification in the form of a mixed initiative dialogue. One capability such a system must have is the ability to ask questions in a coherent order. We will see a number of reasons it is better if such a system produces all the questions it can and has a "dialogue moderator" choose which to ask next, than if the system asks the first question it thinks of. DM [9I, the dialogue moderator of PSI C51, chooses questions by searching a network model of a program, under control of a set of heuristic rules. This technique is simple and flexible. I. Introduction When you need a computer program, it is usually easier to tell a human being what the program should do than to specify the program directly to a computer system (eg a compiler). There are a number of reasons for this, including the knowledge and reasoning ability that a human has. We will concentrate here, however, on another advantage of communicating with humans, their ability to engage in a mixed initiative dialogue, and on one particular capability required for carrying on such a dialogue,*the ability to ask questions in a coherent order. A mixed initiative dialogue is one in which either party may take initiative. From the perspective of the work reported here, to "take initiative" in a dialogue is to alter the structure of the dialogue. This definition is essentially equivalent to that of Bobrow, et al [II, who define taking initiative as establishing or --m---w- ' This work was supported by the Defense Advanced Research Projects Agency at the Department of Defense under contract MDA 903-76-C-0206. The author was also partially supported by an IBM Graduate Fellowship. The work reported here would not have been possible without the environment provided by Cordell Green and the other members of the PSI project. I would also like to thank N. S. Sridharan for his comments on an earlier draft of this paper. violating expectations about what will come next, since it is precisely the structure of a dialogue which gives rise to such expectations. In particular, we will be concerned here with "topic structure", the order and relationships of the topics covered in the dialogue, and with "topic initiative", the ability to affect topic structure. The work described here 191 been done in the context of the PSI program synthesis system [5]. PSI acquires program specifications via mixed initiative, natural language dialogue. II. The General Scheme In order to ask questions, such a system must be able to do two things: it has to decide what aspects of the specification are as yet incomplete, and it has to decide which one of these aspects to ask about next. We will refer to the latter problem, deciding which question to ask next, as the task of "question ordering". A. Order from the Reasoning Process One common way to handle question ordering might be summarized as asking the first question the system thinks of. In this scheme, the system goes through its normal reasoning process, and at some point comes across a fact which it wants to know, but cannot deduce. Whenever this happens, the system stops and asks the user. (See, for example, [11 and 143). Note that the system stops whenever it finds any question to ask. Thus, the system asks each question as it comes up, and the order is determined by the reasoning process. If a system's reasoning process seems natural to the user, then this scheme produces a question order which seems natural, at least to a first approximation. However, there are some problems. The basic problem is that this scheme ties the topic structure of the dialogue to the reasoning procedures of the system. This makes topic structure harder to change, since any change in topic structure requires a change in the reasoning procedure. It can also make it hard to transfer the question ordering methods to another system that uses a different reasoning method. Finally, this method of question ordering assumes that there is a single, sequential reasoning process, and is 61 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. not possible in a system structure such as that of HEARSAY-II [71. B. Order from a Dialog Moderator A better scheme is to have the reasoning process produce as many questions as it can, and to use some other mechanism to select a single one of them to ask next. This scheme largely avoids the problems of the previous one. Its main drawback is that it requires a reasoning process which is able to produce more than one question at a time. An additional advantage of this scheme is that it allows us to implement question ordering in a separate module, with a clearly defined interface to the rest of the system. I have termed such a module a "dialogue moderatortt. Thus, the dialogue moderator is given a list of all the questions currently open, and must choose which one is to be asked next, so as to keep the dialogue well structured. Much recent research (eg [2], C61, [8]> has shown that structure of a dialogue is closely tied to the structure of goals and plans being pursued by the dialogue's participants. One might therefore imagine that the dialogue moderator needs a complete model of goals and plans, both those of the system and those of the user. However, in a program specification dialogue, the goals and plans of both participants are tied very closely to the structure of the program. As will be seen, it has been possible in PSI to use a simple model of the program structure instead of a complex model of goals and plans. (It might be argued that any system which handles natural language will eventually need the full model of goals and plans anyway, so using a simpler model here is no savings in the long run. It should be noted, however, that mixed initiative does not necessarily imply natural language. A useful system might be constructed which handles mixed initiative dialogue in some formal language.) III. J& Specific Method DM is the dialogue moderator of the PSI system. As noted above, DM maintains a simplified model of the program being specified. The program is viewed as a structured set of objects. Each object is either a piece of algorithm or a piece of data structure - the pieces of algorithm correspond roughly to the executable statements of a program, and the pieces of data structure correspond roughly to the variable declarations. A specific loop or a specific input operation might be algorithmic objects, while a set or a 5-tuple might be data structure objects. These objects are structured by two relationships: an object may be a subpart of another (eg an input operation might be a step of a loop, and thus one of its subparts), and an algorithm object may use a data structure object (eg an input operation "usestt the data structure it inputs). DM represents this structure in a standard network form; nodes represent the objects, and arcs represent the relations subpart/superpart and uses/used-by. Each node also has associated with it a list of questions about the object it represents. (A question asks about some attribute of some specific object. The objects, relations, and questions come from other modules of PSI.) In order to choose the next question to ask, DM searches the net, starting at the "present topic". The present topic is the object currently being discussed. Determining which object this is is a difficult and important problem in its own right, involving the syntax of the user's sentences as well as the status of the program specification, and has not been seriously dealt with in this work. Instead, some simple heuristics are used, the main one being to assume that most of the time the user will be talking about the object that the system just asked about. Once the present topic has been chosen, the search proceeds, under control of a set of rules. (The rules are listed in the appendix. See [9] for a discussion of the specific rules.) Each time the search reaches an object, a list of rules is chosen (depending on whether the object is a piece of algorithm or data structure) and these rules are applied in order. Some say to look for a specific kind of question about the current object. Others say to move along some particular kind of arc from the current object, and recursively apply the rules on the object we reach. If no question is found by this recursive application, we come back and continue applying the rules here. If at any point a rule that looks for questions finds one, that question is the one to ask, and the search stops. This scheme of moving through the net and looking for questions, under control of a set of rules, has proven to be simple and flexible. A related technique was used in SCHOLAR [3]. SCHOLAR is a CA1 system which teaches geography by engaging in a mixed initiative dialogue with the student. Both participants may ask and answer questions. SCHOLAR chooses which question to ask by a random (rather than rule directed) walk on a net which encodes its knowledge about geography. As ultimately envisioned, SCHOLAR would teach in a Socratic manner, that is, by asking a carefully designed sequence of questions. However, the structure of goals and plans in such a dialogue is probably very different from the structure of the net as discussed in [31. Because of this, a scheme of moving through this net is unlikely to be useful for producing such a sequence of questions. DM's question ordering behavior has been tested in two ways. First, a log of runs of PSI was surveyed. This log included 42 dialogues which were essentially complete. Each dialogue was checked, both to see if the user complained about the question ordering (there is a comment feature that can be used for such complaints), and also to see if the question order was subjectively acceptable. Except for one instance, later traced to a program bug, DM's behavior was correct. This test was too subjective, however, so a simulated dialogue was recorded, with myself playing the role of PSI and a programmer from outside the PSI group as the user. The inputs DM would have gotten during this dialogue were hand coded and given to DM, and the questions DM chose were compared with those I had chosen. DM had to choose a question at sixteen points, with two to seven questions to choose from. The correct question was chosen at thirteen of these points. An analysis of the errors indicates that they could be removed by some straightforward extensions of the current methodology, particularly by maintaining more history of how the dialogue got to the present topic. IV. Conclusions Thus we see that it is advantageous for a system which engages in mixed initiative dialogue to have the reasoning modules produce all the questions they can at each point in the dialogue, and to have a separate dialogue moderator choose which one to ask next. In such a system, the question ordering mechanism is decoupled from the reasoning process, so that either can be modified without changing the other. A given mechanism for selecting one of the proposed questions can be more easily transferred to a system with very different reasoning mechanism. Also, multiple parallel reasoning processes can be used with this scheme. DM, the dialogue moderator of PSI, represents the program as a simple net of objects and relations. It chooses a question by starting at the node representing the present topic of the dialogue, and searching the net, under control of a set of rules. It is possible to use a simple model of the program, rather than a complex model of goals and plans, because in the program specification task, the participants' goals and plans are so closely tied to the program structure. This general scheme of rule based search is advantageous because it is simple and flexible. These techniques are probably applicable to other settings where the structure of goals and plans can be tied to some simple task related structure. APPENDIX: Question Choice Rules - _I_- (These are slightly simplified versions of the content of the rules. The actual rules consist of LISP code.) Rules for Algorithms Al) Are there questions about the NAME of this object? A2) Look at all objects that are USED-BY this object. A3) Are there questions about this object other than EXIT-TEST, PROMPT, or FORMAT? A4) Are there questions about the PROMPT or FORMAT for this object? A5) Look at all objects that are SUB-PARTS of this object. A61 Are there questions about the EXIT-TEST of this object? A7) Look at all objects that are SUPER-PARTS of this object. Rules for Data Structures Dl) Look at all objects that are SUB-PARTS of this object. D2) Are there questions about the STRUCTURE of this object? D3) Are there OTHER questions about this object? D4) Look at all objects that are SUPER-PARTS of this object. D5) Look at all objects that USE this object. REFERENCES Cl] Bobrow, D., Kaplan, R., Kay, M., Norman, D., Thompson, H., Winograd, T., "GUS, A Frame- Driven Dialogue System." Artificial Intelligence 8 (1977) 155-173. [2] Brown, G., "A Framework for Processing Dialogue", Technical Report 182, MIT Laboratory for Computer Science, June '977. [3] Carbonell, J. R., "AI in CAI: An Artificial Intelligence Approach t.0 Computer-Aided Instruction." IEEE Trans. Man-Machine Svst. J-l- (1970) 190-202. [4] Davis, R., Buchanan, B., Shortliffe, E., Production Rules as a Representation for a Knowledge-Based Consultation Program, Memo AIM- 266, Stanford Artificial Intelligence Laboratory, October, 1975. [5] Green, C., A Summary of the PSI Program Synthesis System, Proceedings of &h& Fifth International Joint Artificial IntelliKZ, Conference on Cambridge Massachusetts, August 1977, 380 - 381. [63 Grosz, B., The Representation and Use of Focus in a System for Understanding Dialogues, Proceedings of &t& Fifth d International Joint Conference on Artificial Intelligence, Cambridge, Massachusetts, August 1977, 67 - 76. [7] Lesser, V., Erman, L., A Retrospective View of the HEARSAY-II Architecture, Proceedin= of the Fifth International Joint Conference on Artificial Intelligence, Cambridge, Massachusetts, August 1977, 380 - 381. 183 Mann, W., Man-Machine Communication Research: Final ReDort, ISI/RR-77-57, USC Information Sciences Institute, February, 1977. [9] Steinberg, L., A Dialogue Moderator for Program Specification Dialogues in t,he psi Svstem, PhD dissertation, Stanford University, in progress. 63

1980

1

AUTOMATIC GENERATION OF SEMANTIC ATTACHMENTS IN FOL Luigia Aiello Computer Science Department Stanford University Stanford, California 94305 ABSTRACT Semantic attachment is provided by FOL as a means for associating model values (i.e. LISP code) to symbols of a first order language. This paper presents an algorithm that automatically generates semantic attachments in FOL and discusses the ad- vantages deriving from its use. I INTRODUCTION In FOL (the mechanized reasoning system developed by R. Weyhrauch at the Stanford A.I. Laboratory [4,5,6 1, the knowledge about a given domain of discourse is represented in the form of an L/S structure. F An L/S structure is the FOL counterpart of the logician notion of a theory/model pair. It is a triple <L,S,F> where L is a sorted first order language with equality, S is a simulation structure (i.e. a computable part of a model for a first order theory), and F is a finite set of facts (i.e. axioms and theorems). Semantic attachment is one of the characteriz- ing features of FOL. It allows for the construction of a simulation structure S by attaching a "model value" (i.e. a LISP data structure) to (some of) the constant, function and predicate symbols of a first order language. Note that the intended semantics of a given theory can be specified only partially, i.e. not necessarily all the symbols of the language need to be given an attachment. The FOL evaluator, when evaluating a term (or wff), uses both the semantic and the syntactic information provided within an L/S structure. It uses the semantic attachments by directly invoking the LISP evaluator for computing the value of ground sub-terms of the term (wff). It uses a simplification set, i.e. a user-defined set of rewrite rules to do symbolic evaluations on the term (wff). Semantic information and syntactic informa- tion are repeatedly used - in this order - until no further simplification is possible. The research reported here has been carried out while the author was visiting with the Computer Science Department of Stanford University on leave from IEI of CNR, Pisa, Italy. Author's permanent address: IEI-CNR, via S. Maria 46, I-56100 Pisa, Italy Semantic attachment has been vital in the generation of many FOL proofs, by significantly increasing the efficiency of evaluations. The idea of speeding up a theorem prover by directly invoking the evaluator of the underlying system to compute some functions (predicates) has been used in other proof generating systems. FOL is different from other systems in that it provides the user with the capability of explicitly telling FOL which semantic information he wants to state and use about a given theory. This approach has many advantages, mostly epistemological, that are too long to be discussed here. II AUTOMATIC GENERATION OF SEMANTIC ATTACHMENTS It is common experience among the FOL users that they tend to build L/S structures providing much more syntactic information (by specifying axioms and deriving theorems) than semantic inform- tion (by attaching LISP code to symbols). In recent applications of FOL, L/S structures are big, and (since the information is essentially syntactic)the dimension of the simplification sets is rather large. The unpleasant consequence is that the evaluations tend to be very slow, if feasible at all. This has prompted us to devise and implement an extension of the FOL system, namely, a compiling algorithm from FOL into LISP, which allows for a direct evaluation in LISP of functions and pre- dicates defined in First Order Logic. The compila- tion of systems of function (predicate) definitions from FOL into LISP allows FOL to transform syntac- tic information into semantic information. In other words, the compiling algorithm allows FOL to automatically build parts of a model for a theory, starting from a syntactic description. Semantic attachment has often been criticised as error prone. In fact, the possibility of directly attaching LISP code to symbols of the language allows the FOL user to set up the se- mantic part of an L/S structure in a language different from that of first order logic. This forbids him to use FOL itself to check the relative consistency of the syntactic and semantic part of an L/S structure. From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. The automatic transformation of FOL axioms (or, in general, facts) into semantic attachments, besides the above mentioned advantage of substanti- ally increasing the efficiency of the evaluator, has the advantage of guaranteeing the consistency between the syntactic and semantic specifications of an FOL domain of discourse, or at least, to keep to a minimum the user's freedom of introduc- ing non-detectable inconsistencies. The semantic attachment for a function (predi- cate)symbol can be automatically generated through a compilation if such a symbol appears in the syntactic part of an L/S structure as a definiendum in a system of (possibly mutually recursive) definitions of the following form: Yx 1 . ..x r.fixl,... X r> =6 i@i,Iyi,q⌧19 . l . +I V Yl l - Ys 4Pj (Yl . Y,) = Zj@jr~j,~j,Yl,...,Ysl) Here the f-s are function symbols and the P-s are predicate symbols. The mare terms in4-s, Y-s, F-s and x-s; the t-s are wffs in the$-s, q-s, ‘E-s and y-s. By ? we denote a tuple of constant sym- bols. By 9 (resp.2) we denote a tuple of function (resp. predicate) symbols. @(resp.2) may contain some of f (resp. P), but it is not necessarily limited to them, i.e. other function and predicate symbols besides the definienda can appear in each definiens. The compilation algorithm, when provided with a system of definitions, 'first performs a well- formedness check, then a compilability check. The well-formedness check tests whether or not all the facts to be compiled are definitions, i.e. if they have one of the two following forms (note that here we use the word "definition" in a broader sense than logicians do): fjxl...xr.fi(xl,...,xr) = . . . if Yl l **Ys.(Pj(yl'...,ys)= . . . The compilability check- consists in verifying that a> each definition is a closed wff, i.e. no free variable occurs in it; b) all the individual constants and the function (predicate) symbols appearing in the definiens either are one of the definienda or are attached to a model value (the first case allows for recursion or mutual re- cursion); c) the definiens can contain logical constants, conditionals and logical connectives but no quantifiers. When the FOL evaluator invokes the LISP evaluator, it expects a model value to be returned; it does not know how to handle errors occurring at the LISP level. This, for various reasons too long to be reported here, justifies the three above restrictions. Actually, the second and the third restrictions can be weakened with and appropriate extension of the FOL evaluator and of the compiler (respectively) to cope with the new situation. More details are presented in [l]. To present a simple example of compilation, consider the following facts: yy x.f(x,y) = if P(x) then g(x,x) else f(y,x) - vy x.g(x,y) = x+y If we tell FOL to compile them in an L/S structure where a semantic attachment exists both for the symbol P and for the symbol + (let them be two LISP functions named C-P and PLUS, respectively), it produces the following LISP code: (DE c-f (X y> (COND ((C-P x) (C-g x x)) (T (C-f y xl>>> (DE C-g (x y) (PLUS x Y> ) and attaches it to the function symbols f and g, respectively. III SOUNDNESS OF THE COMPILATION The compiling algorithm is pretty straight- forward, hence, its correctness should not con- stitute a problem. Conversely, a legitimate question is the following: Is the compilation process sound? In other words: Who guarantees that running the FOL evaluator syntactically on a system of definitions gives the same result as running the LISP evaluator on their (compiled) semantic attachments? The answer is that the two evaluations are weakly equivalent, i.e. if both terminate, they produce the same result. This is because the FOL evaluator uses a leftmost outermost strategy of function invocation (which corresponds to call-by- name) while the mechanism used by the LISP evalua- tor is call-by-value. Hence, compiling a function can introduce some nonterminating computations that would not happen if the same function were eval- uated symbolically. This, however, does not constitute a serious problem and it will be overcome in the next version of FOL. In fact, it will be implemented in a purely applicative, call-by-need dialect of LISP (note that, call-by-need is strongly equivalent to call-by-name in purely applicative languages). IV CONCLUSION FOL is an experimental system and, as is often the case with such systems, it evolves through the experience of its designer and users. Particular attention is paid to extend FOL only with new features that either improve its proving power or allow for a more natural interaction between the user and the system (or both) in a uniform way. The addition of the compiling algorithm sketched in the previous sections is in this spirit. This extension of FOL has been very useful in recent applications (see, for instance [2]). Experience has shown that the largest part of the syntactic information in an L/S structure can be compiled. This suggests a further improvement to be done on FOL evaluations. The use of the compiling algorithm leads to L/S structures where (almost) all the function (predicate) symbols of the language have an attachment. Hence, the strategy of the FOL evaluator to use semantic information first (that was the most reasonable one when semantic attachments were very few and symbolic evaluations could be rather long) is in our opinion no longer the best one. In fact, sometimes, properties of functions (stated as axioms or theorems in the syntactic part of the L/S structure) can be used to avoid long computations before invoking the LISP evaluator to compute that function. Finally, a comment on related work. Recently (and independently), Boyer and Moore have added to their theorem prover the possibility of introducing meta-functions, proving them correct and using them to enhance the proving power of their system [3]. This is very much in the spirit of the use of META in FOL and of the compiling algorithm described here. ACKNOWLEDGMENTS The members of the Formal Reasoning Group of the Stanford A.I. Lab are acknowledged for useful discussions. Richard Weyhrauch deserves special thanks for interesting and stimulating conversa- tions about FOL. The financial support of both the Italian National Research Council and ARPA (through Grant No. MDA903-80-C-0102) are acknowledged. REFERENCES [1] Aiello, L., "Evaluating Functions Defined in First Order Logic." Proc of the Logic -* - - Programming Workshop, Debrecen, Hungary, 1980. [2] Aiello, L., and Weyhrauch, R. W., "Using Meta- theoretic Reasoning to do Algebra." Proc. of -- the 5th Automated Deduction Conf., Les Arcs, France, 1980. [3] Boyer, R.S., and Moore, J.S., "Metafunctions: Proving them correct and using them efficiently as new proof procedures." C. S. Lab, SRI International, Menlo Park, California, 1979. [4] Weyhrauch, R.W., "FOL: A Proof Checker for First-order Logic." Stanford A.I. Lab, Memo AIM-235.1, 1977. [5] Weyhrauch, R. W., "The Uses of Logic in Artificial Intelligence." Lecture Notes of the Summer School on the Foundations of Artificial Intelligence and Computer Science (FAICS '78), Pisa, Italy, 1978. [6] Weyhrauch, R.W., "Prolegomena to a Mechanized Theory of Formal Reasoning." Stanford A.I. Lab, Memo AIM-315, 1979; Artificial Intelligence Journal, to appear, 1980. 92

1980

10

2

ABSTRACT HCPRVR: AN INTERPRETER FOR LOGIC PROGRAMS Daniel Chester Department of Computer Sciences University of Texas at Austin An overview of a logic program interpreter written in Lisp is presented. The interpreter is a Horn clause-based theorem prover augmented by Lisp functions attached to some predicate names. Its application to natural language processing is discussed. The theory of operation is explained, including the high level organization of the PROVE function and an efficient version of unification. The paper concludes with comments on the overall efficiency of the interpreter. An axiom is either an atomic formula, which can be referred to as a fact, or an expression of the form ( <conclusion> < <premissl> .*. <premissN> ) where both the conclusion and the premisses are atomic formulas. The symbol "<" is intended to be a left-pointing arrow. An atomic formula is an arbitrary Lisp ___ - expression beginning with a Lisp atom. That atom is referred to as a relation or predicate name. Some of the other atoms in the expression may be designated as variables by a flag on their property lists. I INTRODUCTION III CALLING LOGIC PROGRAMS - -- HCPRVR, a Horn Clause theorem PRoVeR, is a Lisp program that -interprets a --- simple logical formalism as a programming language. It has been used for over a year now at the University of Texas at Austin to write natural language processing systems. Like Kowalski [II, we find that programming in logic is an efficient way to write programs that are easy to comprehend. Although we now have an interpreter/compiler for the logic programming language Prolog [2], we continue to use HCPRVR because it allows us to remain in a Lisp environment where there is greater flexibility and a more familiar notation. This paper outlines how HCPRVR works to provide logic programming in a Lisp environment. The syntax of logic programs is given, followed by a description of how such programs are invoked. Then attachment of Lisp functions to predicates is explained. Our approach to processing natural language in logic programs is outlined briefly. The operation of HCPRVR is presented by giving details of the PROVE and MATCH functions. The paper closes with some remarks on efficiency. II LOGIC PROGRAM SYNTAX ~___-- ___ A logic program is an ordered list of axioms. -- -e--e * This work was supported by NSF Grant MCS 74-2491 -8. There are two ways to call a logic program in HCPRVR. One way is to apply the EXPR function TRY to an atomic formula. The other way is to apply the FEXPR function ? to a list of one or more atomic formulas, i.e., by evaluating an expression of the form ( ? <formulal> ..- <formulaN> > In either case the PROVE function is called to try to find values for the variables in the formulas that makes them into theorems implied by the axioms. If it finds a set of values, it displays the formulas to the interactive user and asks him whether another set of values should be sought. When told not to seek further, it terminates after assigning the formulas, with the variables replaced by their values, to the Lisp atom VAL. IV PREDICATE NAMES AS FUNCTIONS ~- Occasionally it is useful to let a predicate name be a Lisp function that gets called instead of letting HCPRVR prove the formula in the usual way. The predicate name NEQ*, for example, tests its two arguments for inequality by means of a Lisp function because it would be impractical to have axioms of the form (NEQ* X Y) for every pair of 93 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. constants X and Y such that X does not equal Y. Predicate names that are also functions are FEXPRs and expect that their arguments have been expanded into lists in which all bound variables have been replaced by their values. These predicate names must be marked as functions by having the Lisp property FN set to T, e.g., executing (PUT '<predicate name> 'FN T), so thi: HCPRVR will interpret them as functions. By letting syntactic categories be predicates with three arguments, we can make axioms that pull phrases off of a list of words until we get a sentence that consumes the whole list. In addition, arbitrary tests can be performed on the phrase representations to check whether they can be semantically combined. Usually the phrase representation in the conclusion part of an axiom tells how the component representations are combined, while the premisses tell how the phrase should be factored into the component phrases, what their' representations should be, and what restrictions they have. Thus, the axiom ((S X (U ACTOR V . W) Z) < (NP X V Y) (VP Y (U . W) Z) (NUMBER v ~1) (NUMBER U N2) (EQ N1 N2)) says that an initial segment of word list X is a sentence if first there is a noun phrase ending where word list Y begins, followed by a verb phrase ending where word list Z begins, and both phrases agree in number (singular or plural). Furthermore, the noun phrase representation V is made the actor of the verb U in the verb phrase, and the rest of the verb phrase representation, W, is carried along in the representation for the sentence. After suitable axioms have been stored, the sentence THE CAT IS ON THE MAT can be parsed by typing (? (S (THE CAT IS ON THE MAT) x NIL) The result of this computation is the theorem (S (THE CAT IS ON THE MAT) (IS ACTOR (CAT DET THE) LOC (ON LOC (MAT DET THE))) NIL) VI THEORY OF OPERATION -- A. General Organization HCPRVR works essentially by the problem reduction principle. Each atomic formula can be thought of as a problem. Those that appear as facts in the list of axioms represent problems that have been solved, while those that appear as conclusions can be reduced to the list of problems represented by the premisses. Starting from the formula to be proved, HCPRVR reduces each problem to lists of subproblems and then reduces each of the subproblems in turn until they have all been reduced to the previously solved problems, the "facts" on the axiom list. The key functions in HCPRVR that do all this are PROVE and MATCH. B. The PROVE Function -~ PROVE is the function that controls the problem reduction process. It has one argument, a stack of subproblem structures. Each subproblem structure has the following format: ( <list of subproblems>.<binding list> ) where the list of subproblems is a sublist of the premisses in some axiom and the CAR of the binding list is a list of variables occurring in the subproblems, paired with their assigned values. When PROVE is initially called by TRY, it begins with the stack ( ( ( <formula> ) NIL ) ) The algorithm of PROVE works in depth-first fashion, solving subproblems in the same left-to-right order as they occur in the axioms and applying the axioms as problem reduction rules in the same order as they are listed. PROVE begins by examining the first subproblem structure on its stack. If the list of subproblems in that structure is empty, PROVE either returns the binding list, if there are no other structures on the stack, i.e., if the original problem has been solved, or removes the first structure from the stack and examines the stack again. If the list of subproblems of the first subproblem structure is __ empty, not PROVE examines the first subproblem on the list. If the predicate name in it is a function, the function is applied to the arguments. If the function returns NIL, PROVE fails; otherwise the subproblem is removed from the list and PROVE begins all over again with the modified structure. 94 When the predicate name of the first subproblem in the list in the first subproblem stucture is not a function, PROVE gets all the axioms that are stored under that predicate name and assigns them to the local variable Y. At this point PROVE goes into a loop in which it tries to apply each axiom in turn until one is found that leads to a solution to the original problem. It does this by calling the MATCH function to compare the conclusion of an axiom with the first subproblem. If the match fails, it tries the next axiom. If the match succeeds, the first subproblem is removed from the first subproblem structure, then a new subproblem structure is put on the stack in front of that structure. This new subproblem structure consists of the list of premisses from the axiom and the binding list that was created at the time MATCH was called. Then PROVE calls itself with this newly formed stack. If this call returns a binding list, it is returned as the value of PROVE. If the call returns NIL, everything is restored to what it was before the axiom was applied and PROVE tries to apply the next axiom. The way that PROVE applies an axiom might be better understood by considering the following illustration. Suppose that the stack looks like this: ( ( (Cl C2).<blist> ) ..* ) The first subproblem in the first subproblem structure is Cl. Let the axiom to be applied be (C < Efl P2 P3) PROVE applies it by creating a new binding list blist', initially empty, and then matching C with Cl with the call (MATCH C <blist'> Cl <blist>). If this call is successful, the following stack is formed: ( ( (Pl P2 P3).<blist'> ) ( (C2).<blist> ) e.. ) Thus problem Cl has been reduced to problems Pl, P2 and P3 as modified by the binding list blist'. PROVE now applies PROVE to this stack in the hope that all the subproblems in it can be solved. In the event that the axiom to be applied is cc>, that is, the axiom is just a fact, the new stack that is formed is ( ( ().<blist'> ) ( (C2).<blist> ) . . . ) When PROVE is called with this stack, it removes the first subproblem stucture and begins working on problem C2. c. The MATCH Function -- The MATCH function is a version of the unification algorithm that has been modified so that renaming of variables and substitutions of variable values back into formulas are avoided. The key idea is that the identity of a variable is determined by both the variable name and the binding list on which its value will be stored. The value of a variable is also a pair: the term that will replace the variable and the binding list associated with the term. The binding list associated with the term is used to find the values of variables occurring in the term when needed. Notice that variables do not have to be renamed because MATCH is always called (initially) with two distinct binding lists, giving distinct identities to the variables in the two expressions to be matched, even if the same variable name occurs in both of them. MATCH assigns a value to a variable by CONSing it to the CAR of the variable's binding list using the RPLACA function; it also puts that binding list on the list bound to the Lisp variable SAVE in PROVE. This is done so-that the effects of MATCH can be undone when PROVE backtracks to recover from a failed application of an axiom. VII EFFICIENCY HCPRVR is surprisingly efficient for its simplicity. The compiled code fits in 2000 octal words of binary programming space and runs as fast t: the Prolog interpreter. Although the speed can further improved by more sophisticated programming, we have not done so because it is adequate for our present needs. A version of HCPRVR has been written in C; it occupies 4k words on a PDP11/60 and appears to run about half as fast as the compiled Lisp version does on a DEC KIlO. The most important kind of efficiency we have noticed, however, is program development efficiency, the ease with which logic programs can be written and debugged. We have found it easier to write natural language processing systems in logic than in any other formalism we have tried. Grammar rules can be easily written as axioms, with an unrestricted mixture of syntactic and non-syntactic computations. Furthermore, the same grammar rules can be used for parsing or generation of sentences with no change in the algorithm that applies them. Other forms of natural language processing are similarly easy to program in logic, including schema instantiation, question-answering and text summary. We have found HCPRVR very useful for gaining experience in writing logic programs. REFERENCES [l] Kowalski, R. A. "Algorithm = logic + control." CACM 22, 7, July, 1979, 424-436. [2] Warren, D. H., L. M. Pereira, and F. Pereira. "PROLOG - the language and its implementation compared with lisp." Proc. SymP' AI and Prog. Langs., SIGPLAN 12, -___ 8/sIGARTx4, August, 1977, 109-115. 95

1980

11

3

FIRST EXPERIMENTS WITH RUE AUTOMATED DEDUCTION Vincent J. Digricoli The Courant Institute and Hofstra University 251 Mercer Street, New York, N.Y. 10012 ABSTRACT RUE resolution represents a reformulation of binary resolution so that the basic rules of inference (RUE and NRF) incorporate the axioms of equality. An RUE theorem prover has been imple- mented and experimental results indicate that this method represents a significant advance in the handling of equality in resolution. A. Introduction In (1) the author presented the complete theory of Resolution by Unification and Equality which incorporates the axioms of equality into two inference rules which are sound and complete to prove E-unsatisfiability. Our purpose here is to present systematically the results of experiments with an RUE theorem prover. The experiments chosen were those of McCharen, Overbeek and Wos (2), ahd in particular we are interested in comparing the results achieved by these two theorem provers. In MOW, the equality axioms were used expli- citly for all theorems involving equality and apparently no use was made of paramodulation. In RUE, where proofs are much shorter, the inference rules themselves make implicit use of the equality axioms which do not appear in a refutation and also no use of paramodulation is made. Both systems are pure resolution-based systems. Before considering the experiments, we first review and summarize the theory of resolution by unification and equality as presented in (1). There we define the concept of a disagreement set, the inference rules RUE and NRF, the notion of viability, the RUE unifying substitution and an equality restriction which inhibits redundant inferences. Here we simply introduce the concept of a disagreement set and define the rules of inference. A disagreement set of a pair of terms (tl,t2) is defined in the following manner : is the on y disagreement set and if "If (tl,ts) are identical, the empty ;;st2) differ, the set of one pair { (tl,t2) is the origin disagreement set. i, Furt ermore, if tl has the form f(al,...,ak) and t2 the form f(bl,...,bk), then the set of pairs of corresponding arguments which are not iden- tical is the topmost disagreement set -0 In the simple example : tl = f( at g(WW) 1 t2 = f( a', g(b',h(c')) 1 besides the origin disagreement, there are the disagreement sets : Dl = i (ana'), ( g(b,h(c)) ,g(b',h(c')) 1) D2 = { (a,a'), (b,b'), (h(c) ,h(c') 1) D3 = ( tara'), (b,b'), (c,c') 1 This definition merely defines all possible ways of proving t =t , i.e. we can prove t =t by proving equali 4 12 y ?n every pair of any one disagreement set. An input clause set, for ex- ample, may imply equality in D D3' but not in D2 Or it may most directly phove tl=t2 by or proving equality in D3. We proceed to define a disagreement set of -- complementary literals : W1,...,sn) f;(t, v.,tn) as the union of disagreement sets : D = .U D i=l,n i where D i is a disagreement set of (si,ti). We see immediately that : P(sl"",sn) A ht1 I-O&~) * D where D now represents the disjunction of inequal- ities specified by a disagreement set of P,;, and furthermore, that : f(al~~~~,ak) # f(blto..,bk) --$ D where D is the disjunction of inequalities speci- fied by a disagreement set of f(al,...,ak), f(bl,-wbk) 0 For example, p(f(a,g(b,h(c)))) A F(f(a',g(b',h(c')))) 4 afa' A bfb' A cfc' . The reader is invited to read (1) which states the complete theory of RUE resolution with many examples. Our primary concern here is to discuss experiments with an RUE theorem prover and to begin to assess the effectiveness of this 96 inference system. we Experiments Our experiments deal with Boolean Algebra are asked to prove from the eight axioms : Al :x+0=x A2 :x*1=x A3 :x+Z=l A4 : x *xt=o A5 : x(y+z) = xy +x2 A6 : x + yz = (x+y) (x+2) A7 :x+y=y+x A0 :x*y=y*x (we are denoting logical or by +r logical and by * or juxtaposition, and negation by overbar), the following theorems : and a*0 # 0 I - x*x = 0 a*: # a*0 d = (a/x} - x(y+z) = xy + xz (7 ={a/x} y+z # ;; ay+az # a*0 t- x+0=x o- = E a/y,o/z *a/x j a*: + a*0 # a*0 t- 0+x=x 0- =la*O/x) Tl : Z=l T2 :x+1=1 T3 :x*0=0 T4 : x + xy = x T5 : x(x+y?" = x T6 :x+x=x T7 :x*x=x T8 : (x+y) +z = x+(y+z) T9 : (x*y)*z = x* (y*z) TlO : the complement of x is unique (x*a=O) (x+a=l) (x*b=O) (x+b=l) 4 a= b Tll :z=x -- T12 :x+y =x*y De Morgan's Law I -- T13 :x*y=x+;; De Morgan's Law II These theorems are stated in the order of increasing complexity of proof, with 6= 1 being trivially easy for a human to prove and De Morgan's Laws being very difficult for a human to deduce from the axioms. George and Garrett Birkhoff have a paper on the above proofs published in the Transactions of the American Mathematical Society (3) and Halmos comments on the significantly difficult character of the proofs in his Lectures on Boolean Algebras (4) 0 The following is a machine deduced, five step RUE refutation which proves x*0 = 0 : 0 # a*a - 0 = x*; D o-=<a/x] The above experiments together with many others (dealing with group theory, ring theory, geometry, Henken Models, set theory and program verification) were proposed as benchmarks by McCharen,Overbeek and Wos, who in (2) published the results of their own experiments. We here tabulate the comparative performance of the RUE and MOW theorem provers on the above thearems. The MOW theorem prover uses binary resolution with explicit use of the equality axioms and is implemented in Assembly language on the IBM System 370-Model 195. Great effort was made to enhance the efficiency of their theorem prover and this is described in (2). The RUE theorem prover, on the other hand, represents a first implementation in PLl on a CDC 6600 machine which is much slower than the Model 195. In the experiments each theorem is treated as an independent problem and cannot use earlier theorems as lemmas, so that for example in proving associativity (T8), we need to prove (T2,T3,T4,T5) as sub-theorems. The total number of unifications performed is suggested as the primary measure of comparison rather than time. The comparative results are given in Table 1. From Tl to T7, The RUE theorem prover was very successfull, but at T8 (associativity) results have yet to be obtained since refinements in the heuristic pruning procedure are required and are being developed with the expectation that more advanced results will be available at the conference. RUE represents one of several important methods for handling equality in resolution and it is important to emphasize that it is a complete method whose power is currently being tested in stand-alone fashion. However, it is not precluded that we can combine this method with other tech- niques such as demodulation,paramodulation and reduction theory to achieve a mutually enhanced effect. 97 TABLE 1. THEOREM TOTAL NUMBER OF UNIFICATIONS RUE : MOW Tl 'i=l 77 26,702 T2 x+1=1 688 46,137 T3 x*0=0 676 46,371 T4 x+xy=x 3,152 see below T5 X(X+Y) = x -3,113 tl I! T7 X*X=X 2,145 n m T6rrT7 4,326(l)105,839 T8 (x+y)+z = x+(y+z) IP 413,455 T9 (x*y)*z = x*(y*z) IP NPR TlO ';:;~;I ';;-;I+ a=b IP NPR ( -I( 1 e Tll z =x IP NPR T12 x+y = ;; * ;; IP NPR - - - T13 x*y=x+y IP NPR TIME LENGTH OF PROOF (SECONDS) RUE : MOW RUE : MOW 10.1 27.5 7 51.5 " " 12 102.9 57.0 24 41.6 see below 13 Note 1 : To prove the double theorem, T4,,T5, x+xy=x A x(x+y)=x, we add the negated theorem as a single clause, a+ab#a v a(a+b)#a , to the input clause set. It is evident that the erasure of these two literals in a refutation decomposes into two independent subproblems since no variables appear in the clause. Hence, the refutations for a+ab#a and a(a+b)#a obtained in separate experi- ments T4,T5 can be concatenated and the results of these experiments simply summed which is what we have done to state the RUE results for the double theorem. The same holds true for T6hT7. * The estimated length of MOW proofs with the equality axioms is twice as long as corresponding RUE proofs. REFERENCES : 1. "Automatic Deduction and Equality" by Vincent J. Digricoli, Proceedings of the Oct., 1979, Annual Conference of the ACM, 240-250. 2. "Problems and Experiments for and with Automat- ed Theorem-Proving Programs", by McCharen, Overbeek and Wos, IEEE Transactions on Comput- ers, Vol C-25,No.8,August 1976. 3. "Distributive Postulates for Systems Like Boolean Algebras", Transactions of the American Mathematical Society, Volume 60, July-Dec. 1946. 4. "Lectures on Boolean Algebras", by Paul R. Halmos, Section 2, pages 3-5, D.Van Nostrand. The completion of these Boolean experiments together with other experiments with a more fully delineated comparative analysis with systems other than MOW represents work the author will report upon in the future. 98

1980

12

4

WHAT’S WRONG WITH NON-MONOTONIC LOGIC? David J . Israel Bolt Beranek and Newman Inc. 50 Moulton St. Cambridge, Mass. 02238 ABSTRACT In this paper ’ I ask, and attempt to answer, the following question : What’s Wrong with Non-Monotonic Logic? The answer, briefly’ is that the motivation behind the wonderfully impressive work involved in its development is based on a confusion of proof-theoretic with epistemological issues. ------------ What’s wrong with non-monotonic logic (and fo! that matter, with the logic of default reasoning)? The first question we should ask is: What’s supposed to be wrong with “standard”, monotonic logic? In recent - and extremely impressive - work, Doyle and McDermott [ 1 I, McDermott C21, and Reiter C31 have argued that classical logic - in virtue of its monotoniqity - is incapable of adequately capturing or representing certain crucial features of real live reasoning and inference. In particular’ they note that our knowledge is always incomplete, and is almost always known to be so ; that, in pursuing our goals - both practical and theoretical - we are forced to make assumptions or to draw conclusions on the basis of incomplete evidence ; conclusions and assumptions which we may have to withdraw in the light of either new evidence or further cogitation on what we already believe. An essential point here is that new evidence or new inference may lead us to reject previously held beliefs, especially those that we knew to be inadequately supported or merely presumptively assumed. In sum, our theories of the world are revisable; and thus our attitudes towards at least some our beliefs must likewise be revisable. Now what has all this to do with logic and its monotonicity? Both Reiter and Doyle-McDermott characterize the monotonicity of standard logic in syntactic or proof-theoretic terms. If A and B are two theories, and A is a subset of B, then the ‘The research reported in this paper was supported in part by the Advanced Research Projects Agency, and was monitored by ONR under Contract No. NOOO14-77-C-0378. To remedy this lack, Doyle and McDermott introduce into an otherwise standard first order language a modal operator “M” which, they say, is to be read as “It is consistent with everything that is believed that.. .” (Reiter’s “M”, which is not a symbol of the object language, is also supposed to be read “It is consistent to assume that..“. I think there is some unclarity on Reiter’s part about his “M”. He speaks of it in ways conducive to interpreting it as a metalinguistic predicate on sentences of the object language ; and hence not as an operator at all, either object-language or metalanguage. So his default rules are expressed in a language whose object-language contains sentences of the form l’Mp” , i .e . , in a language which, relative to the original first-order object language, is a meta-meta-language .) Now in fact this reading isn’t quite right. ** The suggested reading doesn’ t capture the notion Doyle-McDermott and Reiter seem to have in mind. What they have in mind is, to put it non-linguistically (and hence, of course, non-syntactically) : that property that a belief has just in case it is both compatible with everything a given subject believes at a given time and remains so when the subject’s belief set undergoes certain kinds of changes under the pressure of both new information and further thought, and where those changes are the result of rational epistemic -----_ policies. 99 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. I’ve Put the notion in this very epistemologically oriented way precisely to hone in on what I take to be the basic misconception underlying the work on non-monotonic logic and the logic of default reasoning. The researchers in question seem to believe that logic - deductive logic , for there is no other kind - is centrally and crucially involved in the fixation and revision of belief. Or to put it more poignantly, they mistake so-called deductive rules of inference for real, honest-to-goodness rules of inference. Real rules of inference are precisely rules of belief fix ation and revision ; deductive rules of transformation are precisely not. Consider that old favorite : modus ( ponendo) ponens. It is not a rule that should be understood as enjoining us as follows : whenever you believe that p and believe that if p then q, then believe that q. This, after all, is one lousy policy. What if you have overwhelmingly good reasons for rejecting the belief that q? All logic tells you is that you had best reconsider your belief that p and/or your belief that if p then q (or, to be fair, your previously settled beliefs on the basis of which you were convinced that not-q); it is perforce silent on how to revise your set of beliefs so as to . . to what? Surely, to come up with a good theory that fits the evidence, is coherent, simple, of general applicability, reliable, fruitful of further testable hypotheses, etc. Nor is it the case that if one is justified in believing that p and justified in believing that if p then q (or even justified in believing that p entails q) , is one justified in be1 iev ing (inferring) that c~. Unless, of course, one has no other relevant ---- beliefs. Butone always does. ---- The rule of modus ponens is, first and foremost, a rule that permits one to perform certain kinds of syntactical transformations on (sets of) formally characterized syntactic entities. (Actually, first and foremost, it is not really a rule at all; it is f’reallyff just a two-place relation between on the one hand an ordered pair of wffs., and on the other, a wff .> It is an important fact about it that, relative to any one of a family of interpretations of the conditional, the rule is provably sound, that is ** Nor is it quite clear. By “consistentff are we to mean syntactically consistent in the standard monotonic sense of syntactic derivability or in the to-be-explicated non-monotonic sense? Or is it semantic consistency of one brand or another that is in question? This unclarity is fairly quickly remedied . We are to understand by If consistencyff standard syntactic consistency, which in standard systems can be understood either as follows: A theory is syntactically consistent iff there is no formula p of its language such that both p and its negation are theorems, or as follows : iff there is at least one sentence of its language which is not a theorem. There are otherwise standard, that is, monotonic, systems for which the equivalence of these two notions does not hold; and note that the first applies only to a theory whose language includes a negation operator. truth (in an interpretation)-preserving . The crucial point here, though, is that adherence to a set of deductive rules of transformation is not a sufficient condition for rational belief; it is sufficient (and necessary) only for producing derivations in some formal system or other. Real rules of inference are rules (better : policies) guiding belief fixation and revision. Indeed, if one is sufficiently simple-minded, one can even substitute for the phrase ” good rules of inference”, the phrase ‘I( rules of) scientific procedure” or even “scientific method”. And, of tour se, there is no clear sense to the phrase “good rules of transformation”. (Unless ffgoodff here means ffcompleteff - but with respect to what? Truth? > Given this conception of the problem to which Doyle-McDermott and Reiter are addressing themselves, certain of the strange properties of, on the one hand, non-monotonic logic and on the other, the logic of default reasoning, are only to be expected. In particular, the fact that the proof relation is not in general decidable. The way the “Mfl operator is understood, we believers are represented as follows: to make an assumption that p or to put forth a presumption that p is to be1 iev e a proposition to the effect that p is consistent with everything that is presently believed and that it will remain so even as my beliefs undergo certain kinds of revisions. And in general we can prove that p only if we can prove at least that p is consistent with everything we now be1 iev e . But, of course, by Church’s theorem there is no uniform decision procedure for settling the question of the consistency of a set of first-order formulae . (Never mind that the problem of determining the consistency of arbitrary sets of formulae of the sentential calculus is NP-complete . > This is surely wrong-headed : assumptions or hypotheses or presumptions are not propositions we accept only after deciding that they are compatible with everything else we be1 iev e , not to speak of having to establish that they won’t be discredited by future evidence or further reasoning. When we assume p, it is just p that we assume, not some complicated proposition about the semantic relations in which it stands to all our other beliefs, and certainly not some complicated belief about the syntactic relations any one of its linguistic expressions has to the sentences which express all those other beliefs. (Indeed, there is a problem with respect to the consistency requirement, especially if we allow be1 ief s about beliefs. Surely, any rational subject will believe that s/he has some false be1 iefs , or more to the point, any such subject will be disposed to accept that belief upon reflection. By doing so, however, the subject guarantees itself an inconsistent belief-set; there is no possible interpretation under which all of its beliefs are true. Should this fact by itself worry it (or us?) .) After Reiter has proved that the problem of determining whether an arbitrary sentence is in an extension for a given default theory is undecidable, he comments: (A)ny proof theory whatever for... the facts? (Are the rules provably sound rules of transformation?) Or are the conclusions legitimate because they constitute essential (non-redundant) parts of the best of the competing explanatory accounts of the original data; the best by our own, no doubt somewhat dim, lights? (Are the rules arguably rules of rational acceptance?) At the conclusion of his paper, McCarthy disambiguates and opts for the right reading. In the context of an imaginative discussion of the Game of Life cellular automaton, he notes that "the program in such a computer could study the physics of its world by making theories and experiments to test them and might eventually come up with the theory that its fundamental physics is that of the Life cellular automaton. We can test our theories of epistemology and common sense reasoning by asking if they would permit the Life-world computer to conclude, on the basis of its experiments, that its physics was that of Life." McCarthy continues: default theories must somehow appeal to some inherently non semi-decidable process. [That is, the -proof-relation, not just the proof predicate, is non recursive; the proofs, not just the theorems, are not Why such a beast recursively enumerable. is to be called a logic is somewhat beyond me - DI.1 This extremely pessimistic result forces the conclusion that any computational treatment of defaults must necessarily have an heuristic component and will, on occasion, lead to mistaken beliefs. Given the faulty nature of human common sense reasoning, this is perhaps the best one could hope for in any event. Now once ag ain substitute in the above "(scient ific or common sense) reasoni ng I1 for "defaulted and then reflect on how odd it is to think that there could be a purely proof-theoretic treatment of scientific reasoning. A heuristic treatment, that is a treatment in terms of rational epistemic More generally, we can imagine a metaphilosophy that has the same relation to philosophy that metamathematics has to mathematics. Metaphilosophy would study mathematical (? - D.1.) systems consisting of an 'fepistemologist'f seeking knowledge in accordance with the epistemology to be tested and interacting with a ffworldff. It would study what information about the world a given philosophy would obtain. This would depend also on the structure of the world and the ffepistemologist'sff opportunities to interact. AI could benefit from building some very simple systems of this kind, and so might philosophy. policies, is not just the best we could hope for. It is the only thing that makes sense. (Of course, if we are very fortunate a "syntactic" encoding we may be able to develop f these policies; but we 0 certainly mustn't expect to come up with rules for rational belief fixation that are actually provably truth-preserving. Once again, the only thing that makes sense is to hope to form ulate a set of rules which, from within our current theory of the world of ourselves as both objects within and inouirers about that world, can be argued to embody rational policies for extending our admittedly imperfect grasp of things.) Inference (reasoning) is non-monotonic: New information (evidence) and further reasoning on old beliefs (including, but by no means limited to, reasoning about the semantic relationships - e.g., of entailment - among beliefs) can and does lead to Amen; but might I note that such a metaphilosophy does exist. Do some substituting again: for " hilosophyf' P (except in its last occurrence), substitute ff~~iencef'; for 'fepistemologistff, ffscientistff; for ffepistemologyff, either "philosophy of science" or "scientific methodology". The moral is, I hope, clear. Here is my constructive proposal: AI researchers interested in "the epistemological problem" should look, neither to formal semantics nor to proof-theory; but to - of all things - the philosophy of science and epistemology. the revision of our theories and, of course, to revision bv f'subtractionff as well as by ffaddition'f. Entailment- and derivability are monotonic. That is, logic - the logic we have, know, and - if we understand its place in the scheme of things - have every reason to love, is monotonic. BRIEF POSTSCRIPT I've been told that the tone of this paper is REFERENCES overly critical; or rather, that it - lacks constructive content. A brief postscript is not [IlMcDermott, D., Doyle, J. "Non-Monotonic Logic I" , AI Memo 486, MIT Artificial Intelligence Laboratory, Cambridge, Mass., August 1978. C2lMcDermott. D. "Non-Monotonic Logic II", Research Report 174, Yale University Department of Computer Science, New Haven, Conn., February 1980. the appropr iate locus for correcting this defect; but it may be an appropri ate place for cas ting my vote for a suggestion made by John McCarthy. In his "Epistemological Problems of Artificial Intelligence" [41. McCarthy characterizes the epistemological part of "the AI problem" as follows: "(it) studies what kinds of facts about the world are available to an observer with given opportunities to observe, how these facts can be represented in the memory of a computer, and what rules permit legitimate conclusions to be drawn - - from these facts." [Emphasis added.] ThisTthough brief, is just about right, except for a perhaps C3lReiter, R. "A Logic for Default Reasoning", Technical Report 79-8, University of British Columbia Department of Computer Science, Vancouver, B.C., July 1979. 143McCarthy' J. "Epistemological Problems of studied ambiguity in that final clause. Are the Artificial Intelligence",-In Proc. IJCAI-77. Cambridge, Mass., August, 1977, pp. 1038-1044. conclusions legitimate because they are entailed by 101

1980

13

5

PATHOLOGY ON GAME TREES: A SUMMARY OF RESULTS* Dana S. Nau Department of Computer Science University of Maryland College Park, MD 20742 ABSTRACT Game trees are widely used as models of various decision-making situations. Empirical results with game-playing computer programs have led to the general belief that searching deeper on a game tree improves the quality of a decision. The surprising result of the research summarized in this paper is that there is an infinite class of game trees for which increasing the search depth does not improve the decision quality, but instead makes the decision more and more random. I INTRODUCTION - Many decision-making processes are naturally modeled as perfect information games between two players [3, 71. Such games are generally represented as trees whose paths represent various courses the game might take. In artificial intelligence, the well-known minimax procedure [2, 71 is generally used to choose moves on such trees. If a correct decision is to be guaranteed using minimaxing, substantial portions of the game tree must be searched, even when using tree- pruning techniques such as alpha-beta [2, 71. This is physically impossible for large game trees. However, good results have been obtained by searching the tree to some limited depth, estimating the minimax values of the nodes at that depth using a heuristic evaluation function, and computing the minimax values for shallower nodes as if the estimated values were correct [2, 71. There is almost universal agreement that when this is done, increasing the search depth increases the quality of the decision. This has been dramatically illustrated with game-playing computer programs [l, 8, 91, but such results are purely empirical. the The author has developed a mathematical ory modeling the effects of search depth on the * This work was supported in part by a National Science Foundation graduate fellowship, in part by a James B. Duke graduate fellowship, and in part by N.S.F. grant number ENG-7822159 to the Laboratory for Pattern Analysis at the University of Maryland. The results discussed in this paper are presented in detail in the author's Ph.D. dissertation [4]. probability of making a correct decision. This research has produced the surprising result that there is an infinite class of game trees for which as long as the search does not reach the end of the tree (in which case the best possible decision could be guaranteed), deeper search does not improve the decision quality, but instead makes the decision more and more random. For example, probability of search depth 0 2 4 6 8 10 12 14 16 FIGURE l.--Probability of correct decision as a function of search depth on the game tree G(l,l), for five different evaluation functions. On G(l,l), the probability of correct decision is 0.5 if the choice is made at random. For each of the five functions, this value is approached as the search depth increases. 102 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Figure 1 illustrates how the probability of correct decision varies. Section 2 of this paper summarizes the mathematical model used in this research, Section 3 presents the main result, and Section 4 contains concluding remarks. II THE MATHEMATICAL MODEL -- Let G be a game tree for a game between two players named Max and Min. Nodes where it is Max's or Min's move are called max and min nodes, respectively. Assume that G has no draws (this restriction can easily be removed, but it simplifies the mathematics). Then if G is finite, every node of G is a forced win for either Max or Min. Such nodes are called "+I' nodes and 1(-11 nodes, respectively. If G is infinite, not every node need be a "+,. or II- II node, but the "+" and M-n labeling can easily be extended to all nodes of G in a way which is consistent with all finite truncations of G. Correct decisions for Max and Min are moves leading to "+" and u-11 nodes, respectively. "+" max nodes (which we call S nodes) may have both "+" and 1(-(1 children; "+" min nodes (T nodes) have only "+" children; *-II min nodes (U nodes) may have both "+" and 11-(1 children; and --II max nodes (V nodes) have only 11-11 children. Thus it is only at the S and U nodes that it makes a difference what decision is made. These nodes are called critical nodes. An evaluation function on G may be any mapping e from the nodes of G into a set of numbers indicating how good the positions are estimated to be. For computer implementation, the range of e must be finite. We take this finite set to be fO,l,...,r}, where r is an integer. Ideally, e(g) would equal r if g were a "+" node and 0 if g were a 11-u node, but evaluation functions are usually somewhat (and sometimes drastically) in error. Increasing the error means decreasing e(g) if g is a "+" node and increasing e(g) if g is a 11-11 node. Thus if we assume that the errors made by e are independent and identically distributed, the p.d.f. f for the values e returns on "+" nodes is a mirror image of the p.d.f. h for the values e returns on 1)-11 nodes; i.e., f(x) = h(r-x), x = O,l,...,r. f may be represented by the vector P= (f(O),f(l),...,f(r>>, which is called the probability vector for e. --- 111 RESULTS -- The probability vector for e induces probability vectors on the minimax values of the nodes of G, and the probability of making a correct decision at any critical node g of G is a function of the probability vectors for the minimax values of the children of g. This probability is thus determined by the structure of the subtree rooted at g, and little can be said about it in general. However, if G has a sufficiently regular structure, the properties of this probability can be analyzed. Let m and n be positive integers, and let G(m,n) be the unique game tree for which 1. the root is an S node (this choice is arbitrary and the results to follow are independent of it); 2. each critical node has m children of the same sign and n children of opposite sign; T . . . T U . . . U U . . . U m+n A-A . . . . . . FIGURE 2.--The game tree G(m,n). Min nodes are indi- cated by the horizontal line segments drawn beneath them. 103 3. every node has m+n children. G(m,n) is illustrated in Figure 2. If moves are chosen at random on G(m,n), the probability that a correct choice is made at a critical node is obviously m/(m+n). If the choice is made using a depth d minimax search and an evaluation function with probability vector P, it is proved [4, 51 that the probability that the decision is correct depends only on m, n, P, and d. We denote this probability by &,n(P,d). The trees G(m,n) have the following surprising property. Theorem 1. For almost every* probability ~- vector P and for all but finitely many values of m and n, lim &,n(P,d) = m/(m+n). d-3 co Thus, as the search depth increases, the probability of correct decision converges to what it would be if moves were being chosen at random. This pathological behavior occurs because as the search depth increases it becomes increasingly likely that all children of a critical node receive the same minimax value, whence a choice must be made at random among them. Figure 1 illustrates Theorem 1 on the game tree G(l,l), using five different values of P. The significance of Theorem 1 for finite games is that infinitely many finite games can be generated by truncating G(m,n) in whatever way desired. Deeper search on these trees will yield increasingly random decisions as long as the search does not reach the end of the tree. Additional theoretical and experimental results reported elsewhere [4, 5, 61 provide additional information about which of the G(m,n) are pathological and why. Theorem 1 almost certainly extends to a much larger class of game trees, but the irregular structure of most game trees would require a much more complicated proof. IV CONCLUSIONS - The author believes that the pathology of the trees G(m,n) indicates an underlying pathological tendency present in most game trees. However, in most games this tendency appears to be overridden by other factors. Pathology does not appear to occur in games such as chess or checkers [l, 8, 91, but it is no longer possible blithely to assume (as has been done in the past) that searching deeper will always result in a better decision. 1. 2. 3. 4. 5. 6. 7. 8. 9. REFERENCES Biermann, A. W. Theoretical issues related to computer game playing programs. Personal Comput. (Sept. 1978), 86-88. Rnuth, D. E., and Moore, R. W. An analysis of alpha-beta pruning. Artif. Intel. 6 - - (1975), 293-326. LaValle, I. H. Fundamentals of Decision Analysis. Holt, Rinehart and Winston, New York, 1978. Nau, D. S. Quality of decision versus depth of search Dissertation, Dull Uni~?m~Aug~~~eSb79). Ph.D. Nau, D. S. Decision quality as a function of search depth on game trees. Tech. Report TR-866, Computer Sci. Dept., Univ. of Md. (Feb. 1980). Submitted for publication. Nau, D. S. The last player theorem. Tech. Report TR-865, Computer Sci. Dept., Univ. of Md. (Feb. 1980). Submitted for publication. Nilsson, N. J. Problem-Solving Methods in Artificial Intelligence. --ram NG York, 1971. Robinson, A. L. Tournament competition fuels computer chess. Science 204 (1979), 1396-l 398. Truscott, T. R. Minimum variance tree searching. Proc. First Int. Symp. on Policy Anal. and IX------- Syst. (1979X 203-209.- - - * A property holds for almost every member of a set if it holds everywhere but on a subset of measure zero. Thus for any continuous p.d.f. on the set, the probability of choosing a member of the set to which the property does not apply is 0. 104

1980

14

6

ISEOC9PFREE S. W. Ng and Adrian Walker Work performed at Rutgers University* ABSTRACT If a system uses assertions of the general form x causes y , (e.g. MYCIN rules) then loop situations in which X, causes X2, X2 causes X3, . . . . , X, causes X,, are, intuitively, best avoided. If an assertion has an attached confidence weight, as in x (0.8)-causes y , then one can choose to say that the confidence in a chain of such assertions is as strong as the weakest link in the chain. If there are several chains of assertions from X to Z, then one can choose to say that X causes Z with a confidence equal to that of the strongest chain. From these cfioices, it follows that the confidence that X causes Z corresponds to a loop-free chain of assertions. This is true even if there are chains from X to Z with common subchains and loops within loops. An algorithm for computing the confidence is described. I INTRODUCTION and TERMINOLOGY There is currently considerable interest in representing knowledge about a practical situation in the form of weighted cause-effect or situation-action rules, and in using the knowledge so represented in decision- making systems. For example, in medical decision mak- ing systems, the rules may represent causal trends in a disease process in a patient [6], or the rules may represent trends in the decision process of a physician who is diagnosing and treating a patient [2,4]. In such representations, the chaining together of rules can be written as a weighted, directed graph. In MYCIN [2] the graphs are like and-or trees, while in OCKHAM [3,4,5] the graphs may have loops. This paper presents a result which appears in [l]. From the result it follows that, using the max and min operations, a graph con- taining loops can be interpreted as though it were loop- free. *Authors’ pres ent addresses: S. W. Ng, 6F Wing Hing Street, Hong Kong. Adrian Walker, Bell Laboratories, Murray Hill, NJ. The kind of graph in question is a stochastic graph (sg) consisting of nodes N = {1,2,...,n) and a function P from N X N to the real numbers W, 0 5 w 11. P is such that, for each i E N, 2 P(i,j) 5 1. If P(i,j) = W, then w J=i is the weight of the arc from node i to node j. A path in an sg is a string n, . . nl c N+ such that P(nk,nk+,) > 0 for lsk-cl. n2,...,ni-l are intermediate nodes of . . . q. A path n, . ;dop if n, = n nl of a graph is said to have a J for some i.j such that either 1 li < j <I or I <i < j I I. Otherwise the path is loop-free. The weight of a path n, n2 . nl of an sg is the minimum over 1 <i < I of the weight of the arc from n, to n,+l. The k-weight w: from node i to node j of a graph, is the maximum of the weights of all the paths from i to j having no intermediate node with number higher than k. The weight w,/ from node i to node j of an sg is w,;. II EXAMPLES This section gives examples of potential causal loops, in MYCIN [2] and in OCKHAM [3,4,5], and it shows how these loops are avoided by the use of the maximum and minimum operations. A. A MYCIN Example Consider a set of MYCIN rules B A C (l.O)- A B (l.O)- D D V E (0.5)- B G AH (0.5)- B and suppose that C, E, G, and H are known with confidences 0.9, 0.8, 0.5, 0.4, respectively. Writing C(X) for the confidence in X, confidences propagate through rules by: 105 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. c(Z) = w . max(c(X),c(Y)) for xv Y (w)+ 2 III ALGORITHM and RESULTS and, c(Z) = w . min(c(X),c(Y)) for x A Y (w)-2. The greatest confidence which can be computed in A is c (A ) = 0.4 by the tree A+((Bt-(D+(B*G AH)) V E) A C) B occurs twice, so the tree can be thought of as a graph with a loop. However, the value of C(A) depends only on the loop-free path EBA. B. An OCKHAM Example The following set of OCKHAM [3,4,5] rules is intended to show a strategy of a person who is deciding whether to stay at home, go to the office, or to try for a standby flight to go on vacation. The external factors project deadline, sno&torm, project completed, and another flight influence the choice by placing the arc(s) so labelled in a stochastic graph. The rules are: HOME (project deadline, 1 .O)+ OFFICE OFFICE (snowstorm, 0.5)--+ HOME OFFICE (project completed, 0.5)-+ AIRPORT-STANDBY AIRPORT-STANDBY (another flight, 0.25)+ AIRPORT-STANDBY AIRPORT-STANDBY (snowstorm, 0.75)-+ HOME These rules make up a stochastic graph with nodes HOME, OFFICE, and AIRPORT-STANDBY. If all of the external factors project deadline, snowstorm, project com- pleted, and another flight are true, then the graph has five arcs and multiple loops. If the weight from HOME to AIRPORT-STANDBY is considered, then it turns out to be 0.5. The corresponding path, HOME-OFFICE-AIRPORT-STANDBY, is loop-free. The algorithm MAXMIN, shown below, computes the weight from a given node to another node in an sg. Note that, by Step 2, MAXMIN runs in o(n3) time. MAXMIN Input: A stochastic graph of n nodes Output: n2 real numbers Step 1: for 1 5 i,j 5 n do B,,O := P(i,j) Step 2: for k:=l to n do for 1 1. i,j I n do B{ := max (BG-‘, min (B,kk-‘, BjJ-‘)) Step 3: for 1 I ij I n do output Bl The properties of paths, path weights, and the values Blf, described in the Lemma below, are esta- blished in Appendix I. Lemma In an sg of n nodes, the following hold for 1 5 i,j 5 n andforork (n: statements (i) If WI; > 0, then there exists i to j whose weight is w;, a loop-free path from (ii) B: = WI:. Setting k=n in parts (i) and (ii) of the Lemma yields the two results: Result I In any sg the weight wlJ, that is, the maximum path weight over all paths from i to j, is equal the max- imum over only the loop-free paths from i to j. Result 2 If MAXMIN receives as input an sg with n nodes, then, for 1 I ij I n, the output Bl; is equal the weight wi, from node i to node j of the graph. Result 1 establishes a property of any sg, namely that the weight from one node to another is the weight of some loop-free path, while Result 2 establishes that MAXMIN is one possible algorithm for finding such weights. 106 This is because (A) and (B) exhaust all possibilities. IV CONCLUSIONS In a system in which weighted causal assertions can be combined into causal paths and graphs, causal loops can occur. Common sense about everyday causality suggests that such loops are best avoided. If the weight of a path is chosen to be the minimum of the individual arc weights, and the net effect of a start node on a final node is chosen to be the maximum of the path weights from the start node to the final node, then the weights (by whatever algorithm they are computed) are independent of the presence of loops in the underlying graph. There is a simple O(n3) algorithm to compute these weights. ACKNOWLEDGEMENT Thanks are due to A. Van der Mude for helpful comments. APPENDIX I Proof of Lemma Let k = 0. If wij > 0, then by definition, there exists a path from i to j having no intermediate nodes, whose weight is wiio. Clearly this path is ij, which is loop-free. So we may write r,$’ = ij, where 76 denotes a path from i to j having no inter- mediate node with number greater than k. Then WijO = P (i j) = Biio. If wi! = 0, then there is no such path, and BijO = 0. Suppose, by way of inductive hypothesis, that for 1 5 ij I n and for some (k-l) -c n, (i) if wi’f-’ > o then there is a loop-free path r$-‘, from i to j with each intermediate node at most k-l, whose weight is w,$-l, and (ii) B,!j-l = ~4~~. If w,$-’ > o then there is a path 7 from i to j whose weight is w$. y is such that either (A) each intermediate node of 7 is at most (k-l), or (B) y goes from i to k; from k to k some number of times, then from k to j, with each intermediate node of each subpath being at most (k-l). In case (A) it is clear that wlf = WI:-‘, and the induc- tive step for part (i) of the Lemma is completed with k- -f?I k-‘. - Y,, In case (B), it follows from our induction hypothesis that there exist loop-free paths r,kk-‘, r&‘, rkk,-’ with weights wi-‘, wrtk- ’ , wkJ k-’ respectively. Let w = min(w,i-‘,wl,-‘) and w’ = w,&‘, and consider the sub- cases (Bl) in which y goes from k to k zero times, and (B2) in which y goes from k to k one or more times. In (Bl) the weight of y is clearly W, while in (B2) it is min(w ,ti). Hence, from the definition of WI:, we have wif = max(w ,min (w ,ti)), which is simply w. So part (i) of the Lemma holds with 7,: = r,kk-’ riJ-‘. From part (ii) of the inductive hypothesis, and from Step 2 of the MAX- MIN algorithm, it follows that B; = max(wJ-‘,w). So Bk = max(wk-’ wkj = wk since it dlfinition zf k{that ‘z; zz w,:-‘. follows from the So in either of the cases (A) and (B) B[ = w,f, which establishes part (ii) of the Lemma for the case W; > O. If w; = 0 then there is no path from i to j. Suppose Bt f 0. Then either w,t-’ # 0, or both of w;-‘, wiJ-’ are nonzero. In each case there is a path from i to j, a con- tradiction. So if wlf = o then B: = w:. q HI PI [31 [41 [51 h5 REFERENCES Ng, S. W., and A. Walker “Max-min Chaining of Weighted Assertions is Loop-free”, CBM-TR-73, Dept. of Comp. Sci., Rutgers University, 1977. Shortliffe, E. Computer Based Medical Consulta- tions: MYCIN. American Elsevier, 1976. Van der Mude, A. and A. Walker “On the Infer- ence of Stochastic Regular C-7dmmars” Information and Control, 38:3 (1978) 310-329. Walker, A. “A Framework for Model Construc- tion and Model Based Deduction in a System with Causal Loops” In Proc. Thi ,i Illinois Conf. Med. Info. Syst., 1976. Walker, A. “On the Induction of a Decision Mak- ing System from a Data Base”, CBM-TR-80, Dept. of Comp. Sci., Rutgers University, 1977. Weiss, S. “Medical Modeling and Decision Mak- ing”, CBM-TR-27, Dept. of Comp. Sci., Rutgers University, 1974. 107

1980

15

7

Applying General Induction Methods to the Card Game Eleusis Thomas G. Dietterich Department of Computer Science Stanford University Stanford, CA 94305 Abstract Research was undertaken with the goal of applying general universally-applicable induction methods to complex real-world problems. The goal was only partially met. The chosen domain-the card game Eleusis-was still somewhat artificial, and the universally-applicable induction methods were found to be lacking in important ways. However, the resulting Eleusis program does show that by using knowledge-based data interpretation and rule evaluation techniques and model-fitting induction techniques, general induction methods can be used to solve complex problems. Introduction Work in the area of computer induction is characterized by a continuum from general, universally-applicable methods [5, 6, 7, 9, 10, 121 to specific, problem-oriented methods [2, 8, 111. The general-purpose methods have been criticized for lacking the power to operate in real-world domains. Problem-oriented methods have been criticized for being too specialized to be applied to any problems outside their original domains. This paper describes an attempt to bridge this gap by applying general-purpose induction algorithms to the problem of inducing secret rules in the card game Eleusis. Further details are available in [3]. A Program for Eleusis Eleusis (developed by Robert Abbott [l, 41) is a card game in which players attempt to induce a secret rule invented by the dealer. The secret rule describes a linear sequence of cards. In their turns, the players attempt to extend this sequence by playing additional cards from their hands. The dealer gives no information aside from indicating whether or not each play is correct. Players are penalized for incorrect plays by having additional cards added to their hands. The game ends when a player empties his hand. A record of the play is maintained as a layout (Figure 1) in which the top row, or mainline, contains all of the correctly-played cards in sequence. Incorrect cards are placed in side lines below the main line card which they follow. mainline: 3H QS 4C JD 2C 10D 8H 7H 2C 5H sidelines: JD AH AS IOH 5D 8H 10s QD Rule 1: “If the last card is odd, play black, if the last card is even, play red.” Figure 1. Sample Eleusis Layout (after El]). This research sought to develop a program which could serve as an intelligent assistant to a human Eleusis player. The program needed to be able to: ) discover rules which plausibly describe the layout, ) accept rules typed by the user and test &hem against the layout, ) extend the layout by suggesting cards to be played from the player’s hand. Although Eleusis is artificial and noise-free, it is sufficiently complex to provide a reasonable test bed for inductive techniques. The development of an intelligent assistant required not only basic induction methods but also extensive deduction techniques for testing rules and extending the layout. Problems with Existing Induction Methods While designing the rule-discovery portion of the Eleusis program, existing induction algorithms [5, 6, 7, 9, 10, 121 were examined and found to be lacking in three fundamental ways. The first major problem with some of these algorithms is their emphasis on conjunctive generalizations. Many Eleusis rules are disjunctive. For example, Rule 1 can be written as: tli {odd(cardi-1) A black(cardi) V even(cardi-1) A red(cardi)} The second major problem with these algorithms is that they make implicit assumptions concerning plausible generalizations-assumptions which are not easily modified. Of the algorithms examined, only Mitchell’s version space algorithm [lo] maintains information concerning all rules consistent with the data (and his algorithm is still oriented toward conjunctive generalization). The algorithms of Hayes-Roth and Vere both seek the most specific rule consistent with the data, while Michalski’s Aq algorithm seeks a disjunctive description with the fewest conjunctive terms. In contrast, the plausibility heuristics for Eleusis are: Choose rules with intermediate degree of generality. (Justification: the dealer is unlikely to choose a rule which is overly general because it would be too difficult to discover. Conversely, overly specific rules are easily discovered because they lead to the creation of numerous counter-examples during play.) Choose disjunctive rules based on symmetry. (Justification: Rule 1 is an excellent example of a symmetric disjunctive rule. Most often in Eleusis, the terms of a disjunction define mutually exclusive cases which have some symmetric relationship to each other. The dealer is very likely to choose such rules because they are not too hard-nor too easy--to discover.) (These plausibility heuristics are based on the assumption that the dealer is rational and that he is attempting to maximize his own score (according to the rules of the game). This is an artificial assumption. It is very rare in science that we have such insight into nature. However, in all domains plausibility criteria must be available-otherwise, we don’t know what we are searching for.) The third major problem with using general-purpose induction techniques in Eleusis is that the raw data of the Eleusis layout are not in a form suitable for generalization. (Many researcliers [2, 111 have pointed out this problem in other domains.) One aspect of this problem is evident in Rule 1: neither color nor parify is explicit in the representation of the cards. Another difficulty is that the sequential ordering of the cards is implicit in their position in the layout. It must be made explicit in order to discover rules like Rule 1. Two techniques were developed to address these problems. First, in order to avoid an exhaustive search of rule space and at the same 218 time avoid the “tunnel vision” of existing algorithms, rule models were developed to guide the induction process. Secondly, in order to transform the input data info a form appropriate for generalization, a series of knowledge-based processing layers were created. Induction by Model-Fitting By analogy with traditional statistical time-series analysis, the program uses a model-fitting approach to induction. The term model denotes a syntactic or functional skeleton which is fleshed out by the induction algorithms to form a rule. In traditional regression analysis, for example, the model is the regression polynomial whose coefficients must be determined by induction from the data. Properly chosen models can strongly constrain the search required for induction. After looking at several Eleusis games, the following models were designed for the Eleusis program: B Decomposition. This model specifies that the rule must take the form of an exclusive disjunction of if-lhen rules. The condifion parts of the rules must refer only to cards prior to the card to be predicted. The action parts of the if-then rules describe correct plays given that the condition parts are true. The condition parts must be mutually exclusive conjunctive descriptions. The action parts are also conjunctions. Rule 1 fits the decomposition model: Vi odd(cardi-1) * black(cmdi) V even( cardi- 1) =P red(cardJ ) Periodic. A rule of this model describes the layout as a periodic function. For example, Rule 2 (Figure 2) is a periodic rule. The layout is split into phases according to the length of the proposed period. The periodic model requires that each phase have a conjunctive description. JC 4D QH 3s QD 9H QC 7H QD 9D QC 3H KC 5s 4s 10D 7s M phase 0: JC QH QD QC QD QC 5s 4s 1OD 7s phase 1: 0 4D 3s 9H 7H 9D 3H KC Rule 2: (periodic rule with length 2): phase 0: Vi faced(cardi) phase 1: tli nonfaced(cardi) Figure 2. A Periodic Rule. ) Disjunctive Normal Form (DNF) with fewest terms. The Aq algorithm (Michalski [9]) is used to discover rules which have the fewest number of separate conjunctions. The Aq algorithm was given heuristics to guide it towards symmetric, disjoint disjunctive terms. By definition, not all Eleusis rules can be represented using these three models. But, these models, when combined with segmentation (see below), cover all but one or two of the Eleusis rules which I have seen. For each of these models, an algorithm was developed to fit the data to the model. In order to fit the data to the decomposition model, the program must determine which variables to decompose on, i.e. which variables to test in the condition part of the rule (Rule 1 decomposes on parity E {odd, even)). The program must also decide how far into the past this decomposition should apply (i.e. do we look at just the most recent card, or the two most recent cards, . . . . etc.). Once the decomposition variables and the degree of lookback are determined, the algorithe must find a conjunctive description for the action parts of the rules. The program uses a generate-and-test approach. First, it considers rules which look back only one card, then two cards, and so on until a rule consistent with the data is found. To determine the 219 decomposition variable(s), it generates trial rules by decomposing on each variable in turn and chooses the variable which gives the simplest rule. If the resulting rule is not consistent with the data, the layout is decomposed into sub-layouts based on the chosen variable, and a second decomposition variable is again determined by generating trial decompositions and selecting the simplest. This process is repeated until a rule consistent with the data is found. (This is a beam search with a beam width of 1). In order to fit the periodic model, the program chooses a length for the period, splits the layout into separate phases, and finds a conjunctive description of each phase. Since the rule is more plausible if the descriptions of each pha’se are mutually exclusive, the algorithm attempts to remove overlapping conditions in the descriptions of the different phases. Again, a generate-and-test approach is used to generate periodic rules with different length periods (from length 1 upwards) until an acceptable rule is discovered. The Aq aigorithm is used to fit dati to the DNF model. Knowledge-layer Structure Like many other AI systems, the Eleusis program is structured as a set of layers, or more accurately, rings, based on the kinds of knowledge used to solve the problem (Figure 3). Each layer takes input data from the outside, transforms the data using knowledge appropriate to this layer, performs generalization by searching the space of rules at this level of abstraction, and evaluates the obtained rules. Rules which are sufficiently plausible are returned to the outer layers. Each layer calls the inner layers to perform tasks requiring knowledge appropriate to those inner layers. Figure 4 shows the five layers of the Eleusis program. Notice that in Eleusis, the outermost ring is very specific to Elcusis, while the inner-most rings contain only the general model-fitting induction algorithms. This design is intended to allow the program to be easily applied to similar problems. Since all Eleusis-specific knowledge is in the outer-most two layers, these could be stripped off and replaced by layers which apply different kinds of data transformations to solve different problems (e.g. letter series completion, sequence extrapolation). Figure 3. The Knowledge-layer Scheme. 5 User Interface 4 Eleusis KnowledPe , ; Seg~ Sequential Analysis 1 Basic induction Most Specific CL Most General Figure 4. Layered Structure of Eleusis Program. The five layers fiinction as follows. The outer-most layer provides an interface for the user. Layer 4 transforms the layout by making explicit such things as the color, parity, prime-ness, and faced-ness of the cards. Layer 3 segments the layout. Segmentation is used to discover rules such as Rule 3 (Figure 5) which involve first splitting up the layout into segments according to some criterion (like constant color) and deriving a new layout based on the lengths of these segments. Layer 2 makes the order of the events explicit [4] Gardner, Martin, “On Playing the New EIeusis, the game that either by creating “first difference” variables (e.g. Avalue(caQ = simulates the search for truth,” Scientific American, 237, value(c&$ - value(cardi_l)) or by breaking the layout into separate October. 1977, pp 18-25. phases (for periodic rules). The result of the preprocessing of layers 5 through 2 is that layer 1 is called with a specific model for which [S] Hayes-Roth, F:; J. McDermotf “An Interference Matching Technique for Inducing Abstractions”, Communicufions of fhe the degree of Zookback and (optionally) length of period have been specified and with a set of unordered events to which the model is to be fitted. Layer 1 actually performs the model-fitting using one of the three model-fitting induction algorithms. Once the rules have been induced, they are passed back up through the layers for evaluation. Layer 2 evaluates the rules using knowledge about ordering (e:g. guaranteeing that the rule doesn’t lead to a dead end). Layer 3 checks that the rules are consistent with the scgmcntation it performed (in particular, the boundary values cause some problems). Layer 4 evaluates the rules according to the heuristics for plausible rules in Elcusis. Finally, layer 5 prints any rules which survive this evaluation process. ACM, 21:5, 1978, pp. 401-410. [6] Hunt, E.B.. Experiments in Induclion, Academic Press, 1966. f’7] Larson, J., “Inductive Inference in the Variable Valued Predicate Logic System VL21 : Methodology and Computer Implementation”, Rept. No. 869, Dept. of Comp. Sci., Univ. of III., Urbana, May 1977. [8] Lenat, D., “AM: An Artificial Intelligence Approach to Discovery in Mathematics as Heuristic Search,” Comp. Sci. Dept., Rept. STAN-CS-76-570, Stanford University, July 1976. 191 Michalski. R. S., “Algorithm Ag for the Quasi-Minimal Solution of the Covering Problem,‘* Archiwum -Automafyki i Telemechaniki, No. 4, Polish Academy of Sciences, 1969 (in Polish). [lo] Mitchell, T. M., “Version Spaces: an Approach to Concept Learning,” Comp. Sci. Dept. Rept. STAN-CS-78711, Stanford University, December 1978. I AH 7C 6C 9s 10H 7H 1 OD JC AD 4H 8D 7C KD 6s QD 3s I JH I Rule 3: “Play odd-length strings of cards where color is constant within each string.” 1 The segmenled layout looks like this (color, length): (red, 1) (black, 3)’ (red, 3) (black, 1) (red, 3) Figure 5. A Segmentation-based Rule. The program works well. The three rule models, when combined with segmentation, span a search space of roughly 1O183 possible rules (several control parameters affect the size of this space). The program generates and tests roughly 19 different parameterizations of the three models in order to choose three to five plausible rules. It runs quite quickly (less than seven seconds, on a Cyber 175, in the worst case so far). The rules developed are similar to those invented by humans playing the same games (15 complete games have been analyzed). Conclusion General induction techniques can be used to solve complex learning tasks, but they form only part of the solution. In the Eleusis domain, data interpretation, rule evaluation, and model-directed induction were all required to develop a satisfactory program, A degree of generality was obtained by segregating the functions of the program into layers according to the generality of the knowledge they required. This should allow the program to be applied to similar tasks merely by “peeling off’ and replacing its outer layers. Acknowledgments Thanks go to R. S. Michalski, my M.S. thesis advisor, for suggesting Eleusis as a domain and for providing numerous ideas including the basic idea for the decomposition algorithm. Thanks also to David Shur for proofreading this paper. NSF grant no. MCS-76-22940. This research was supported by References [l] Abbott, Robert, “The New Eleusis,” Available from Abbott at Box 1175, General Post Office, New York, NY 10001 ($1.00). [2] Buchanan, B.G., D. H. Smith, W. C. White! R. J. Gritter, E. A. Feigenbaum, J. Lederberg, C. Djerassr, Journal of the American Chemical Society, 98 (1976) p. 6168. [ll] Soloway, E., “Learning = Interpretation + Generalization: a case study in knowledge-directed learning,” PhD Thesis, COINS TR 78-13, University of Massachusetts, Amherst, MA., 1978. [12] Vere, S. A., “Induction of Relational Productions in the Presence of Background Information,” In Proceedings of the Fifrh International Joint Conference on Artificial Intelligence, MIT, Cambridge, MA., 1977. [3] Dietterich, Thomas G., “The Methodology of Knowledge Layers for Inducing Descriptions of Sequentially Orderid Events,” MS Thesis, Dent of Corn. Sci., Univ. of Illinois, Urbana, October, 1979. - 220

1980

16

8

MODELLING STUDENT ACQUISITION OF PROBLEM-SOLVING SKILLS Robert Smith Department of Computer Science Rutgers University New Brunswick, N. J. 08903 ABSTRACT This paper describes the design of a system that simulates a human student learning to prove tneorems in logic by interacting with a curriculum designed to teach those skills. The paper argues that sequences in this curriculum use instructional strateu, and that the student recognizgs these strategies in driving the learning process. I. INTRODUCTION A central issue in the design of learning systems (LS's) is the classification of the sources of information that the system uses for its acquisition. The general notion is that an LS begins with certain knowledge and capabilities, and then extracts information from training sequences or experiments in the acquisition process. Implicit within this general characterization is the idea that the LS enforces some kind of interpretation on the training sequence by way of driving the acquisition process. Often the nature of the interpretation is left implicit. An example of an LS that makes somewhat explicit the interpretation of its training sequence is Winston's program for learning structure descriptions, where the program interprets the near m example as providing key information about the structure being learned [lr]. We speculate that much human learning takes place in a more richly structured environment, wherein the human learner is interpreting the instructional seauences provided to him in a richer way than LS's typically envision. Indeed, most LS's nave made few if any explicit assumptions about the structure of the environment in which the training sequence occurs. One particularly rich environment is learning & teaching. We suggest that teachers use certain instructional strategies in presenting material, and that students recognize these strategies. This paper describes the motivation for an LS called REDHOT. REDHOT is a simulation of a student -------- * I would like to thank Phyllis Walker for analysis of the curriculum and student protocols; Saul Amarel, Tom Mitchell, Don Smith, and N. Sridharan for many ideas and assistance. The research reported here is sponsored by the Office of Naval Research under contract N00014-79-C-0780. We gratefully acknowledge their support for this work. acquiring the skill of constructing proofs in elementary logic. We characterize this skill as consisting of (1) primitive operators in the form of natural-deduction rules of inference, (2) "macro moves" consisting of several rules of inference, and (3) application heuristics that describe when to use the rules. The central theme of this research is to model the acquisition of these skills around the recognition of instructional strategies in a curriculum designed to teach the student. II. CURRICULUM FOR REDHOT We are using the curriculum from the computer- assisted instruction (CA11 course developed at Stanford University by Patrick Suppes and co- workers. (See [3] for details.) This CA1 system is used as the sole mode of instruction for approximately 300 Stanford students each year. We chose a CA1 curriculum because we thought that the explicitness inherent in a successful CA1 system developed and tested over a number of years might make the instructional strategies relatively clear. The curriculum contains explanatory text, examples, exercises, and hints. The explanatory text is rather large, and uses both computer- generated audio and display as modes of presentation. The presentation strategy used by the actual CA1 system is linear through the curriculum. For use with REDHOT, we have developed a stylized curriculum in an artificial language CL. It contains the examples, exercises, partially formed rules, and hints. The exercises are the central part of the curriculum. There are approximately 500 theorems that the student is asked to prove, with about 200 in propositional logic, 200 in elementary algebra, and 100 in the theory of quantification. The human student performs these exercises by giving the steps of the developing proof to an interactive proof checker. This proof checker is the heart of the original instructional system. We developed a version of this proof checker for use with the REDHOT student simulation. III. THE DESIGN OF REDHOT REDHOT learns rules for proving theorems. These rules are initially the natural deduction 331 rules of many logic systems. The student improves upon these rules by building macro operators and by adding heuristics to existing rules--i.e., giving strategic advice in the left-hand-sides of the production rules.* For example, the rule AA ("affirm the antecedent", the system's version of modus ponens) can be stated as the following rule: Rule AA I GOAL: Derive Q I Prerequisites: f P already on some line i I P -> Q on some line j I Method: f AA command on lines i and j I Heuristics: None (yet) I Effectiveness: Perfect I In the above, we have adopted a style of rule presentation that is similar to the rules of a production system. The letters P and Q stand for arbitrary formulas, and i and j for arbitrary lines of the already existing proof. The aoal tells what the rule will produce. The mreauisites tell us that two formulas of the appropriate forms are needed, and the method gives the schematic command to the proof checker that will accomplish this. The heuristics associated with a rule are learned by the system, and indicate how the rule should be used. Effectiveness of the rule is also learned, and indicates how effective the rule will be in achieving its goal given its prerequisites. The effectiveness of this rule is "perfect" since the rule is given as such in the curriculum. The underlying problem solver for REDHOT is a depth-first, heuristic problem solving over the existing rules. It is assumed that: (a) the system has sufficient memory and CPU facilities for this search, if necessary; and (b) that the underlying pattern matchers are sufficient to match arbitrary formulas, line numbers, etc. Both of these assumptions are open to criticism on the grounds of being psychologically unrealistic. One of the goals of the construction of REDHOT is to decide on plausible ways to restrict the problem solver and pattern matcher to make a more realistic system. REDHOT learns the heuristics for the application of the rule. These heuristics are stated in a heuristic-language HL, which is strongly tied to the curriculum language CL. The heuristics associated with a rule guide the student as to when to try that rule or not to try it. For example, the curriculum appears to teach the student that the AA rule is a plausible thing to try when the prerequisites are available -------- * See [II for a conceptual discussion of the levels through which this process might proceed. One way to regard this research is a suggestion of the mechanism for this acquisition of heuristics and macro moves. (whether or not the goal is immediately desired). This is one of the primitives of the HL heuristics language. An example of a macro operator that is not a primitive rule is the "multiple AA" macro move. A paraphrase of this macro operator might be: I Multiple-AA Macro Move I I IF you want to prove Q : I AND I I have P, P -> P1, PI -> P2, . . . . Pn -> Q I : THEN I I make multiple AA applications, I I which is guaranteed to succeed I We discuss below t.he training sequence that teaches this macro move. IV. m RECOGNITION @ INSTRUCTIONAL STRATEGIES REDHOT bases its acquisition of application heuristics and macro operators on its recognition of instructional strategies in the training sequence. For example, consider the sequences of exercises in Figure I, taken from the actual curriculum. The sequence, which is at the beginning of the whole curriculum, gives the primitive rule of inference AA, then shows successive elaborations of the use of that rule of inference. REDHOT detects this to be a use of a strategy called focus a& elaborate, in which a rule is first focussed, and then a particular elaboration is given. Teacher: Here is a rule called AA. Teacher: Here are some exercises involving AA: 1. Goal: Q Premises: S -> Q, S 2-5 [Several more exercises with different formulas. 1 6. Goal: W Premises: S -> Q, Q -> W, S 7. Goal: S Premises: R -> S, Q -> W, W -> R, Q 899 [Two more similar exercises involving multiple applications of AA.] Figure _1_ Sequence of Exercises for Learning Multiple Application of AA Command In the above training sequence, REDHOT takes steps l-5 as focussing on the AA rule, and steps 6- 9 as providing a useful elaboration of that rule, in the form of the macro operator for multiple application. 222 A second example of the use of an instructional strategy concerns removing possible bugs in learned heuristics and macro operators. We illustrate this with the macro rule for conditional proof, a common strategy in logic and mathematics, which we paraphrase as follows: I Conditional Proof MACRO MOVE I 1 IF you want to prove P -> Q I I THEN I I ASSUME P as a working premise I I PROVE Q (likely from P); I I APPLY "CP" rule, removing premise I I The actual instructional sequence goes to great length to teach this principle, and students typically have a difficult time with the principle; one defective version of the rule that students seem to learn is the following: I "Defective" MACRO MOVE I I IF you have a formula P -> Q I I AND you want to prove Q I I THEN I I ASSUME P as a working premise; I I PROVE Q using AA; I This is a very poor strategy; but the evidence suggest that over half of the students learn it. The following exercise seems to help students debug the rule "Defective" . Derive: S -> (Q OR R) Premise: (R -> R) -> (Q OR R) In examining student protocols, we see that many students will try several times to apply the "Defective" rule to this exercise. Finally, (we speculate) they realize that (R -> R) is already something that they know how to prove, using a previously learned macro operator. Then, the actual proof becomes obvious, the student corrects the defective rule, and goes on to handle similar exercises correctly. We call this instructional strategy "focus and frustrate" wherein a student discovers --somewhat traumatically--that a rule he learned is defective. Therefore, an exercise such as the above is not just randomly selected, but instead tests possible student "bugs" in an exact way. Notice that it is one of the simplest exercises that will discriminate between the correct and defective formulations of the macro rule for conditionaL proof. (See [2] for a discussion of "debugging" student skills.) to V. REDHOT AND LEARNING SYSTEMS Like many LS's, REDHOT starts wi th the ab ility state everything it will "learn" I in some sense at least. The initial rules for solving the problem (the natural deduction rules for logic) are complete with respect to the underlying problem solver --unless it is restricted in time/space (in practice it is). The heuristic and macro languages are also given in advance, and they of course define a space of the possible rules that might be learned. So, the object is to select among heuristics and macro rules in this space. One way to formulate doing this is by experimentation or exploration. REDHOT selects objects from this meta-space by being taught. Learning by being taught consists of the "teacher" laying out exercises in an organized and structured way, and the student recognizing something of that structure. The student makes-- believes that he is entitled to make--fairly bold hypotheses about the rules he is learning, and relies on the training sequence to contain exercises that will check for common errors that he the student may have made in formulating these rules. REDHOT compares somewhat radically to many LS's that rely on a somewhat slow, computationally coherent delimitation of the rule (or concept) involved. We speculate that learning by "discovery" or "experimentation" is a slow process for even humans, done over the eons of time and through social interaction. Most human learning is by being taught, and one can argue that AI should give attention to the relation between learning and teaching, in terms of modelling the acquisition of concepts, problem-solving skills, and natural language. We further speculate that learning by "discovery" will be aided by extracting as much information as possible from the structure of the environment in which the LS operates. IllI L-21 [31 II41 REFERENCES Amarel, Saul, "An Approach to Problem Solving and Theorem Proving in the Propositional Calculus", in Svstems and Computer Science, (Hart and Takasu, eds.), Toronto: University of Toronto Press, 1967. Brown, John Seely, Burton, Richard R., and Larkin, Kathy M., "Representing and Using Procedural Bugs for Educational Purposes", in Proceedings of the 1977 Annual Conference of the Association for Computing Machinerv, New York, 1977. Suppes, P., Smith, R. L., and Beard, M., "University-level Computer-assisted Instruction at Stanford: 1975", in Instructional Science, 1977, 4, 151-185. Winston, Patrick Henry, "Learning Structural Descriptions from Examples", Ph.D. thesis, in The Psvchology of Commuter Vision, (Patrick Henry Winston, ed.), McGraw-Hill, New York, 1975. 223

1980

17

9

A Computer Model of Child Language Learning Mallory Selfridge Yale University Abstract A computer program modelling a child between the ages of 1 and 2 years is described. This program is based on observations of the knowledge this child had at age 1, the comprehension abilities he had at age 2, and the language experiences he had between these ages. The computer program described begins at the age 1 level, is given similar language experiences, and uses inference and learning rules to acquire comprehension at the age 2 level. Introduction This paper describes a computer model of the development of comprehension abilities in a child, Joshua, between the ages of one and two years. The program begins with the kind of knowledge that Joshua had at age 1, when he under- stood no language, and learns to understand com- mands involving action, object, and spatial rela- tion words at Joshua's age 2 level. It does so by being given the kind of language experiences Joshua had between the ages of 1 and 2, and making use of rules to 1) infer the meaning of utter- ances, 2) attend to words, and 3) learn language meaning and structure. The program passes through a reasonable developmental sequence and makes the same kind of errors that children make at inter- mediate stages. This work suggests that language learning to the 2 year old level can be accounted for primari- ly by the learning of word meaning and structure, that world knowledge is crucial to enable the child to infer the meaning of utterances, and that children hear language in situations which enable them to perform such inferences. The success of the program in modelling Joshua's language development -- both its progression and its errors -- suggests that it embodies a plausible theory of how Joshua learned to understand language. While there are several aspects of the model which are unrealistic (for example, segmented input, no am- biguous words, no simultanious conceptual develop- ment), there is reason to believe that future work can sucessfully address these issues. Further de- tails can be found in Selfridge (1980). This paper first considers Joshua's initial state of knowledge at age 1, and then his comprehension abilities at age 2. It describes the kind of language experiences he had, and several kinds of learning rules which can account for Joshua's development. The computer program incor- porating these observations and rules * described, and finally some conclusions ai: presented. Joshua's Initial Knowledge The first component of --- a computer model of the development of Joshua's comprehension is Joshua's knowledge prior to his language learning. Observations like the follow- ing suggest that Joshua had considerable knowledge of objects, actions, spatial relations, and ges- tures at age 1 (ages are given in YEARS:MONTHS:DAYS): 0:11:19 Joshua and I are in the playroom. I build a few block towers for him to knock down, but he doesn't do so; rather, he dismantles them removing the blocks from the top, one at a time 1:0:16 Joshua and I are in the playroom . Joshua takes a toy cup, and pretends to drink out of it. 1:2 Joshua is sitting in the living room playing with a ball. I hold my hand out to him, and he gives me the ball. The above observations show that Joshua knew the properties and functions of objects like blocks, cups and balls. He knew actions that could be per- formed with them, and various spatial relations that they could enter into. Finally, he knew that behavior can be signaled through gestures by other people. Thus, a language learning program must be equipped with this kind of knowledge. Joshua's Comprehension Abilities at Age2 At age 2, Joshua could respond correctly to commands with unlikely meaning and structure. His correct responses suggests full understanding of them. For example, consider the following: 2:0:5 We walk into the living room and Joshua shows us his slippers. His mother says "Put your slippers on the piano." Joshua picks up the slippers and puts them on the piano keys, looking at his mother. She laughs and says "Thats silly." Joshua removes the slippers. The meaning of this utterance is unlikely since slippers do not generally go on piano keys, and piano keys don't generally have things put on them. His response suggests that he was guided by full understanding of the meanings of the words in "Put your slippers on the piano." At age 2 Joshua also understood language structure, as the following example shows: 2:o:o Joshua and I are in the p layroom, my tape recorder i on the floor in front of me. I say "Get on the tape recorder, Joshua". Joshua looks at me oddly, and looks at the the tape recorder. I repeat "Get on the tape recorder." Joshua moves next to the tape tape recorder. I once more repeat 'Get on the the tape recorder." Joshua watches me intently, and lifts his foot up and slowly moves it over the tape recorder to step on it. I laugh and pull the tape recorder away. It seems that Joshua understood "Get on the tape recorder" the first time I said it, and that his reluctance to comply reflected his knowledge that what I was asking was very unlikely. That is, Joshua understood that the tape recorder was the object to be underneath him, although this is un- likely given his experience with it. This, in turn, suggests that Joshua understood the struc- ture of the word "on", namely, that the word whose meaning is the supporting surface follows "on". Thus a program modelling Joshua at age 2 must understand utterances using language structure. Joshua's Language Experiences In the year between -- the ages of 1 and 2, Joshua experienced situations which allowed him to make inferences concerning the utterances he heard. In this section, three examples of such situations are given, and infer- ence rules accounting for Joshua's response and attention to words are presented. In the first example, I am using an utterance and simultaniously signalling the meaning of that utterance through gestures: 1:2:17 We are sitting in the living room, Joshua is holding a book. I look at Joshua, maintain eye contact for a moment, hold my hand out to him and say "Give me the book, Joshua." Joshua holds the book out to me. In this situation, Joshua probably inferred that the meaning of "Give me the book, Joshua." was the same as that signalled by the gestures. The fol- lowing rule captures this idea: Gestural Meaning Inference If an utterance is accompanied by gestures with associated meanings then infer that the the utterance means the same as the gestures. Knowledge of object function and properties helped Joshua infer responses in other situations. In the following, Joshua used his knowledge that books can be opened in his response: 1:0:9 Joshua has a book in his hand, and is looking at it, turning it over, and examining it. His mother says 'open the book, open the book..." Joshua opens the book. She says, "Good Joshua, good." A rule summarizing this inference is the follow- ing: Function/Property Inference If an utterance is heard while interacting with an object then the meaning of the utterance involves a function or property of that object. Parent speech to children posesses many attention-focussing characteristics (e.g. Newport, 1973). The following example is typical: 1:18:0 Joshua's father is trying to demonstrate that Joshua knows the names of the upstairs rooms, and has put a toy lawnmower in the bathroom. He says "Where is the lawnmower, Josh? Its in the BATHROOM. The LAWNMOWER is in the BATHROOM. BATHROOM!" Joshua's attention to "bathroom" in this example can be explained by the following rule: Attention Inference If a word is emphasised, repeated, or said in isolaytion, then attend to it. These are the kind of rules which I postulate enabled Joshua to infer the meaning of utterances from context, and attend to part of the utterance. The program must be equipped with such rules and must be given input in similar contexts. Learning Rules This section will consider Joshua's learning of action, object, and relation words, and language structure. It presents accounts of how Joshua might have learned each of these. Most of the rules have their roots in the learning strategies proposed by Bruner, Goodnow, and Austin ('956). One way Joshua learned the names of ob- jects is by having them named for him, as in the following example: 1:0:0 Joshua is crying. His mother picks him up and goes over to the refrigerator. She gets some juice, holds it up, and asks, "Do you want some JUICE?" Joshua keeps crying. She gets a banana and asks, "Do you want some BANANA, Joshua?" Joshua reaches for it. The following rule models Joshua's ability to learn by having objects named: Direct Naming Inference If a word and an object are both brought to attention, infer the word is the object's name. This rule, and other object word learning rules, can account for how Joshua learned object words such a "slippers", "piano", "ball", and "table". Action words can be learned via inferences about other known words in the utterance. In the following example, summarized from Schank and Sel- fridge (1977), Hana could have inferred the mean- ing of "put" based on her knowledge of the mean- ings of "finger" and "ear." (age 1) Hana knows the words "finger" and "ear", but not "put." She was asked to "Put your finger in your ear," and she did so. The following two rules can account for learning "Put" in situations like this. The first suggests that "put" would initially be learned as "put something in something else." The second, applied after the first in a slightly different situation, would refine the meaning of "put" to "put some- thing someplace". Response Completion Inference Infer the meaning of an unknown word to be the meaning of the entire utterance with the meanings of the known words factored out. Meaning Refinement Inference If part of the meaning of a word is not part of the meaning of an utterance it occurs in, remove that part from the word's meaning. Rules like the above can account for Joshua learning action words like "put", "bring", "give", and so on. However, they can also account for Joshua learning relation words, such as "on" and “in”. If Joshua knew "put", "ball", and "box", say, and was asked to "put the ball in the box", these rules would account for his learning that "in" referred to the "contained" relation. These, then, are the sort of rules the pro- gram uses to learn word meanings. The program's rule for learning language structure is more direct. It is based around the two structural predicates, PRECEDES and FOLLOWS, which relate the positions of words and concepts in short-term memory. This rule models Joshua's acquisition of structural information upon hearing utterances he understands, and appears below: Structure Learning Rule If a slot filler occurs preceding or fol lowing a word' s meaning then update the word's defini .tion that information. This rule accounts for Joshua learning that the filler of the VAL slot of "in"'s meaning -- (CONT VAL (NIL)) -- is found FOLLOWing "in" in the utterance. The Program This section presents four excerpts from a run of the program, written in LISP on a DECSYSTEM-20. Each represents the program at a different stage in development as it progresses from Joshua's age 1 abilities to Joshua's age 2 abilities, using the inference rules described previously. The knowledge representation used is Conceptual Dependancy (Schank, 19731, and the language understanding process embedded in the program is similar to that in Birnbaum and Sel- fridge (1979). The first stage of the program corresponds to Joshua at age 1. At this stage, the program had only the knowledge ascribed to Joshua at that age. In the excerpt below, the "parent" types a lower- case utterance to the program, and the program responds with a message stating its lack of under- standing. When the parent provides gestures via simulated visual input, however, the program understands, and prints the CD representation of its response. IPARENT SAYS: give me the ball ICHILD STARES BLANKLY AT PARENT ICHILD RETURNS TO PLAY ICHILD sms: (PARENT HOLDS OUT HAND) (PARENT ~0oKs AT BALLS) ICHILD INFERS RESPONSE USING RULE: I GESTURAL MEANING ICHILD RESPONDS: (ATRL~NS ACTOR (CHILD) I OBJECT (BALLS) TO (~0~s VAL (PARENT)) In the second stage, shown in the excerpt below, the program has learned the meaning of several words, and understands some utterances correctly. In this case, it has learned the words "put", "ball", and "box". However, notice that although it responds correctly to the first utter- ance given by the parent, it misunderstands the second. This sort of error is reported in Hoogen- raad et al. (1976). Not knowing "on", the program incorrectly infers that the appropropriate rela- tionship is containment. IPARENT SAYS: put the ball in the box I ICHILD 11wERs RESPONSE USING RULE: I UTTERANCE UNDERSTANDING, FUNCTION/PROPERTY ICHILD RESPONDS: (PTRANS ACTOR (CHILD) I OBJECT (BALLS) TO (CONT VAL (BOXl)) I PARENT SAYS: put the ball on the box I iCHILD INFERS RESPONSE USING RULE: I UTTERANCE UNDERSTANDING, FUNCTION/PROPERTY ICHILD RESPONDS: (PTRANS ACTOR (CHILD) OBJECT (BALLS) TO (CONT VAL &0X1)) The transition from the second stage to the third is accomplished by teaching the program more words. In this case it has learned the additional words "slippers", "on", "piano", “ball”, and “table. ” At this stage, the program can now understand "Put the slippers on the piano", whereas at any earlier stage it would not have. The program also prints out a message showing that it recognizes this as an unusual request. However, although this stage represents Joshua's age 2 understanding of word meaning, the program has not yet learned language structure. The program interprets the second utterance in- correctly, however, in accord with its knowledge of the usual relationships between objects. This sort of error is similar to that reported in Stroher and Nelson (1974). IPARENT SAYS: put the slippers on the piano ICHILD LOOKS AT PARENT STRANGELY I CHILD INFERS RESPONSE USING RULE: I UTTERANCE UNDERSTANDING ICHILD RESPONDS: I (PTRANS ACTOR (CHILD) I OBJECT (SLIPPERS~) TO (TOP VAL (~1~~01))) IPARENT SAYS: put the table on the ball I iCHILD INFERS RESPONSE USING RULE: I UTTERANCE UNDERSTANDING ICHILD RESPONDS: (PTRANS ACTOR (CHILD) I OBJECT (BALLS) TO (TOP VAL (TABLET))) The fourth stage is shown in the excerpt below. The program has now learned the structure of "on", and can hence correctly understand "Put the table on the ball." In addition, it prints out 226 a message indicating its liarity of this reque st. awareness of IPARENT SAYS: put the table on the ball ~CHILD LAUGHS AT UNUSU~~L REQUEST ICHILD INFERS RESPONSE USING RULE: I UTTERANCE UNDERSTANDING ICHILD RESPONDS: (PTRANS ACTOR (CHILD) I OBJECT (TABLET) TO (TOP VAL (BALLS))) the pecu- At the four th stage, the program has success- fully learned to understand a subset of language at Joshua's age 2 level. It began with world knowledge similar to that Joshua began with, was equipped with reasonable learning and inference rules, and progressed as he did by being given language experiences similar to those he experi- enced. Conclusions This paper has described a computer model of a child learning to understand commands involving action, object, and relation words. The program learns language meaning and structure to the level attained by the child at age 2, by being initially given the same kind of knowledge the child had and by being exposed to language in the same kind of contexts as the child did. gram learned language according to a The pro- reasonable progress ion, making the same sort of error s that children do a t intermediate stages. No par ts of speech or traditional grammatical constructions are learned. It also acquires structural knowledge after knowledge of meaning, because no structural knowledge can be-associated with a word until the meaning of that word is learned. This This aspect of the model offers an explanation for why children learn structure following meaning (Wetstone and Friedlander, 1973). In addition to English, the program has been tested on comparable subsets of Japanese, Russian, Chinese, Hebrew, and Spanish. Its performance with these languages was equivalent to its learning of English, suggesting that the program has no English-specific mechan- isms. This research suggests several conclusions. It suggests that a large part of the language learning problem lies in accounting for how the child infers the meaning of the language he hears. It argues that the mechanisms underlying the learning of meaning and structure are the same. It questions the role of traditional grammatical models both in language learning and language understanding, and suggests that models of language learning must be based on strong models of language understanding. In particular, it ques- tions Chomsky's (1980) position that language is not learned. This work suggests that plausible learning models of language development are possi- ble. Further research should proceed in many directions. In particular, the program discussed here should be extended to model the development of comprehension of more complex constructions, such as relative clauses, and the generation of language. Acknowledgements Dr. Roger Schank's assistance in this work was in- valuable. Peter Selfridge provided useful com- ments on this paper. Bibliography Birnbaum, L., and Selfridge, M. (1979). Prob- lems in Conceptual Analysis of Natural Language. Research Report 168, Department of Computer Sci- ence, Yale University. Bruner, J.S., Goodnow, J. J., and Austin, G.A., (1956). A Study of Thinking. John Wiley and Sons, New York - Chomsky, N., (1980). Rules and Representations, excerpted from Rules and Representations. Colum- -- bia University Press, New York Hoodenraad, R., Grieve, R., Baldwin, P., and Campbell, R. (1976). Comprehension as an Interac- tive Process. In R. N. Campell and P. T. Smith (eds.) Recent Advances in the Psychology of Language., Plenum Press, NeFYor - Newport, E.L., (1976). Motherese: the Speech of Mothers to Young Children. in N.J. Castellan, D.B. Pisoni, and G.R. Potts, (eds.) Cognitive Theory: VI II., Lawrence Erlbaum Assoc., Hilsdale, N.J. Schank, R. C., (1973). Identification of Con- ceptualizations Underlying Natural Language. In R. C. Schank and K. M. Colby (eds.) Computer Models of Thought and Language W.H. Freeman and Co., San Fransisco. Schank, R. C., and Selfridge, M. (1977). How to Learn/What to Learn. in Proceedings of the International Joint Conference on Artificial In- telligence, Cambridge, Mass. Selfridge, M. (1980). A Process Model of Language Acquisition. Computer Science Technical Report 172, Yale University, New Haven, Ct. Strohrer, H. and Nelson, K.E., (1974). The Young Child's Development of Sentence Comprehen- sion: Influence of Event Probability. Non-verbal . . Context, Syntactic Form, and Strategies. Child m., 45: 567-576 Westone, H. and Friedlander, (1973). The Ef- fect of Word Order on Young Children's Responses to Simple Questions and Commands. Child Dev 44:734-740 - -0 227

1980

18

10

APPROACHES TO KNOWLEDGE ACQUISITION: THE INSTRUCTABLE PROQUCTION SYSTEM PROJECT Michael D. Rychener Carnegie-Mellon University Department of Computer Science Schenley Park Pittsburgh, PA 15213 Abstract Progress in building systems that acquire knowledge from a variety of sources depends on determining certain functional requirements and ways for them to be met. Experiments have been performed with learning systems having a variety of functional components. The results of these experiments have brought to light deficiencies of various sorts, in systems with various degrees of effectiveness. The components considered here are: interaction language; organization of procedural elements; explanation of system behavior; accommodation to new knowledge; connection of goals with system capabilities; reformulation (mapping) of knowledge; evaluation of behavior; and compilation to achieve efficiency and automaticity. A number of approaches to knowledge acquisition tried within the Instructable Production System (IPS) Project are sketched.* 1. The lnst ructable Production System Project The IPS project [6] attempts to build a knowledge acquisition system tinder a number of constraints. The instructor of the system gains all information about IPS by observing its interactions with its environment (including the instructor). Interaction is to take place in (restricted) natural language. The interaction is mixed initiative, with both participants free to try to influence the direction. Instruction may be about any topic or phenomenon in the system’s external or internal environment. Knowledge accumulates over the lifetime of the system. *This research was sponsored by the Defense Advanced Research Projects Agency (DOD), ARPA Order No. 3597, monitored by the Air Force Avionics Laboratory Under Contract F33615-78-C-1551. The views and conclusions contained in this document ate those of the author and should not be interpreted as representing the official polictes, either expressed or implied, of the. Defense Advanced Research Projects Agency or the US Government. Throughout these IPS experiments, the underlying knowledge organization has been Production Systems (PSs) [2], a form of rule-based system in which learning is formulated as the addition to, and modification of, an unstructured collection of production rules. Behavior is obtained through a simple recognize-act cycle with a sophisticated set of principles for resolving conflicts among rules. The dynamic short-term memory of the system is the Working Memory (WM), whose contents are matched each cycle to the conditions of rules in the long-term memory, Production Memory. Study of seven major attempts to construct instructable PSs with various orientations leads to recognizing the centrality of eight functional components. Listing the components and their embodiment in various versions of IPS can contribute to research on learning systems in general, by clarifying some of the important subproblems. This discussion is the first overview of the work of the project to date, and indicates its evolutionary development. Members of the IPS project are no longer working together intensively to build an instructable F ‘S, but individual studies that will add to our knowledge about one or more of these components are continuing. Progress on the problem of efficiency of PSs has been important to the IPS project [3], but will not be discussed further here. 2. Essential Functional Components of Inst ructable Systems The components listed in this section are to be interpreted loosely as dimensions along which learning systems might vary. Interaction The content and form of communications -A between instructor and IPS can have a lot to do with ease and effectiveness of instruction. In particular, it is important to know how closely communications correspond to internal IPS structures. Similarly, we must ask how well the manifest behavior of IPS indicates its progress on a task. An IPS can have various orientations towards interactions, ranging from passive to active, with maintenance of consistency and assimilation into existing structures. Organization. Each version of IPS approaches the issue of obtaining coherent behavior by adopting some form of organization of its ‘procedural’ knowledge. This may involve such techniques as collecting sets of rules into ‘methods’ and using signal conventions for sequencing. Whether IPS can explain its 228 static organization and whether the instructor of procedural control are important subissues. can see the details Explanation. A key operation in an instructable system is that of explaining how the system has arrived at some behavior, be it correct or incorrect. In the case of wrong behavior, IPS must reveal enough of its processing to allow the more intelligent instructor to determine what knowledge IPS lacks. Accommodation. Once corrections to IPS’s knowledge have been formulated by the instructor, it remains for further interactions with IPS to augment or modify itself. In the IPS framework, these modifications are taken to be changes to the rules of the system, rather than changes to the less permanent WM. As with interaction, IPS can take a passive or active approach to this process. Connection. Manifest errors are not the only way a system indicates a need for instruction: inability to connect a current problem with existing knowledge that might help in solving it is perhaps a more fundamental one. An IPS needs ways to assimilate problems into an existing knowledge framework, and ways to recognize the applicability of, and discriminate among, existing methods. Reformulation. Another way that IPS can avoid requiring instruction is for it to reformulate existing knowledge to apply in new circumstances. There are two aspects to this function: finding knowledge that is potentially suitable for mapping, and performing the actual mapping. In contrast to connection, this component involves transformation of knowledge in rules, either permanently or dynamically. Evaluation. Since the instructor has limited access to what IPS is doing, it is important for IPS to be able to evaluate its own progress, recognizing deficiencies and errors ai they occur so that instruction can take place as closely as possible to the dynamic point of error. Defining what progress is and formulating relevant questions to ask to fill gaps in knowledge are two subissues. Compilation. Rules initially formed as a result of the instructor’s input may be amenable to refinements that improve IPS’s efficiency. This follows from several factors: during instruction, IPS may be engaged in search or other ‘interpretive’ execution&; instruction may provide IPS with fragments that can only be assembled into efficient form later; and IPS may form rules that are either too general or too specific. Improvement with practice is the psychological analog of this capability. Anderson et al [l] have formulated several approaches to compilation. 3. Survey of Approaches Kernell, ANA, Kernel2 and IPMSL have been fully implemented. The others were suspended at various earlier stages of development, for reasons that were rarely related to substantive or conceptual difficulties. Kernel Version 1 -ZL The starting point for IPS is the adoption of a pure means-ends strategy: given explicit goals, rules are the means to reducing or solving them. Four classes of rules are distinguished: means rules; recognizers of success; recognizers of failure; and evocation of goals from goal-free data. The Kernel1 [6] approach further organizes rules into methods, which group together (via patterns for the same goal) a number of means, tests and failure rules. Interaction consists of language strings that roughly correspond to these methods and to system goals (among which are queries). Keywords in the language give rise to the method sequencing tags and also serve to classify and bound rules. Explanation derives from the piecing together of various goals in WM, along with associated data. The major burden of putting together raw data that may be sufficient for explanation rests on the instructor, a serious weakness. Additive Successive Aouroximations [ASA). Some of the drawbacks of I<ernell can be remedied* by orienting instruction towards fragments of methods that can be more readily refined at later times. Interaction consists of having the instructor point at items in IPS’s environment (especially WM) in four ways: condition (for data to be tested), action (for appropriate operators), relevant (for essential data items), and entity (to create a symbol for a new knowledge expression). These designations result in methods that are very loose collections of rules, each of which contributes some small amount towards achievement of the goal. Accommodation is done as post-modification of an existing method in its dynamic execution context, through ten method-modification methods. Analoclv. A concerted attempt to deal with issues of connection and accommodation is represented by McDermott’s ANA program [4]. ANA starts out with the ability to solve a few very specific problems, and attacks subsequent similar problems by using the methods it has analogically. The starting methods are hand-coded. Connection of a new goal to an existing method takes place via special method description rules that are designed to respond to the full class of goals that appear possible for a method to deal with by analogy. An analogy is set up by finding paths through a semantic network containing known objects and actlons. As a method operating by analogy executes, rules recognize points where an analogy breaks down. Then general analogy methods are able either to patch the method directly with specific mappings or to query the instructor for new means-ends rules. Each attempt to build an IPS has been based on the idea of an initial hand-coded kernel system, with enough structure in it to support all further growth by instruction, A kernel establishes the internal representations and the overall approach to instruction. The following are presented in roughly chronological order. *These ideas were introduced by A. Newell in October, 1977. 229 Problem Soaces. Problem spaces [5]* provide a novel basis for IPS by embedding all behavior and interactions in search. A problem space consists of a collection of knowledge elements that compose states in a space, plus a collection of operators that produce new states from known ones. A problem consists of an initial state, a desired state, and possibly path constraints. Newell’s Problem Space Hypothesis (ibid.) claims that all goal-oriented cognitive activity occurs in a problem space, not just activity that is sufficiently problematical. Interaction consists of giving IPS problems and search control knowledge (hints as to how to search specific spaces). Every Kernel component must be a problem space too, and thus subject to the same modification processes. The concrete proposal as it now stands concentrates on interaction, explanation (which involves sources of knowledge about the present state of the search), and organization. Schemas. The use of schemas as a basis for an IPS kerner* make slot-filling the primary information-gathering operation. A slot is implemented as a set of rules. The slots are: executable method; test of completion; assimilation (connects present WM with the schema for a goal); initialization (gathers operands for a method); model (records the instruction episode for later reference); accommodation (records patches to the method); status (records gaps in the knowledge); monitoring (allows careful execution); and organization (records method structure). Orientation towards instruction is active, as in ASA. Explanation consists of interpreting the model slot, and accommodation, of fitting additions into the model. Connection is via a discrimination network composed of the aggregated assimilation slots of all schemas. Compilation is needed here, to map model to method. Kernel Version 2. An approach with basic ideas similar to ASA and to Waterman’s Exemplary Programming [8], Kernel2 [7] focusses on the process of IPS interacting with the instructor to build rules in a dynamic execution context. The instructor essentially steps through the process of achieving a goal, with IPS noting what is done and marking elements for inclusion in the rules to be built when the goal is achieved. Kernel2 includes a semantic network of information about its methods, for use as a ‘help’ facility. Kernel2 is the basis from which the IPMSL system, below, is built. Semantic Network. Viewing accumulation of knowledge as additions to a semantic network is the approach taken by the IPMSL system [7]. Interaction consists of definition and modification of nodes in a net, where such nodes are represented completely as rules. Display and net search facilities are provided as aids to explanation and accommodation. The availability of traditional semantic network inferences makes it possible for IPMSL to develop an approach to connection and reformulation, since they provide a set of tools for relating and mapping knowledge into more tractable expressions. *This 1978. approach was formulated by A. Newell and J. Laird in October of 4. Conclusions The IPS project has not yet succeeded in combining effective versions of components as discussed above, to produce an effective IPS. The components as presently understood and developed, in fact, probably fall short of complete adequacy for such a system. But we have explored and developed a number of approaches to instructability, an exploration that has added to the stock of techniques for exploiting the advantages of PSs. We are encouraged by the ability of the basic PS architecture to enable explorations in a variety of directions and to assume a variety of representations and organizations. Acknowledqments. Much of the work sketched has been done jointly over the course of several years. Other project members are (in approximate order of duration of commitment to it): Allen Newell, John McDermott, Charles L. Forgy, Kamesh Ramakrishna, Pat Langley, Paul Rosenbloom, and John Laird. Helpful comments on this paper were made by Allen Newell, Jaime Carbonell, David Neves and Robert Akscyn. References 1. Anderson, J. R., Kline, P. J., and Beasley, C. M. Jr. A Theory of the Acquisition of Cognitive Skills. Tech. Rept. 77-1, Yale University, Dept. of Psychology, January, 1978. 2. Forgy, C. L. OPS4 User’s Manual. Tech. Rept. CMU-CS-79-132, Carnegie-Mellon University, Dept. of Computer Science, July, 1979. 3. Forgy, C. L. On the Efficient Implementation of Production Systems. Ph.D. Th., Carnegie-Mellon University, Dept. of Computer Science, February 1979. 4. McDermott, J. ANA: An Assimilating and Accommodating Production System. Tech. Rept. CMU-CS-78-156, Carnegie-Mellon University, Dept. of Computer Science, December, 1978. Also appeared in IJCAI-79 5. Newell, A. Reasoning, problem solving and decision processes: the problem space as a fundamental category. In Attention and Performance VIII, Nickerson, R., Ed.,Erlbaum, Hillsdale, NJ, 1980. 6. Rychener, M. D. and Newell, A. An instructable production system: basic design issues. In Pattern-Directed Inference Systems, Waterman, D. A. and Hayes-Roth, F., Eds., Academic, New York, NY, 1978, pp. 135-153. 7. Rychener, M. D. A Semantic Network of Production Rules in a System for Describing Computer Structures. Tech. Rept. CMU-CS-79-130, Carnegie-Mellon University, Dept. of Computer Science, June, 1979. Also appeared in IJCAI-79 8. Waterman, D. A. Rule-Directed Interactive Transaction Agents: An Approach to Knowledge Acquisition. Tech. Rept. R-21 71 -ARPA, The Rand Corp., February, 1978. d*Schemas were first proposed for IPS by Hychener, May, 1978 230

1980

19

11

SOME ALGORITHM DESIGN METHODS Steve Tappel Systems Control, Inc., 1801 Page Mill Road Palo Alto, California 94304 and Computer Science Department, Stanford University Abstract Algorithm design may be defined as the task of finding an efficient data and control structure that implements a given input- output specification. This paper describes a methodology for control structure design, applicable to combinatorial algorithms involving search or minimization. The methodology includes an abstract process representation based on generators, constraints, mappings and orderings, and a set of plans and transformations by which to obtain an efficient algorithm. As an example, the derivation of a shortest-path algorithm is shown, The methods have been developed with automatic programming systems in mind, but should also be useful to human programmers. 1. introduction The general goal of automatic programming research is to find methods for constructing efficient implementations of high- level program specifications. (Conventional compilers embody such methods to a very limited extent.) This paper describes some methods for the design of efficient control structures, within a stepwise refinement paradigm.. in stepwise refinement (see for instance Cl,Zl), we view the program specification itself as an algorithm, albeit a very inefficient one. Through a chain of transformation steps, we seek to obtain an efficient algorithm. Specification -+ Alg -+ . . . 3 Algorithm Each transformation step preserves input-output equivalence, so the final algorithm requires no additlonal verification. Algorithm design is a difficult artificial intelligence task involving representation and planning issues. First, in reasoning about a complicated object like an algorithm it is essential to divide it into parts that interact in relatively simple ways. We have chosen asynchronous processes, communicating via data channels, as an appropriate representation for algorithms. Details are in Section II. Second, to avoid blind search each design step must be clearly motivated, which in practice requires organization of the transformation sequence according to high-level plans. An outline of the plans and transformations we have developed is given .in Section III, followed in Section IV by the sample derivation of a shortest path algorithm. Sections V and VI discuss extensions and conclude. This methodology is intended for eventual implementation within the CHI program synthesis system 131, which is under development at Systems Control Inc. This research is supported in part by the Defense Advanced Research Projects Agency under DARPA Order 3687, Contract N00014-79-C-0127, which is monitored by the Office of Naval Research. The views and conclusions contained in this paper are fhe of the author and should not be interpreted as necessarily representing the oflicial policies, either expressed or implied, of Xl, DARPA, ONR or the US Government. II. Process graph representation of algorithms Our choice of representation is motivated largely by our concentration on the earlier phases of algorithm design, in which global restructurings of the algorithm take place. Most data structure decisions can be safely left for a later phase, so we consider only simple, abstract data types like sets, sequences and relations. More importantly, we observe that conventional high- level languages impose a linear order on computations which is irrelevant to the structure of many algorithms and in other cases forces a premature committment to a particular algorithm. To avoid this problem, we have chosen a communicating process representation in which each process is a node in a directed graph and processes communicate by sending data items along the edges which act as FIFO queues. Cycles are common and correspond to loops in a conventional language. The use of generators (or producers) in algorithm design was suggested by 141. Our representation is essentially a specialized version of the language for process networks described in I51 Rather than strive for a general programming language we use only a small set of process types, chosen so that: (I) the specifications and algorithms we wish to deal with are compactly represented, and (2) plans and transformations can be expressed in terms of adding, deleting or moving process nodes. The four process types are: Generator: produces elements one by one on its output edge. Constraint: acts as a filter; elements that satisfy the constraint pass through. Mapping: takes each input element and produces some function of it. If the function value is a set its elements are produced one by one. Ordering: permutes its input elements and produces them in the specified order. The representation is recursive, a very important property. There can be generators of constraints, constraints on constraints, mappings producing generators, etc. Most of the same design methods will apply to these “meta-processes”. To illustrate the representation, we encode the specification for our sample problem of finding the shortest path from a to b in a graph. Idotation and terminology for shortest path. A directed graph is defined by a finite vertex set V and a binary relation Edge(u,v). A path p is a sequence of vertices (pi . . . p,), in which Edge(pi.pbr) holds for each pair. The “.” operator is used to construct sequences: (u . . . v).w = (u . . . v w). Every edge of the graph is labelled with a positive weight W(u,v) and the weight of an entire path is then Weight(p) = W(pi,pz)+...+W(p,l,p,,). The shortest path from a to b is just the one that minimizes Weight. 64 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. A specification should be as simple as possible to ensure correctness. Shortest path can be simply specified as: generate ail paths from a to b, and select the shortest. We express selection of the shortest path in a rather odd way, feeding all the paths into an ordering process whose very first output will cause the algorithm to stop. The point is that by using a full ordering for this comparatively minor task, we can apply all the plans and transformations for orderings. As for the paths from a to b, they are defined as a certain kind of sequence of vertices, so we introduce a generator of all vertex sequences and place constraints after it to eliminate non-paths. This completes the specification. Constraint methods. The goal of constraint methods is to reduce the number of elements generated. The top level plan for constraints says to: I. propagate constraints through the process graph to bring them adjacent to a generator, 2. incorporate constraints into a generator whenever possible, and if the results are not satisfactory, 3. deduce new constraints beyond those given in the specification, and repeat. Each of the three subtasks is nontrivial in itself and is carried out according to a set of (roughly speaking) intermediate- level methods. For (2) an intermediate-level method that we use several times in the Shortest Path derivation is: Selection of an appropriate internal structure for the generator of Sequences(V) is actually part of the design process, but to simplify the example we will take as a default the usual recursive definition of sequences. The recursion in the definition corresponds to a cycle in the process graph. The constraint incorporation plan ConstrainComponent. ConstrainComponent applies when a constraint on composite objects x (sets, sequences, not numbers) is reducible to a constraint on a single component c of x, i.e. P(x) e P’(x,). ConstrainComponent then gives a plan: I. Inside the generator, find the sub-generator of values for component c. If necessary, manipulate the process graph to isolate this generator. Again, other methods must be called upon. 2. Remove constraint P and add constraint P’ to the sub- generator. 0 csequences (VI e Ordering methods. Another group of methods is concerned with the deduction, propagation and incorporation of orderings i f scSe 4 uences(VI I on a generated set. These methods are analogous to the methods r I and vc for constraints but more complicated. In the Shortest Path then s. vcsequences (VI Map: is.v 1 WV) derivation we use a powerful transformation, explained rather A sketchily here: The generation process starts when the empty sequence () is produced on the 3” edge. From the “s” edge it goes to the constraint and also to the mapping, which produces the set of all one-vertex sequences ().v, for vcV. These are fed back to generate two-vertex sequences, and so on. A mapping cycle like this is a very common kind of generator. 111. hods for algorithm design The program specification from which design starts is typically written as an exhaustive generate-and-test (or generate- and-minimize) process, and bears little resemblance to the algorithm it will become. The design methods all have the goal of incorporating constraints, orderings or mappings into the generator, or else the goal of planning or preparing to do so. To incorporate a constraint means to modify the generator so that it only generates items which already satisfy the constraint; to incorporate an ordering means to modify the generator so it generates elements directly in that order; and to incorporate a mapping f means to generate elements f(x) instead of elements x. Accordingly, the methods fall into three main classes, briefly described below. Sunerimnosed upon this class division is a heirarchy (not strict) &th multi-step’plans at the higher levels and a large number of specific syntactic transformations at the bottom. The heirarchy is organized according to goals and subgoals. Heuristics and deduction rules are required to support the planning activity. At the time of writing, a total of about 20 methods have been formulated not counting low-level syntactic transformations. The ordering incorporation transformation InterleaveOrder. InterleaveOrder applies when an ordering R is adjacent to a generator consisting of a mapping cycle, in which the mapping f has the property R(x f(x)) for all x. In other words, f(x) is greater than x under the ordering R. InterleaveOrder moves the ordering inside the mapping cycle and adds a synchronization signal to make the ordering and mapping operate as coroutines. The ordering produces an element x, the mapping receives it and produces its successors f(x) (there would be no need for the ordering at all if f were single-valued), then the ordering produces the next element and so on. Mapping methods. The methods for incorporating a mapping into a generator are mostly based upon recent work in the “formal differentiation of algorithms” [6] and are related to the well-known optimizing technique of reduction in operator strength. (They are not used in our sample design.) Some syntactic transformations and other methods described in this section will appear in the derivation. Example: Design of a shortest path ~~go~~t~rn not In the design which follows, the specification will be transformed from an inefficient generate-and-minimize scheme into a dynamic programming algorithm. The final algorithm grows paths out from vertex a, extending only the shortest path to each intermediate vertex, until reaching b. Of necessity we omit many details of the design. IV. 1. Constraint methods. Since the specification’s constraints are already next to the generator (step i), the overall plan for constraints says to try to incorporate them (step 2.) We will follow the heuristic of incorporating the strongest constraint first. Right now, the algorithm reads Incorporate the Edge constraint. More detail will be shown in this first step than in later derivation steps. ConstrainComponent applies because once a vertex Si has been added to a skquence, -the constraint Edge(Si,Sk,) reduces to a constraint on the single component Si+lm (This reasoning step is really the application of another method, not described here.) Step (I) in the Co ns raincomponent t plan says to find the generator of values for components Si+l* Though we have written it in linear form for convenience, the expression (s.v 1 vcV} is really a generator followed by a mapping. Unfortunately “vcV” generates SI as well as the desired siri values, so we have to unroll one cycle of the graph to isolate the generator of Si.1 values. (Agaln, we have applied methods not described in this paper.) Step (2) is now possible and consists in constraining v to satisfy Edge(s,,v). With the Edge constraint inzorporated, only paths are now being generated so we change s to p in the diagram. I Hap: (p.v 1 vcV A Edge(pn,vll I Incorporate the constraiut that pl=a. Since the pl=a constraint refers only to a component of p, ConstrainComponent applies again. We constrain v in the first “vcV” generator to be equal to a. After simplifying, we obtain Incorporate the constraint that pn=b. Once again ConstrainComponent applies. This time, however, we are unable to isolate a generator for the last vertex of paths. The last vertex of one path is the next-to-last vertex of another, and so on. ConstrainComponent fails, other methods fail too; we leave the pn=b constraint unincorporated. Deduce new constraint. In accordance with the general constraint plan (step 3) we now try to deduce more constraints. One method for deducing new constraints asks: do certain of the generated elements have no eflecf whatroever upon the result of the algorithm? If the answer is “yes”, try to find a predicate that is false on the useless elements, true on others. Motive: if we later succeed in incorporating this constraint into the generator, the useless elements will never be produced. NOW consider the Order + STOP combination. Because all it does is select the shortest path, any path which is not shortest will have no effect! The corresponding constraint says: p is a shortest path from a to b. A further deduction gives the even stronger constraint that every subpalh of p must be a shortest path (between its endpoints). Incorporation of this constraint is complex and is deferred till after incorporation of the Weight ordering. IV.2. Ordering methods. So far paths are generated according to the partial order of path inclusion; path p is generated before path q if q = p.u . . .V for some vertices u,..,v. We may generate a lot of paths to b before generating the shortest one - possibly an infinite number. However if the Weight ordering can be incorporated into the path generator, then only a single path to b (the shortest one) will ever be generated. Propagate Ordering. Before applying an incorporation method we need to bring the Weight ordering next to the generator. Constraints and orderings commute so this is easy. Incorporate the ordering into the generator. The InterleaveOrder method applies, because Weight(p.v) is greater than Weight(p). It moves the ordering from outside the generator cycle to inside and also causes the ordering to wait for the mapping to finish extending the previous path before it produces another. Incorporate new constraint. The “p is a shortest path” constraint is readily incorporated now: the shortest path to any vertex will be the jirsi path to that vertex. Any later path q, with the same last vertex q,=p,, can be eliminated by a new constraint C(p) = xq. q,=p,. We introduce a mapping to produce these new constraints C(p), and now we have a generator of consfrainfs. The result of the last three steps is The algorithm is now a breadth-first search for a path to b, with elimination of non-optimal paths at every vertex. Despite various inefficiencies that remain, the essential structure of a dynamic programming algorithm is present. One interesting improvement comes from incorporating the generated constraints C(p) Into the generator of paths, using ConstrainComponent. To complete the derivation would require data structure selection and finally a translation into some conventional programming language. 66 V. Other results and limitations Besides the Shortest Path algorithm shown here (and variants of it) the algorithm design methods have been used [71 to derive a simple maximum finding algorithm and several variants on prime finding including the Sieve of Eratosthenes and a more sophisticated linear-time algorithm. In these additional derivations, no new process types and only a few new methods had to be used. Single and multiple processor implementations have informally been obtained from process graph algorithms, for both prime finding and Shortest Path. More algorithms need to be tried before specific claims about generality can be made. The intended domain of application is combinatorial algorithms, especially those naturally specified as an exhaustive search (possibly over an infinite space) for objects meeting some stated criteria, which can include being minimal with respect to a defined ordering. Backtrack algorithms, sieves, many graph algorithms and others are of this kind 181. The methods described here are quite narrow in the sense that a practical automatic programming system would have to combine them with knowledge of: 1. Standard generators for different kinds of objects. Our methods can only modify an existing generator, not design one. 2. Data structure selection and basic operations such as searching a list. 3. Efficiency analysis to determine if an incorporation really gives a speedup. 4. Domain specific facts, e.g., about divisibility if designing a prime finding algorithm. 5. How to carry out the final mapping of process graph into a conventional programming (or multiprogramming) language. VI. Discussion and Conclusions The main lesson to be learned from this work is the importance of using an abstract and modular representation of programs during algorithm design. Details of data structure, low- level operations and computation sequencing should be avoided, if possible, until the basic algorithm has been obtained. (Since some algorithms depend crucial)y upon a well-chosen data structure, this will not always be possible.) Further, it is advantageous to represent algorithms in terms of a small number of standard kinds of process, for which a relatively large number of design methods will exist. The results so far indicate that just four standard processes suffice to encode a moderate range of different specifications and algorithms. Presumably others will be required as the range is extended, and it is an important question whether (or how long) the number can be kept small. A similar question can be asked about the design methods. One would not expect methods based upon such general constructs as generators, constraints, orderings and mappings to have much power for the derivation of worthwhile algorithms. For instance, if we had explicitly invoked the idea of dynamic programming, our derivation of a shortest path algorithm would have been shorter. For really difficult algorithms, the general methods may be of little use by themselves. We suggest that they should still serve as a useful complement to more specific methods, by finding speedups (based on incorporation of whatever constraints, orderings and mappings may be present) in an algorithm obtained by the specific methods. As a final issue, it is interesting to speculate how the stepwise refinement approach to programming might be used by human programmers. Use of a standard set of process types and correctness-preserving transformations would be analogous to the formal manipulations one performs in solving integrals or other equations. If that were too restrictive, perhaps one could use the methods as a guide, without attempting to maintain strict correctness. After obtaining a good algorithm, one could review and complete the design, checking correctness of each transformation step. The result would be a formally correct but also well-motivated derivation. Acknowledgements. Many helpful ideas and criticisms were provided by Cordell Green, Elaine Kant, Jorge Phillips, Bernard Mont-Reynaud, Steve Westfold, Tom Pressburger and Sue Angebranndt. Thanks also to Bob Floyd for sharing his insights on algorithm design. References I. Baiter, Robert; Goldman, Neil: and Wile. David. “On the transformational implementation approach to programming”, Proc. 2nd Int’l Conference on Software Engineering (1976) 337-344. 2 3 4. 5 6, 7. 8. 9. Barstow. David R. Knowledge Based Program Construction, Elsevier North-Holland, New York, 1979. Phillips, Jorge and Green, Cordell. “Towards Self-Described Programming Environments”, Technical Report, Computer Science Dept., Systems Control, Inc., Palo Alto, California, April 1980. Green, Cordell and Barstow, David R. “On Program Synthesis Knowledge”, Arti.cial Intelligence, 103 ( 1978) 24 I - 279. Kahn, Gilles and MacQueen, David B. “Coroutines and Networks of Parallel Processes”, lnformafion Processing 77, IFIP, North-Holland Publishing Company, Amsterdam, ( 1979) 993-998. Paige, Robert. “Expression Continuity and the Formal Differentiation of Algorithms”, Courant Computer Science Report x5, (1979) 269-658. Tappel, Steve. “Algorithm Design: a representation and methodology for control structure synthesis”, Technical Report, Computer Science Dept., Systems Control, Inc., Palo Alto, California, August 1980. Reingold. Edward M., Nievergelt. Jurg, and‘ Deo, Narsingh. Combinatorial Algorithms: Theory and Practice, Prentice-Hall Inc., Englewood Chffs, New Jersey, 1977. Elschlager, Robert and Phillips, Jorge. “Automatic Programming” (a section of the Handbook of Artificial Intelligence, edited by Avron Barr and Edward A. Feigenbaum), Stanford Computer Science Dept., Technical - --_ _--_ Report 758, 1979. 10. Floyd, R. W. “The Paradigms of Programming” (Turing Award Lecture), CACM 22:s (1979) 455-460. 67

1980

2

12

USING A MATCHER TO MAKE AN EXPERT CONSULTATION SYSTEM BEHAVE INTELLIGENTLY* Rene' Reboh Artificial Intelligence Center SRI International, Menlo Park, CA 94025 ABSTRACT This paper describes how even a simple matcher, if it can detect simple relationships between statements in the knowledge base, will support many features that will make a knowledge- based consultation system appear to behave more intelligently. We describe three features that are useful during the knowledge acquisition phase (involving the building and testing of the knowledge base), and four features that are of assistance during a consultation. Although these features are described in terms of the Prospector environment [2], it will be clear to the reader how these features can be transferred to other environments. I INTRODUCTION Partial-matching (also referred to as interference-matching, correspondence-mapping, . ..) touches many issues representation and efficiency in many AI syZ:ems [ 31. Its role in Prospector is significant because it is involved in many aspects of consultation and knowledge acquisition. Given any two statements Sl and S2 of the knowledge base, the Semantic Network Matcher of Prospector determines which of the following situations applies: Sl and S2 are identical (Sl = S2) Sl is a restriction of S2 (SlC S2), (or S2 is a generalization of Sl (s2 3 Sl)) Sl and S2 are disjoint statements (Sins2 = $ > Sl overlaps S2 (otherwise) For instance, suppose the knowledge base contains the following statements: Sl : "rhyolite is present" s2: "a rhyolite plug is present" s3: I(an igneous intrusive is present" s4: "rhyolite or dacite is present" s5: "pyrite is present" As these statements are being added to the knowledge base, the Matcher will conclude that: S2 is a restriction of Sl, S2 is a restriction of S3 (rhyolite is an igneous rock and a plug is a special kind of intrusive), Sl and S3 overlap (rhyolite is an igneous rock, but need not be an intrusive), Sl is a restriction of S4, S2 is a restriction of S4 (transitivity from the first and third conclusion), S3 and S4 overlap, S5 is disjoint from Sl, S2, S3 and S4. A detailed description of how the Matcher operates in the Prospector environment can be found in [6]. We mention briefly here that the Matcher views each statement as a set of constraints corresponding to a set of assertions about the existence of physical entities or processes and their attributes. semantic networks [:i l5-;spez;;;, partitioned to represent statements in the knowledge base whereby these assertions are expressed in terms of relations and entries in taxonomies of the domain of application (in this case geology). Let us examine some of the features of a knowledge-based system that can be supported by such a Matcher. --a----- * This work was supported by the Office of Resource Analysis of the U. S. Geological Survey under Contract No. 14-08-0001-15985. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the author, and do not necessarily reflect the views of the U. S. G. S. 231 II USE OF THE MATCHER IN KNOWLEDGE ACQUISITION ----- A. Aid in Maintaining Probabilistic Consistency -- of the Inference Network -- The knowledge bases of many expert systems are organized as explicit or implicit networks of statements connected by rules or logical constructs. Because such networks provide the framework for judgmental reasoning, various numerical values, such as probabilities or certainties, are often maintained in them. A major concern of expert systems is the difficulty of keeping the knowledge base error free and consistent in form and content as it grows. Let us examine how the Matcher assists Prospector in maintaining probabilistic consistency in the case where Sl is the most recently entered statement, and S2, which already exists in the knowledge base, is a restriction of Sl. (a) Because S2 is a restriction of Sl, the probability of S2 can never exceed that of Sl. In particular, if the prior probabilities supplied by the domain specialist (DS) do not satisfy this constraint, the Matcher will detect the violation and a correction will be required. Thus, before a consultation begins, we can assume that P(S2) < P(S1). (b) Unfortunately, even though all the probabilistic constraints are initially satisfied, the probability changes .that follow from the use of inference rules may not maintain them. For example, if Sl and S2 are the hypotheses (right- hand sides) of two rules El --> Sl and E2 --> S2, and if the evidence (left-hand side) El is sufficiently unfavorable for Sl, we may have P(Sl[El) < P(S2). Similarly, if the evidence E2 is sufficiently favorable for S2, we may have P(S1) < P(S21E2). In essence, the problem is that when the DS provided the rule saying that the evidence El is unfavorable for Sl (rhyolite), he overlooked the fact that El is also unfavorable for S2 (a rhyolite plug), and did not supply a rule of the form El --> s2. Similarly, when he supplied the rule saying that the evidence E2 is favorable for a rhyolite plug, he overlooked the fact that E2 is also favorable for rhyolite, and did not supply a rule of the form E2 --> Sl. Indeed, the DS should not be asked to supply rules for such obvious deductions. The Matcher helps to detect these situations. It is the responsibility of the consultation system to take the appropriate actions to maintain the probabilistic consistency in the inference networks. In [l] it is shown how Prospector uses results from the Matcher to create additional rules ensuring that the probabilistic constraints will be maintained at run time when inference rules are applied. B. Aid in Designing the Semantic Representation -- Because statements in the knowledge base may be arbitrarily complex, their semantic encoding is often entered manually during the knowledge acquisition phase. During a consultation, however, the user is allowed to volunteer information to the system, and semantic re;r;;~;;:t,~~] is used to create the corresponding to the volunteered statements. The kinds of statements that can be translated by the parser depend upon taxonomy contents and an external grammar. Whether the semantic representation of statements is entered manually or constructed by a parser, the knowledge engineer needs to determine if the resulting representation is adequate. He must ensure that it reflects the intentions of the DS in all situations that could occur during of a consultation. Statements can be combined to form larger statements or broken into smaller units, and their semantic representation need not always be elaborate. Which representation is finally chosen depends upon what other statements are in the knowledge base and how they are related, as well as what the DS thinks is the most appropriate for each particular situation. Because the Matcher can be used to analyze how statements are related, it can assist in choosing an appropriate representation. In particular, no elaborate semantic representation may be needed for a statement (or a portion of a statement) that is unrelated to any other statement in the knowledge base. Because such a statement is unlikely to have a major effect on the consultation, a simple text-string representation would be adequate for most purposes. In addition to determining if a restriction/generalization relation exists between two statements, the Semantic Network Matcher in Prospector can identify corresponding elements of the statements and point out the nature of their differences. This feature has been exploited to some extent in the knowledge acquisition module of Prospector [2] where it was used to choose a representation for a statement from possible alternative representations. For instance, a conjunction "X and Y" can be encoded either as a single statement or as two statements, "X" and "Y," connected by a logical AND in the inference network. The first alternative is chosen if a statement already exists in the knowledge base that is equal to or is a restriction of "X and Y," or if "X and Y" is not related to any existing statement. The second alternative is chosen otherwise. We believe this approch can be generalized, and that an automatic procedure using the Matcher can be devised to assist in the uniform, and perhaps optimal, encoding of all statements in the knowledge base. C. As a Search Feature--Accessing by Contents --- - Development and testing of a knowledge base typically extend over long periods. The knowledge engineer cannot be expected to remember all the statements (or any labels assigned to them) that he 232 or another knowledge engineer developing the same knowledge base has already entered. The Matcher can be used as a search feature allowing the knowledge engineer to access statements by specifying a partial description of their (semantic) contents. In effect, the Matcher-search (a command of the knowledge acquisition system) will allow the knowledge engineer to say something like: "Now I want to work on that portion of the inference network that deals with sulfide mineralization." The search-by-content feature is accomplished by matching the partial description specified by the knowledge engineer with statements currently in the knowledge base. III USE OF THE MATCHER IN CONSULTATION ----- A. As a Tool for Maintaining Consistency of the -- User'-s Answers: -- 1 . Discovering Inconsistencies If a user is allowed to volunteer information to an expert system, logical inconsistencies in the input could result. For example, suppose the user volunteers the two statements Sl and S2 concerning rhyolite and a rhyolite plug. Because S2 is a restriction of Sl, if his statements imply that P(S2) > P(Sl), he will be reminded that his answers are contradictory. This is the case, for instance, if the user says that "there are igneous rocks" with some degree of certainty, but later volunteers that "there is a rhyolite plug" with a higher degree of certainty. The contradictions occurring in a consultation often involve several levels of inference and long chains of restriction/generalization links, which sometimes have embarrassed our expert users while being impressed by Prospector's ability to detect the inconsistencies. 2. Changing Answers A significant advantage of the Bayesian method used in Prospector for computing probabilities is the ease with which answers to questions can be changed without having to repeat all previous calculations. Basically, all that is required in changing an answer to a question about any evidence E is to change the probability for E and to propagate the results through the inference network. The possibility of violating the restriction/generalization probabilistic constraints causes the only complication in this process. However, by keeping a record of how statements are related as computed by the Matcher , the answer-changing program knows that it may also have to change the probabilities of some of the related statements in order to maintain consistency. For instance, if the inference network contains the two rules Sl --> Hl and S2 --> H2, and the user gives a negative answer to a question about Sl, the probability of Hl will be updated (in accordance with the rule strengths associated with the first rule). In addition, because S2 is a restriction of Sl, the probability of H2 must also be updated (in accordance with the rule strengths of the second rule) as if a negative answer had been given for S2 as well. When the user then changes his answer for Sl, the probabilities of both Hl and H2 will be automatically updated and propagated through the inference network. By changing an answer, the user may contradict some of his earlier assertions, and changing these assertions may give rise to still further contradictions. This can confuse the user, but poses no problem for the answer-changing program, which is recursive and will make sure no contradictions are introduced before resuming a consultation. B. Use of the Matcher as a Dialog Management Tool ---____-- 1. Mixed Initiative, Volunteering Information Prospector can work in either of two modes -- the consequent mode or the antecedent mode. In the consequent mode Prospector attempts to establish (or rule out) an hypothesis, and it traces backward through the inference network and searches for appropriate evidences to ask about. In the antecedent mode, the inference network is used in a forward direction to propagate the consequences of input information. Prospector is a mixed initiative system whereby the user has the option of taking control any time during the consultation session to inform Prospector about facts he believes to be relevant. The Matcher makes this possible by relating the volunteered information to the current knowledge base in the same fashion as it did for the knowledge acquisition phase. 2. Control Strategy and Goal Selection -- The information volunteered by the user is often relevant to several hypotheses; In Prospector, a simple scoring criterion is used to select the goal hypothesis. Among other things, this criterion takes into account the volunteered statements (whose effect on the hypotheses may be encouraging or discouraging) that are linked to each hypothesis as recorded by the Matcher. 7. Interaction Psychology Before the user is asked about the evidence E selected by the control strategy, Prospector 233 reminds him about any facts it thinks relevant. The information needed to recognize these facts are the links relating E to other statements in the knowledge base computed by the Matcher and recorded at some earlier phase of the consultation or during knowledge acquisition. How these facts are presented to the user depends upon the current llstatew of the statements involved. The state of a space is determined by its certainty value and how that certainty was established --whether it was inferred by using rules, volunteered by the user, Or inferred through restriction/generalization links through the Matcher. Depending upon the actual situation, one of several standard phrases is displayed before the question is asked, and an appropriate phrase is selected to ask the question. used The fol lowing are some of the in a Pr ospector consultation: standard phrases - You told me about . . . - You suspected . . . - I know you doubted . . . - Your statements imply . . . - I know there is reason to doubt . . . - I have reason to suspect . . . - I have reason to doubt . . . Thus, the program might preface a question about a piece of evidence E by saying: "I have reason to doubt E. What is your degree of belief about that?" Clearly, these stock phrases are simple attempts to inform the user about the implications of his earlier statements. Although they have no effect on the function of Prospector and are not necessary in any logical sense, they enhance communication between the user and the consultation system and often serve to make the logical processes of the consultation system more evident. The Matcher has been an important tool for the design of the interaction environment in all phases of development and use of the Prospector knowledge- based system. It is particularly important in the "psychology" of man-machine interaction in consultation systems that the user does not feel ignored and that the dialogs are not totally dictated by the system. Whenever possible, the user should be shown evidence that the system listens to him, understands what he says, and sometimes can even use the information he supplied! IV CONCLUSION w M [31 [41 [53 b1 L-71 REFERENCES Duda, R.O., P.E. Hart, N.J. Nilsson, R. Reboh, J. Slocum and G.L. Sutherland, "Development of a Computer-Based Consultant for Mineral Exploration," Annual Report, SRI Projects 5821 and 6415, SRI International, Menlo Park, California (October 1977). Duda, R.O., P.E. Hart, K. Konolige, R. Reboh, "A Computer-Based Consultant for Mineral Exploration," Final Report SRI Project 6415 SRI International, Menlo Park, California (September 1979). Hayes-Roth, F., "The Role Of Partial and Best Matches in Knowledge Systems," in Pattern- Directed Inference Systems, D.A-Waterman and F. Hayes-Roth, eds., pp. 557-574 (Academic Press, New York, 1978). Hendrix, G.G., "LIFER: A Natural Language Interface Facility," SIGART Newsletter, No. 61, pp 25-26 (Februarym. Hendrix, G.G., "Encoding Know1 edge in Partitioned Ne tworks." in Associative - The Representation and Use Networks of Knowledge in -- - Computers, N. V. Findler, ed., Academic Press, New York (1979). Reboh, R., "A Knowledge Acquisition Environment for Expert Consultation Systems," Ph.D. Dissertation, Department of Mathematics, Linkoping University, Sweden (to appear 1980). Waterman, D.A. and F. Hayes-Roth, eds., Pattern-Directed Inference Systems (Academic -New York, 1978). - By providing means to relate the statements in the knowledge base to each other, the semantic network Matcher in Prospector has been an important instrument in supporting many of the features that constitute the AI contents of the system. We believe that the approach is a general one, and can enhance the intelligent behaviour of any knowledge- based system. 234

1980

20

13

AN APPROACH TO ACQUIRING AND APPLYING KNOWLEDGE Norman Haas and Gary G. Hendrix SRI International 333 Ravenswood Avenue Menlo Park, California 94025 ABSTRACT The problem addressed in this paper is how to enable a computer system to acquire facts about new domains from tutors who are experts in their respective fields, but who have little or no training in computer science. The information to be acquired is that needed to support question- answering activities. The basic acquisition approach is "learning by being told." We have been especially interested in exploring the notion of simultaneously learning not only new concepts, but also the linguistic constructions used to express those concepts. As a research vehicle we have developed a system that is preprogrammed with deductive algorithms and a fixed set of syntactic/semantic rules covering a small subset of English. It has been endowed with sufficient seed concepts and seed vocabulary to support effective tutorial interaction. Furthermore, the system is capable of learning new concepts and vocabulary, and can apply its acquired knowledge in a prescribed range of problem-solving situations. I INTRODUCTION Virtually any nontrivial artificial intelligence (AI) system requires a large body of machine-usabge knowledge about its domain of application. Construction of a knowledge base is currently a tedious and time-consuming operation that must be performed by people familiar with knowledge representation techniques. The problem addressed in this paper is how to enable computer systems to acquire sets of facts about totally new domains from tutors who are experts in their own fields, but have little or no training in computer science. In an attempt to find a practical solution to this problem, we have developed a pilot system for knowledge acquisition, which, along with -------- * This research was supported by the Defense Advance Research Projects Agency under contract N00039-79-C-0118 with the Naval Electronic Systems Command. The views and conclusions contained in this document are those of the-authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the Defense Advanced Research Projects Agency of the United States Government. several related research issues, is discussed below. The kinds of information we are most interested in acquiring are those needed to support what have been called "question-answering" or "fact-retrieval" systems. In particular, our interest is in collecting and organizing relatively large aggregations of individual facts about new domains, rather than in acquiring rules for judgmental reasoning. This is in contrast to previous work on such systems as those of Davis [I] and Dietterich and Michalski [2], that treat knowledge not so much as a collection of facts, but as a set of instructions for controlling the behavior of an engine. The type of acquisition process we are exploring is "learning by being told," in contrast to the idea of "learning by example." It is this latter concept which has formed the basis of research by other investigators in this area, such as Winston [ll] and Mitchell [8]. Our interest in knowledge acquisition is motivated by the desire to create computer-based systems that can aid their users in managing information. The core idea is that of a system that can talk to a user about his problems and subsequently apply other types of software to meet his needs. Such software would include data base management systems, report generators, planners, simulators, statistical packages, and the like. Interactive dialogs in natural language appear the most convenient means for obtaining most of the application-specific knowledge needed by such intelligent systems. II KNOWLEDGE ACQUISITION THROUGH ENGLISH DIALOGS Systems that acquire knowledge about new domains through natural-language dialogs must have two kinds of special capabilities. First, they must be capable of simultaneously learning both new concepts and the linguistic constructions used to express those concepts. (This need for simultaneous acquisition of concepts and language reflects the integral connection between language and reasoning.) Second, such systems must support interactive, mixed-initiative dialogs. Because a tutor may provide new knowledge in an incremental 235 and incomplete manner, the system must keep track of what it has already been told so that it can deduce the existence of missing information and explicitly ask the tutor to supply it. We are exploring the feasibility of such ideas by developing a series of Knowledge-Learning and -Using Systems (KLAUS). A KLAUS is an interactive computer system that possesses a basic knowledge of the English language, is capable of learning the concepts and vocabulary of new subject domains, and has sufficient expertise to apply its acquired knowledge effectively in problem-solving situations. III RESEARCH ISSUES FOR KNOWLEDGE ACQUISITION -- To create systems capable of acquiring knowledge through tutorial dialogs in English, several fundamental research problems must be resolved: A powerful natural-language processing capability is required. Although much progress has been made in recent years, previous work has assumed a complete knowledge base. Knowledge-acquisition dialogs require several adaptations and extensions. Seed concepts and seed vocabulary must be identified for inclusion in a core system. It is not at all obvious which words and concepts will be most useful in helping tutors describe the concepts of new domains. A structure for lexical entries must be specified so that the system can acquire new lexical information. Because such information provides a key link between surface linguistic form and underlying meaning, structural specification is a very challenging task for certain categories of words, particularly verbs. The linguistic constructions that people use in introducing new concepts must be identified and analyzed so they can be interpreted correctly by the natural- language processing system. Such constructions range from simple syntactic patterns to complex uses of analogy. A flexible scheme of knowledge representation is necessary. The representation must have general expressive power, since it may be applied to many different domains and must support the addition of new information. It should include inherent features that can aid in organizing knowledge and in supporting the incremental acquisition of knowledge. An efficient problem-solving capability is needed to answer questi6ns and to draw inferences for integrating newly acquired information. This capability must be based on general principles, because no application-specific problem-solving procedures will be included in the system. (How to acquire such procedures is a separate and interesting research question.) methodology is needed for integrating new concepts into the system's knowledge base. Because tutors will often provide only partial descriptions of new concepts, methods have to be devised for ascertaining what additional facts must be sought from the tutor to insure proper linkage between the new concepts and those previously acquired. A set of readily understandable questions is needed for eliciting information from tutors. The length and number of questions should be minimized to impose as small a burden on tutors as possible. These problems must be dealt with in an integrated manner, balancing the requirements of one facet of the system against those of others. Our initial attempts to cope with this complex of issues are presented below. OK, now I have assimilated LENGTH. 'i'4> A foot is a unit of length ---____- FOOTS are the first units of LENGTH you've defined. Now I have assimilated FOOT. 75> The plural of foot is feet ----- FEET - got it. Thanks. 76> A meter is a unit of length ----__-- How is it related to FOOT? >> There are 3.3 --- feet in a meter ---- Now I have assimilated METER. 77> A physical object has a length So PHYSICAL OBJ~a~L~N~ Exactly what to use as seed concepts is problematical for a system that must bootstrap its entry into a new domain. Most of the seed concepts in NANOKLAUS are classes of THINGS and RELATIONS. They have been included in the system either simply to avoid forcing tutors to relate everything to the most general concept (THING), or because they have a special status in English. For example, because pronominal usage depends on gender, the class MALE- BEING is defined and associated with the pronoun "he." One might consider defining as seed concepts a set of primitives, to which all other concepts must be reduced, but such a reductionist approach is probably unworkable [3]. NANOKLAUS uses five principles of knowledge organization to integrate new knowledge: (1) there are things; (2) th ere are subclasses of things (i.e., things can be subclassified); (3) there are relations among things; (4) there are subclasses of relations; (5) some of the relations are functions. The concepts of uniqueness and equality also play important roles. NANOKLAUS is not programmed to hold explicit conversations about these concepts, but rather to use them in its internal operations. C. The Natural-Language Component The natural-language component of NANOKLAUS uses a pragmatic grammar in the style of LADDER b1. Although most of the linguistic processing performed by the system follows fairly standard practice, the pragmatic grammar is distinguished by its explicit identification of a number of syntactic structures used principally to define new concepts. As an oversimplified example, NANOKLAUS might be thought of as looking for the syntactic pattern <s> => <A> <NEW-WORD> <BE> <A> <KNOWN-COUNT-NOUN> to account for such inputs as A CARRIER IS A SHIP. When one of these concept-defining patterns is recognized, an acquisition procedure associated with the pattern is invoked. 'This procedure generally adds new facts to the system's set of wffs and generates new entries in its lexicon. The various acquisition procedures also have provi sions for responding to the tutor. Response generation is accomplished through the use of preprogrammed phrases and templates. D. Using Dialog to Aid Acquisition and ---- Assimilation By and large, it is unreasonable to expect tutors to volunteer all the information NANOKLAUS needs to assimilate new concepts. In particular, tutors cannot be expected to know what conclusions NANOKLAUS will draw about a newly taught concept from its previous knowledge, since they know neither the details of its knowledge nor the even more esoteric details of its assimilation procedures. NANOKLAUS must ask the tutor for the information it needs. But wording the system's questions in an understandable form involves a number of human engineering difficulties. The acquisition of new verbs offers a cogent illustration of the problem. English verbs are highly idiosyncratic; making proper entries for them in a lexicon can be a formidable task. Among other considerations, one must ascertain a verb's passivity and transitivity, determine whether its indirect object can become the object of a FOR or TO prepositional phrase, whether it is reflexive or nonreflexive, and how the syntactic cases of its arguments may be "normalized" when the verb appears in different syntactic constructions. NANOKLAUS's tutors, of course, cannot be expected to describe verbs in linguistic terms, so the system asks a series of questions tutors can understand to elicit the same information. The following is a typical verb-acquisition exchange: 92>An officer can assign a unit ------ Give the -ed and -en forms of ASSIGN (for example, for GO: WENT GONE): >>assigned assigned If an OFFICER can ASSIGNanUNIT something, type the most general description of that something, otherwise <return> >> (user typed a carriage return) If an OFFICER can ASSIGN SomethinganUNIT, type the most general description of that something, otherwise <return> >>a task group --- Can an OFFICER ASSIGN an UNIT TO a TASK GROUP ? >>yes Could an UNIT be ASSIGNED a TASK GROUP by an OFFICER ? >>yes Can an OFFICER ASSIGN an UNIT with a TASK GROUP ? >>no Can an OFFICER ASSIGN a TASK GROUP ? >>no - Can an OFFICER ASSIGN ? >>no Verbs that produce rather differenTanswers to the above questions include GIVg, BUY, RUN (run a machine), SERVE and SUPPLY. E. The Knowledge Representation System NANOKLAUS's knowledge representation system uses a many-sorted first-order logic, that combines features from [g] and [5]. The backbone of the system is a treelike data structure reflecting the NANOKLAUS does not consider modal verbs (e.g., "want" and "know"), or verbs with particles (e.g., "pick up", "preside over") or sentential objects (e.g., "The captain requested that the ship change ---- coursell). 237 hierarchy of sorts (or classes) used by the system. This tree is "tangled" in that any given sort may if any, together with CARRIER, span the class of all SHIPS (with or without have multiple ancestors. It is no accident that a sort hierarchy should serve as the primary data structure for an English-speaking acquisition system. "ISA" hierarchies are used by many natural-language processing systems; it appears that something very similar plays a central role in the their knowledge [7]. to a sort hierarchy way humans organize to determine whether or not a given entity to a particular sort . To a id in this task A fundamental task of the deductive system is belongs NANOKLAUS maintains information about the direct descendants of each sort, such as whether some or all of the descendants of an ancestor exhaust (span) it, and whether are mutually exclusive. two or more sibling sorts The utilization of sort information is clearly illustrated in the following dialog with NANOKLAUS: 80>A ship is a unit YouTrGyGg-that anything that is a SHIP is also an UNIT. What is an UNIT? >>a thing You're saying that anything that is znm is also a THING. Which of the following, if any, are proper subclasses of UNIT: PHYSICAL OBJECT, LEGAL PERSON, or MEASURE? >>none of them --- Which of the following classes, if any, could never have any members in common with UNIT: PHYSICAL OBJECT, LEGAL PERSON, or MEASURE ? >>none OK, now I have assimilated UNIT. OK, now I have assimilated SHIP. 81>A ship is a physical object YouTrEyzg-that anything that is a SHIP is also a PHYSICAL OBJECT. Is ANIMATE OBJECT a proper subclass of SHIP? >>no IS ANIMATE OBJECT necessarily<omposed of entirely different members from SHIP? >>yes Do ANIMATE OBJECT and SHIP span the set of all PHYSICAL OBJECTS >>no OK, now I have assimicted SHIP. 82>A submarine is a ship --- YouTre saying that anything that is a SUBMARINE is also a SHIP. OK, now I have assimilated SUBMARINE. 83>A destroyer is a ship --- YouTre saying that anything that is a DESTROYER is also a SHIP. Is SUBMARINE a proper subclass of DESTROYER? >>no Is SUBMARINE necessarily cozosed of entirely different members from DESTROYER? Do SUBMARINE and DESTROYER span the set- all SHIPS? >>no OK, now I ha= assimilated DESTROYER. 84>A carrier is a ship Yo?rwgThgt anything that is a CARRIER is also a SHIP. Which of the following, if any, are proper subclasses of CARRIERI SUBMARINE or DESTROYER? >>neither of them _I__-- Which of the following classes, if any, could never have any members in common with CARRIER: SUBMARINE or DESTROYER ? >>both of them --- What combination of the following subclasses, overlapping): SUBMARINE or DESTROYER ? >>neither of them --- OK, I have assimilated CARRIER. 85>A Kittyhawk is a carrier YouTre saying that-anything that is a KITTYHAWK is also a CARRIER. OK, now I have assimilated KITTYHAWK. . 98>A carrier has a doctor who is a person --------- So CARRIERS have PEOPLE for DOCTORS. Must a CARRIER have exactly one DOCTOR? >>no - 99>= Constellation is a Kittyhawk -- You're saying that CONSTELLATION is one instance of a KITTYHAWK. OK, now I know about CONSTELLATION. lOO>The JFK is a Kittyhawk --- You'resaying that JFK is one instance of a KITTYHAWK. OK, now I know about JFK. l>The JFK is commanded by Moffet AS-T= THA-(MOFFET C~MMANDFK) 2>The length of the JFK is 1072 feet --------- V FUTURE PROSPECTS At this time NANOKLAUS can be best described as a fragile proof-of-concept system still in its early developmental stage. During this coming year, we plan to greatly expand its linguistic coverage by replacing our current pragmatic grammar with Robinson's [lo] DIAGRAM grammar. Once this has been accomplished and NANOKLAUS's verb acquisition package extended to accept particles and prepositional phrases, we believe NANOKLAUS can serve as a useful tool for aiding AI researchers in the construction of knowledge bases for other AI systems --a task that currently consumes an inordinate proportion of research effort. As suggested in the introduction, one of our long-term objectives is the extension of KLAUS to knowing about diverse types of external software packages. Given knowledge of such packages, a KLAUS could serve as an agent that interacts with them on a user's behalf. To explore these possibilities, we plan in the near future to provide NANOKLAUS with the capability of using a conventional data base management system. In this configuration, a user should be able to tell NANOKLAUS about a new domain, about a data base containing information pertaining to that domain, and about the interrelationship of the two. The new system would then be able to use the data base in answering questions regarding the domain. Our work in the area of knowledge acquisition per se has really just begun. As development proceeds, we plan to turn our attention to making provisions for learning by analogy, for acquiring and reasoning about the internal structures of processes, for dealing with causality, and for dealing with mass terms. 238 ACKNOWLEDGMENTS The deduction system supporting NANOKLAUS was developed in large part by Mabry Tyson, with Robert Moore, Nils Nilsson and Richard Waldinger acting as advisors. Beth Levin made major contributions to NANOKLAUS's verb-acquisition algorithm. Paul Asente assisted in the testing of the demonstration system. Barbara Grosz, Earl Sacerdoti, and Daniel Sagalowicz provided very useful criticisms of early drafts of this paper. 1. 2. 3* 4. 5. 6. 7. 8. 9. REFERENCES R. Davis, "Interactive Transfer of Expertise: Acquisition of New Inference Rules," Proc. 5th International Joint Conference on Artmar Intelligence, Cambridge, Massaczsetts, pp. 321-328 (August 1977). T. G. Dietterich and R. S. Michalski, "Learning and Generalization of Characteristic Descriptions: Evaluation Criteria and Comparative Review of Selected Methods," Proc. 6th International Joint Conference on Artificial Intelligence, Tokyo, Japan, pp. 223-231 (August 1979). J. A. Fodor, The Language Of Thought, pp. 124- 156, (Thomas Y. Crowell CoFaz York, New York 1975). G. G. Hendrix, "The LIFER Manual: A Guide to Building Practical Natural Language Interfaces," Technical Note 138, Artificial Intelligence Center, Stanford Research Institute, Menlo Park, California (February 1977). G. G. Hendrix, "Encoding Knowledge in Partitioned Networks." in Associative Networks I - The Representation and Use of Knowledge in -- Computers, N. V. Findler, ed.-(Academic Press, New York, New York 1979). G. G. Hendrix, E. D. Sacerdoti, D. Sagalowicz and J. Slocum, "Developing a Natural Language Interface to Complex Data," ACM Transactions on Database Systems, Vol. 3, No. 2 (June 1978).------ P. H . Lindsay and D. A. Norman, Human Information Processing. (Academicxs, New York, New York, 1972.) T. M. Mitchell, "Version Spaces: a Cand Elimination Approach to Rule Learning," 5th International Joint Conference on - idate Proc. Artificial Intelligence, Cambridge, Massachusetts, pp. 305-310 (August 1977). R. Moore, "Reasoning from Incomplete Knowledge in a Procedural Deduction System," MIT Artificial Intelligence Laboratory, AI-TR-347, Massechusetts Institute of Technology, Cambridge, Massechusetts (1975). IO. 11. J. J. Robinson, "DIAGRAM: an Extendable Grammar for Natural Language Dialogue," Technical Note 205, Artificial Intelligence Center, SRI International, Menlo Park, California (February 1980). P. H. Winston, "Learning Structural Descriptions from Examples," Chapter 5 in The Psychology of Computer Vision, P. H. Wins%%, ed. (McGrawxi.11 Book Company, New York, New York (1975). 239

1980

21

14

SELF-CORRECTING GENERALIZATION Stephell B. Wilitehill Deyartment of Irlformation and Computer Science University of California at Irvine Irvine, Ca 92717 ABSTRACT A system is described which creates and generalizes rules from examples. The system can recover from an initially misleading input sequence bY keeping evidence which supports (or doesn't support) a given generalization. By undoing over-generalizations, the system maintains a minimal set of rules for a given set of inputs. I GENERALIZATION Many programs have been written which generalize examples into rules. Soloway[5] generalizes the rules of baseball from examples. Hayes-Roth[3] and Winston[8] yeneralize common properties in structural descriptions. Vere[G] has formalized yeneralization for several applications. If a program maintains a current hypothesis about the rule set as it sees new examples it is said to generalize incrementally. A program that incrementally forms generalizations may be sensitive to the order in which examples are presented. If exceptional examples are encountered first, the program may over-generalize. If the program is to recover and undo the over-generalization it must have a certain amount of knowledge about why the over-generalization was made. The system to be described here has this type of self-knowledge. BY associating positive and negative evidence with each generalization, the system is able to reorganize its rules to recover from over-generalizations. Even if it is initially misled by an unusual sequence of inputs it still discovers the most reasonable set of rules. II THE PROBLEM DOMAIN The problem domain chosen for the system is learning language rnorphology from examples. For example, from the words " j urnped" , "walked" and "kissed" we can deduce the rule that the English past tense is formed by adding "ed". The Concept Description Language consists of a set of rules. Each rule is a production consisting of a pair of character strings. When the left-hand side is matched the right-hand side is returned. The left-hand string may optionally contain a I*# which will match zero or more cllaracters. In this case the right hand cklaracter string may contain a '*I and the value the star on the left hand side which was matched is suhstltuted for the star on the riyht hana side. For example the protiuction for t11e example above looks like: *->*Nl. 'l'his *->*ED rule does not always work. From "baked", "related" and "hoped" we see that for words ending in "e" we need only to add an 'Id". This rule is written as *E->*ED. For this domain, the problem bears resemblence Grammatical Inferznce Problem[4]. to the III RELATIONSHIPS BETWEEN RULES Rule Containment. Given rules Pl and p2 I let Sl be the set of strings which match the left-hand side of Pl and S2 the set of strings which match left-hand side of P2. If Sl is a subset of S2 then we say that Pl is contained by P2. This forms a partial ordering of the rules. The Is-a-generalization-of Operator. Given rules Pl and P2, let Sl be the set of strinys which match the left-hand side of Pl and S2 the set which match the left-hand side of P2. If Pl contains P2 and if Pl and P2 produce the same result for every element in S2 trlen Pl is d yeneralization of P2. This is also a partial ordering. Note the distinction between the containment operator and the is-a-qeneralization-of operator. Basically, containment deals with the left-hand side of a rule. Is-a-generalization-of deals with both sides. An example will clarify this: *->*S contains *K->*KS *->*S is a generalization of *K->*KS *->*S contains *CH->*CHES *->*S and *CH->*CHES are unrelated by generalization By definition, if Pl is a generalization of P2, Pl contains P2. The converse is not necessarily true. is If Pl is a generalization of P2 and Pl a generalization of P3 then Pl is a a generalization of no other common generalization. Roughly, the maximal common generalization is the one which captures all common features of the rules being generalized. E'o r example, given WALK->WALKED and TALK->TALKED, possible generalizations are: *->*ED, *K->*KED, *LK->*LKED and *ALK->*ALKED. The last one, *ALK->*ALKED is the maximal one. In the concept description language we are using all common generalizations are related on the is-a-generalization-of operator. Therefore in our domain the maximal common yeneralization is unique. IV ORGANIZATION OF - RULES The rules and their evidence are organized in a tree structure. At the top level the rules are organized as a rule list. A rule list is a list of rules partially ordered on the containment operator. No rule may be contained by a rule which precedes it in the list. Associated with most rules is some evidence which is itself in the form of another rule list. The only rules without evidence are the example pairs whose evidence is fact. These correspond to terminal nodes in the evidence tree. If a rule Rl would contain a rule R2 which follows it then Rl is marked as being blocked by R2. If Rl blocks R2 then evidence for Rl is negative evidence for H2. The positive evidence consists of those rules which were generalized into the current generalization. Negative evidence for a generalization G is all the evidence of generalizations that are blocked by G. .Thus when *->*ES blocks (*N->*NS + *K->*KS == > *->*S) that is negative evidence for *->*ES. The evidence described here is much like evidence used for explanation in [ll or to maintain beliefs in [2]. In our system the evidence is used for reorganization, but it could be used for these other purposes as well. Hule Application and Conflict Hesolution. When the rule interpreter produces a response, it is as if it finds a11 rules which match the given input and tnen uses the one which doesn't contain any of the others (informally, the one with the least general left-hand side). In reality the rules and rule interpreter are organized so that the first rule that matches is the desired one. Inserting New Rules. If a rule has produced the ----EOrrect result, the new example pair is inserted into the evidence list for the rule. If the rule system has not produced the correct result the rule is inserted in the main rule list before the first rule witn which it will generalize. If it will not generalize with any rule, it is inserted before the first rule that contains it. The same rule insertion algorithm is used to insert new rules or evidence. This means that generalizations take place in an evidence list in the same way that they do in the main rule list. V SYSTEM REORGANIZATION Each blocked generalization has knowledge about which generalization is blocking it. Whenever evidence for a blocked generalization Gl is entered into the rule structure, we check to see if there is now more evidence for Gl than for the blocking generalization G2. If so, G2 is moved to the position in the rule list immediately preceding Gl, G2 is marked as being blocked by Gl and Gl is no longer marked as being blocked. There are several choices on how to compare positive and negative evidence. The one chosen is to count how much direct evidence there is for a rule. Direct evidence is that evidence found in the top level rule list in the evidence tree. Another metnod which was reJected for this application is to count the total number of pieces of evidence in the evidence tree. The first method was chosell because *Ch->*CHES and *X->*XES are exceptions to *-> *s (rather than "A->*AS, *B->*BS, *C->*CS, etc. being exceptions to *->*ES) because there is more direct evidence for *->*S (rules like *A->*AS) than for *->*ES. Even if half the words in English used *CH->*CHES this would still be an exception to *->*s. This method produces the most reasonable set of rules. The system has been tested on pluralizing French adjectives. French has a much more complicated morphology than English, having not only exceptions to rules but also exceptions to exceptions. The system was found to produce the same rules for pluralizing French adjectives as those found in a French-English dictionary . A detailed example of this appears in [7]. VI UNDOING GENERALIZATIONS - AN EXAMPLE - The system is written in MLISP and runs on UC1 LISP. The following example was annotated from a trace of the system in operation. The rules and evidence are printed as a tree. The evidence for a node is indented from the parent node in the printout. INPUT WORD? church WHAT IS RESPONSE? churches INPUT WORD? match WHAT IS RESPONSE? matches 241 INPUT WORD? bus WHAT IS RESPONSE? buses RULES: *->*ES BUS->BUSCS *CH->*CHES MATCH->MATCHES ChUHCH->CHURCHES At this point we have over-generalized. We will find this out later. The only rule seen by tile rule interpreter is *-> *ES. BUS->BUSES arid *CH->*CdES are evidence for *->*ES. MATCH->MATCHES and CHURCH->CHURCHhS are evidence for the *CH->*CHES rule (which is itself evidence). INPUT WORD? book IS RESPONSE BOOKES? n WHAT IS RESPONSE? books INPUT WORD? back IS RESPONSE BACKES? n WHAT IS RESPONSE? backs RULES: B*K->B*KS BACK->BACKS BOOK->BOOKS *->*ES BUS->BUSES "CH->*CHES MATCH->MATCHES CHURCH->CHURCHES What should be regular cases are treated as exceptions. INPUT WORD? car IS RESPONSE CARES? n WHAT IS RESPONSE? cars RULES: (*->*s> CAR->CAKS B*K->B*KS BACK->BACKS BOOK->BOOKS *->*ES Bus->i3usLs *CH->*CfiES MATCH->MATCHES CHURCH->ChURCHES At this point we want to make the generalization *->*s but this generalization is blocked by *->*ES. We make the generalization but mark it as blocked. The parentheses indicated that the rule is blocked. The only productions seen by the production system are: CAR->CARS, B*K->B*KS and *->*ES. The blockage of *->*S is negative evidence for *->*lzs. The system will detect that the rules are in the wrong order when there is more evidence for *->*S (and hence against *->*ES) than there is for *->*ES. At this point there is just as much negative evidence as positive (looking down one level in the evidence tree). INPUT WORD? bat IS HESPGNSE BATES? n WHAT Is RESPONSE? bats RULES: (*->*ES) BUS->BUSES *CH->*ChES MATCH->kATCHES CHURCH->CHUHCHES *-> *s CAR->CAhS B*K->B*KS BACK->BACKS BOOK->BOOKS BAT->BATS This addition negative evidence for *->*ES has caused a reordering of the rules. *->*ES is now blocked by *->*s (as it should be). INPUT WORD? house IS RESPONSE HOUSES? y INPUT WORD? bunch IS RESPONSE BUNCHES? y The system now has a properly ordered rule set and can handle both regular and irregular cases. VII CONCLUSIONS BY giving a generalization program some self-knowledge it can recover from initially misleading input sequences. This introspection can be achieved by associating positive and negative evidence with generalizations. Without knowledye about what led to a generalization, it is not possible to undo the generalization. The system described here discovers the most reasonable set of morphological rules for a yiven language construct (the set found in a dictionary) reyardless of the input sequence. Tne choice of language morpllology as the problem domain was arbitrary. Any domain with a concept descriptiorl language whose maximal common generalization is unique would serve just as well. Furtlier work is needed for concept description languages whose maximal common generalization is not necessarily unique. Any incremental generalization program could irnprove its ability to recover from misleading input by applying the techniques described. [II [21 c31 [41 c51 [61 [71 [al REFERENCES Bechtel, R., Morris, P. and Kibler, D., "Incremental Deduction in a Real-time Environment", Canadian Society for the Computational Studies of -- Intelligence (May 1980). Doyle, J., "A Glimpse of Truth Main- terlance", 6-IJCHI (1979), 232-237. Hayes-Rotll, F . and McDermott, J . , “Knowledye Acquisition from Structurai Descriptions", Department of Computer Sciellce, Carnegie-Mellon Univ. (1976). Hunt, E. "Artificial Intelligence" 1975 Soloway, E. and Riseman, E., "Levels of Pattern Description in Learning", 5-IJCAI (1977), 801-811. Vere, S., "Induction of Relational Productions in the Presence of Background Information", 5-IJCAI (19771, 349-355. Whitehill, S., "Self-correcting Gener- alization". U.C. Irvine C.S. Tech Report no. 149. (June 1980). Winst on, P - , " Le Descr iptio ns From Techn ical Report 231 ar EX ning ample (1970 Structural s " , MIT-AI 242

1980

22

15

intelligent Retrieval Planning Jonathan J. King Computer Science Department Stanford University A. introduction artificial intelligent retrieval planning is the application of intelligence techniques to the task of efficient retrieval of information from very large databases. ’ Using such techniques, significant increases in efficiency can be obtained. Some of these improvements are not available through standard methods of database query optimization. lntelligent retrieval planning presents interesting issues related to other artificial intelligence planning research: planning with limited resources[2], optimizing the combined planning and execution process[9], and pursuing plans whose success depends upon the current contents of the database[S]. An experimental system has been implemented to demonstrate the novel kinds of query optimizations and to test strategies for controlling the inference of constraints. The problem of query optimization has arisen wlth the development of high level logical data models and nonprocedural query languages ([I], [3]). These free a user from the need to understand the physical organlzation of the database when posing a query. However, the user’s statement of the query may lead to very inefficient processing. Standard techniques of query optimization WI, Cl 11, Cl211 manipulate the set of retrieval operations contained in the query to find a relatively inexpensive sequence. The manipulations are independent of the meaning of the query, depending entirely on such factors as the size of the referenced files. The essential advance of intelligent retrieval planning over standard techniques of database query optimization is to combine knowledge about the semantics of the application domain with knowledge about the physical organization of the database. Domain knowledge makes it possible to use the constraints in a database query to infer additional constraints which the retrieved data *must satisfy. These additional constraints may make it possible to use more efficient retrieval operations or permit the execution of a sequence of operations that has a lower cost. Knowledge of the physical organization of the database can be used to limit the attempts to make such inferences so that the combined process of retrieval and inference is cost effective. ” The research described here is part of the Knowledge Base Management System Project at Stanford and SRI, supported by the Advanced Research Projects Agency of the Department of Defense under contract MDA908-77-C-0822. 8. Findino semantic eauivaients of a database auerv The techniques of intelligent retrieval planning will be illustrated with a simple example relational database with data about the deliveries of cargoes by ships to ports. The database contains three files, SHIPS, PORTS, and VISITS, with the attributes indicated: SHIPS: (Shipname Type Length Draft Capacity) PORTS: (Portname Country Depth Facilities) VISITS: (Ship Port Date Cargo Quantity) Semantic knowledge of the application domain is represented as a set of rules. The database is forced, via update restrictions, to conform to this set of rules. The general semantic knowledge for our sample database consists of these rules: Rule RI. “A ship can visit a port only if the ship’s draft is less than the channel depth of the port.” Rule R2. “A ship can deliver no more cargo than its rated capacity.” Rule R3. “Only liquefied natural gas (LNG) is delivered to ports that are specialized LNG terminals.” Rule R4. “Only tankers deliver oil”. Rule R5. “Only tankers can be over 600 feet long.” During intelligent retrieval planning, the use of the rules is shifted from checking updates to inferring constraints. That is, given certain query constraints, it is Possible to infer new constraints that the desired items must meet. For example, suppose a query requests the names of all ships that are longer than 650 feet. By rule R5, it can be inferred that a semantically equivalent retrieval request is for the names of tankers that are longer than 860 feet. This inferred description of the items to be retrieved may permit more efficient processing than the origlnal description. 243 C. The physical orqanization of a database Inferred semantically equivalent sets of constraints can be exploited for intelligent retrieval only if the physical organization of the database, and hence the cost of processing queries, is taken into account. Often, the physical organization has been arranged so that the cost of retrieving a restricted subset of data depends upon the data attributes that have been restricted. For instance, a file may have an auxiliary “index” on one of its attributes. If such an index exists, then the data pages that contain items that meet a constraint on that attribute can be identified directly and only those pages will be fetched. An indexed scan will be much less expensive than a scan through an entire file, measured in terms of pages fetched from disk. A discussion of retrieval costs for-different physical database organizations is contained in [4]. Thus, given a query that constrains only unindexed attributes. a reasonable semantic retrieval strategy (subject to qualifications discussed in [4]) is to attemdt to infer constraints on indexed attributes. Suppose that the SHIPS file has an index on the Type atf%bute. In that case. the best way to retrieve all the ships longer than 650. feet would be to fetch all the tankers by means of an indexed scan on Type, and then to check the Length value of each record fetched into main memory by that- scan. D. Novel auery optimization based on the use of domain semantics A query optimization method that uses domain semantics is interesting to the extent that it achieves significant increases in efficiency that are not available by other methods. One unique strategy that can arise when semantics are considered is the inclusion of an extra file in the set of files examined when a query is processed. For example, suppose a query requests the quantity of liquefied natural gas delivered for each known visit to ports with a channel depth of less than 20 feet. With no inference, a typical query processor would retrieve all PORTS records with a Depth value of less than 20. For each one, it would retrieve all VISITS whose Port attribute was the same as the Portname for the PORTS record and whose Cargo attribute was liquefied natural gas. The cost of the retrieval varies as the product of the sizes of the PORTS and VISITS files. However, with appropriate rules and indexes, intelligent retrieval planning can provide a much faster retrieval method. Suppose that the VISITS file has an index on the Ship attribute. In effect, this means that the database has been set up to provide inexpensive access from each ship to the set of its visits, while the set of visits to a specific port is much costlier to retrieve. Using rule Rl, it can be inferred that the visits requested by the query could have been made only by ships with a draft of less than 20 feet. It is now possible to retrieve SHIPS with Draft less than 20, then retrieve their associated VISITS (using the index), and finally, for each VISITS record with a Cargo value of liquefied natural gas, retrieve the associated PORTS record to check the value of Depth. If the Draft constraint substantially restricts SHIPS (and therefore the associated VISITS as well), then the overall cost will be much lower than that of the straightforward method, despite the fact that an extra file and an extra retrieval operation have been added. In a simulation test of this method using a cost model based on the System R relational database system[7] in which the VISITS file Is much larger than the PORTS and SHIPS files, the simulated retrieval cost was reduced by more than order of magnitude. E. Controllinq the inference of additional constraints Intelligent retrieval planning is complicated by the need to weigh possible gains in retrieval efficiency against the cost of performing inferences. The amount of planning done in the intelligent retrieval planning system in processing a particular query is determined by the cost of answering the unimproved query, and the possible improvements. The inference control mechanism has these key features: (1) The specific retrieval problem determines which constraints to try to infer (for example, an attempt is made to add constraints to indexed fields). (2) Knowledge about both the structure and the content of the database determines the effort to devote to attempting some inference. (3) Retrieval from the database is an inherent part of the inference process. The ability to carry out an inference (and hence the shape of the whole retrieval plan) may depend upon the current contents of the database. These features can be illustrated briefly in another example. Suppose the VISITS file is indexed only on Cargo, and a query requests data on visits to the port of Zamboanga. The retrieval strategy mentioned in section 3 suggests an attempt to infer a constraint on Cargo from the given constraint on Port. Given the number of records in the VISITS file, it is possible to compute the effort needed to perform a sequential scan. The effort alloted to inference will be a function of this. There is no guarantee that a helpful constraint can be found for any particular query. This suggests a policy to allot to the inference process a fixed small fraction of the effort which the original retrieval would take. With such a policy, the effort to plan the retrieval will result in a minor increase in response time if the inference attempt fails, but may provide a major improvement if It succeeds. Although the policy is intuitively plausible, other strategies for alloting effort during problem solving under uncertainty, such as those discussed in [S], are being investigated. Control of the inference process can be viewed as control of the moves in a space of constraints on attributes. Constraints can be moved either by applying a rule, by retrieving items restricted on one attribute and observing their values on other attributes, or by matching constraints on attributes defined on the same underlying set of entities. Continuing the example, starting with a constraint on the Port attribute of VISITS, new constraints can be found by retrieving from VISITS or by assigning the 244 value “Zamboanga” to the Portname field of PORTS. The first choice is rejected because the objective is to reduce the cost of that very retrieval. With a constraint on Portname in PORTS, a retrieval from PORTS can be performed. In this case, just a single record will be obtained because Portname is the unique identifier in that file. With appropriate access methods, such as hashing, the retrieval will be very inexpensive. When the PORTS record for l@Zamboanga” has been obtained, rules Rl and R3 may apply. if rule R3 applies, that is, if Zamboanga is a specialized liquefied natural gas terminal, then a strong constraint will be obtained on the goal attribute Cargo, and retrieval from VISITS will take place by means of an indexed scan rather than by means of a more expensive complete scan. If the data on Zamboanga does not support that inference, then other inference paths will have to be considered. This illustrates the possible dependence of retrieval planning on the current contents of the database. References 1. Codd, E.F., A relational model for large shared dala banks Commun ACM 13:6 (19701, 377-387. IA- 2. Garvey, Thomas D., Perceptual strategies for purposive vision, Technical Note 117, SRI International, Menlo Park, California, September 1976. 3. Kim, Won, Relational database systems, ACM Computinq Surveys 11:3 (19791, 186-212. 4. King, Jonathan J., Exploring the use of domain knowledge for query processing efficiency, Technical Report HPP-79-30, Heuristic Programming Project, Computer Science Department, Stanford University, December 1979. The cost of each inference step: generating new inference path nodes, testing rules, and retrieving from the database itself, is taken from the allotment of planning resources. Planning terminates if a strong goal constraint is found, if no potential inference path can be extended, or if planning resources are exhausted. F. Conclusion 6. Klahr, Philip, Planning techniques for rule selection in deductive question-answering, In Pattern Directed Inference Systems, D.A.Waterman and F. Hayes-Roth (Eds.), Academic Press, 1978. 6. Moore, Robert C., Handling complex queries in a distributed data base, Technical Note 170, SRI International, Menlo Park, California, October, 1979. Intelligent retrieval planning can provide novel and significant improvements in query processing efficiency. it draws on a knowledge of the physical organization of the database and on semantic knowledge of the application modelled in the database. The outcome of retrieval planning, both the retrieval method chosen and its cost, can depend upon the current contents of the database. An experimental system exists that performs inferences on queries stated in a subset of the SODA relational database query language[6]. The system uses a simple retrieval cost model to select the least expensive semantically equivalent expression of the retrieval request. The cost model is used in conjunction with a planning executive to limit the inference of additional constraints. Work Is under way to codify intelligent retrieval strategies which, though they are specific to a given class of physical database organizations, are independent of the application domain. The eventual aim of this work is to develop a system which, given the set of domain rules and the description of the physical organization for a database, can provide the functions of intelligent retrieval planning described in this paper, much as the EMYCiN system[lO] provides knowledge acquisition functions independently of the knowledge base to be built. 7. Selinger, P. Griffiths et. al. Access path selection in a relational database management system, In Proc. ACM-SIGMOD 1979 Boston, Mass., pp. 23-34. -9 8. Smith J.M. and P. Chang, Optimizing the performance of a relational algebra data base interface, Commun ACM 18:lO (1975), 568-579. -- 9. Sproull, Robert F., Strategy construction using a synthesis of heuristic and decision-theoretic methods, Report CSL-77-2, Xerox Palo Alto Research Center, Palo Alto, California, July 1977. 10. Van Melle, William, A domain-independent production- rule system for consulation programs, IN Proc. IJCAI-79 Tokyo, Japan, 1979, pp. 923-926. 11. Yao, S. Bing, Optimization of query evaluation algorithms, ACM Transactions on Database Systems 4:2 (1979) 133-l 55. 12. Youssefi, Karel A. Allen, Query processing for a relational database system, Memorandum UCB/ERL M78/3, Electronics Research Laboratory, University of California, Berkeley, California, January 1978. Acknowledgments Many thanks for perceptive comments by Jim Bennett, Jim Davidson, Larry Fagan and Jerry Kaplan of Stanford University, and Barbara Grosz of SRI International. 245

1980

23

16

A THEORY OF METRIC SPATIAL INFERENCE Drew McDermott Yale University Department of Computer Science New Haven, Connecticut ABSTRACT* Efficient and robust spatial reasoning requires that the properties of real space be taken seriously. One approach to doing this is to assimilate facts into a "fuzzy map" of the positions and orientations of the objects involved in those facts. Then many inferences about distances and directions may be made by "just looking" at the map, to determine bounds on quantities of interest. For flexibility, there must be many frames of reference with respect to which coordinates are measure. The resulting representation supports many tasks, including finding routes from one place to another. ***Jr In the past, AI researchers have often sought to reduce spatial reasoning to topological reasoning. [4, 6, 81 For example, the important problem of finding routes was analyzed as the problem of finding a path through a network or tree of known places. This sort of formulation throws away the basic fact that a route exists in real physical space regardless of our knowledge of any of the places along the way. So a network-based algorithm will fail to exhibit two important phenomena of route-finding: > Often you know roughly what direction to go in without having any idea of the details of the path, or even if the path is physically possible. > You can tell immediately that you don't know how to get to a place, just by verifying that you don't know the direction to that place. There are many other problems that a topological approach fails to treat adequately. Here are some of the problems we (Ernie Davis, Mark Zbikowski and I> have worked on: > How are metric facts, such as "A is about 5 miles from B" or “The direction from A to B is north" to be stored? > How are queries such as "Is it farther A to B than from A to C?" to be answered? from > Given a large set of objects and facts relating them, how do you find the objects that might be near some position? or with some orientation? Some of these problems have received more of our attention than others. In what follows, I will sketch our approach, the details of various algorithms and data structures, and the results we have so far. All of our solutions revolve around keeping track of the fuzzy coordinates of objects in various frames of reference. That is, to store -- metric facts about objects, the system tries to find, for each object, the ranges in which quantities like its X and Y coordinates, orientation and dimensions lie, with respect to convenient coordinate systems. The set of all the frames and coordinates is called a fuzzy map. We represent shapes as prototypes plus modifications. 13, 51 The domain we have used is the map of Yale University, from which most of my examples will be taken. To tasks: date we have written programs to do (1) Given a stream of metric relationships, create a fuzzy map of the objects involved. Research supported by NSF under contract MCS7803599 246 (2) Given a fuzzy map, test the consistency of a relationship or find the value of a term. (3) Given a fuzzy map, find objects with a position and orientation near some given value. (4) Plot a course around objects or through conduits discovered using (3). So far we have invested most of our effort in the study of task (21, what I described as "just looking" at the map to see what's true. This actually involves using hill climbing to see if a relationship can be satisfied, or to find the possible range of values of a term. So, in Figure 1, to answer the query 'What's the distance from Kline to Sterling in meters?" the system plunks down two points in the fuzz boxes of Kline and Sterling, and moves them as close together, then as far apart, as it can. To answer the query "Is Kline closer to Dunham than to Sterling?" it looks for a configuration of points from the fuzz boxes of Kline, Dunham and Sterling in which Kline is further from Dunham than Sterling. (Since it fails to find it, the answer to the query is "Yes.") The same hill-climbing algorithm is used for task (11, the assimilation of facts into a map. In this case, the object is to find the smallest and largest possible values of each "primitive term" involved in a relationship. (A primitive term is a quantity like (X A) or (LENGTH A) that characterizes an object's position, orientation or dimensions. More complicated terms, like (DIST A B), are functions of primitive terms.) The new, smaller range of a primitive term is then stored for future reference. This device, called fuzz constriction, suffices to capture many spatial facts* However, it can happen that the range of a primitive term does not change at all after a new fact is assimilated, especially when the fact relates objects about which little is known. For example, if we tell the system that the orientation of Sterling Library is the same as the orientation of Becton Library, when it knows nothing about their orientations with respect to Yale, this new fact doesn't constrain them any further. In this case, mere fuzz constriction has failed to capture the new fact. The solution is to introduce a new frame of reference F, and to indicate that (ORIENT STERLING) = (ORIENT BECTON) = 0.0 in this frame, while the orientation of F is completely fuzzy (from 0 to 2pi) with respect to Yale. The machinery introduced so far enables us to retrieve characteristics of given objects. It is also important to be able to retrieve objects given their spatial characteristics (task (3)). For example, if you are trying to get from one place to another in a city, YOU will want to know what streets to use, i.e., how to find the nearest 'long porous objects" with approximately the right position and orientation. This is a job for k-d trees of the kind devised by Bentley. [l, 21 In these trees, a large set of objects is broken down into manageable chunks by an obvious generalization of binary search: repeatedly discriminate on each coordinate. An example is shown in Figure 2. The original version of k-d trees was designed to work on data bases in which all primitive terms have known values. In our application, most primitive terms can only be assigned ranges. To deal with this problem, we take the following tack: if a given attribute of an object is "very fuzzy" (e.g., its orientation is known only to lie between 0 and 2 pi>, then we do not index it on that attribute. But if it is only moderately fuzzy, 247 I I then we index it as though its value were the midpoint of its range. This requires that on retrieval we be willing to look around a little for objects fitting our requirements. That is, if we need to find a street with a given orientation, we keep relaxing the requirement until we find one. Obviously, a street found after many relaxations is only a plausible candidate, and must be proven to actually work; and the process must be terminated when it is unlikely to give worthwhile results. 02 route means depends on the density of the region to be traversed. If it is mostly open, then the problem is to plan to avoid obstacles; if it is mostly obstacle, then the problem is to plan to use conduits. Either way, the system must find objects with known properties (e.g., "open and pointing in the right direction" or "filled and lying in my way"). To summarize, our main results so far are these: representing space as a structure of multiple frames of reference, within which objects have fuzzy positions, is efficient and robust. In this context, assimilation is the process of constricting fuzz or creating new frames to capture a new fact. Route finding involves computing a fuzzy vector from where you are to where you want to be, then finding objects which can help or hinder your progress, and altering the plan to take them into account. Many problems still remain. The assimilation algorithm needs improvement. The route finder has not yet been completed or connected with the assimilation and retrieval algorithms. As yet we have not implemented a (simulated) route executor, although this is a high priority. AcknowledPements: Ernie Davis and Mark Zbikowski have helped develop many of the ideas in this paper, and made suggestions for improving the exposition. [ll 121 [31 [41 [51 [61 [71 181 REFERENCES Jon Bentley 1975 Multidimensional binary search trees used for associative searching, Comm. ACM 18, no. 9, PP. 509-517 Jon Bentley and Jerome Friedman 1979 Data I structures for range searching, Comnut. Surveys ll, no. 4, pp. 397-409 John Hollerbach 1975 Hierarchical shape description of objects by selection and modification of prototypes, Cambridge: MIT AI Laboratory Technical Report 346 Benjamin Kuipers 1978 Modeling spatial knowledge, Cognitive Science 2_, no. 2, p. 129 David Marr and H. Keith Nishihara 1977 Representation and recognition of the spatial organization of three dimensional shapes, Cambridge: MIT AI Laboratory Memo 416 Drew McDermott 1974 Assimilation of new information by a natural language-understanding w-w=-, Cambridge: MIT AI Laboratory Technical Report 291 Drew McDermott 1980 Spatial inferences with ground, metric formulas on simple objects, New Haven: Yale Computer Science Research Report 173 James Meehan 1976 The metanovel: writing stories by computer, New Haven: Yale Computer Science Research Report 74 This algorithm for finding objects that might have given characteristics is used by our route-finding programs. Exactly what finding a 248

1980

24

17

DESIGN SKETCH FOR A MILLION-ELEMENT NETL MACHINE Scott E. Fahlman Carnegie-Mellon University Department of Computer Science Pittsburgh, Pennsylvania 152 13 Abstract This paper describes (very briefly) a parallel hardware implementation for NETL-type semantic network memories. A million-element system can be built with about 7000 IC chips, including 4000 64K RAM chips. This compares favorably with the hardware cost of holding the same body of knowledge in a standard computer memory, and offers significant advantages in flexibility of access and the speed of performing certain searches and deductions. 1. Introduction In [l] I presented a scheme for representing real-world knowledge in the form of a hardware semantic network. In this scheme, called NETL, each node and link in the network is a very simple hardware processing element capable of passing single-bit markers through the network in parallel. This marker-passing is under the overall control of an external serial processor. By exploiting the parallelism of this marker-passing operation, we can perform searches, set intersections, inheritance of properties and descriptions, multiple-context operations, and certain other important operations much faster than is possible on a serial machine. These new abilities make it possible to dispense with hand-crafted search-guiding heuristics for each domain and with many of the other procedural attachments found in the standard Al approaches to representing knowledge. In addition to the ditficuity of creating such procedures, and the very great difficulty of getting the machine to create them automatically, I argue that the heuristic systems are brittle because they gain their efficiency by ignoring much of the search space. NETL, on the other hand, looks at every piece of information that might be relevant to the problem at hand and can afford to do so because it does not have to look at each piece of information serially. NETL has been viewed by many in the Al community as an interesting metaphor and a promising direction ‘for future research, but not as a practical solution to current Al problems because of the apparently impossible cost of implementing a large NETL system with current technology. The problem is not that the hardware for the nodes and links is too costly -- hundreds or even thousands of these elements can be packed onto a single VLSI chip. Rather, the problem is in forming new private-line connections (wires) between particular nodes and links as new information is added to the network. These connections cannot be implemented as signals on a single shared bus, since then all of 1 This research was sponsored by the Defense Advanced Research Projects Agency (DOD), ARPA Order No. 3507, monitored by the Air Force Avionics Laboratory Under Contract F33615-78-C-1551. The views and conclusions contained in this document are those of the author and should not be interpreted as representing the official policies, either expressed or implied, of the Defense Advanced Research Projects Agency or the US Government. the parallelism would be lost. Indeed, it is in the pattern of connecting wires, and not in the hardware nodes and links, that most of the information in the semantic network memory resides. A large switching network, similar to the telephone switching network, can be used in place of physical wires, but for a network of a million elements one would need the functional equivalent of a crossbar switch with 4 x IO’* crosspoints. Such a switch would be impossibly expensive to build by conventional means. In the past year I have developed a multi-level time-shared organization for switching networks which makes it possible to implement large NETL systems very cheaply. This interconnection scheme, which I tail a hashnet because some of its internal connections are wired up in a random pattern, has many possible uses in non-Al applications; it is described in its general form in another paper [2]. In this paper I will briefly describe a preliminary design, based on the hashnet scheme, for a semantic network memory with lo6 NETL elements. (An “element” in NETL is the combination of a single node and a four-wire link.) A million-element NETL system is lo-20 times larger than the largest Al knowledge bases in current use, and it offers substantial advantages in speed and flexibility of access. It is an open question whether a knowledge-base of this size will be adequate for common-sense story understanding, but a system of lo6 NETL elements should hold enough knowledge for substantial expertise in a variety of more specialized domains. In a paper of this length I will be able to sketch only the broad outlines of the design -- for a more complete account see [3]. The NETL machine itself, excluding the serial control computer, requires about 7000 VLSI chips, 4000 of which are commercial 64K dynamic RAM chips. (See the parts list, table 1.) As we will see later, with the same 64K memory technology it would require a comparable number of chips to store the same body of information in a conventional Planner-style data base, assuming that the entire data base is kept in a computer’s main memory and not on disk. So, far from being impossibly expensive, this scheme is quite competititve with standard random-access memory organizations. I am about to seek funding to build a million-element prototype machine within the next two or three years. The 64K RAM chips are not available today in sufficient quantities, and may be quite expensive for the next couple of years. The prototype machine will be designed so that 16K RAMS can be substituted if necessary, giving us a 256K element machine to use until the 64K RAM chips are obtained. 2. Requirements of NETL Figure 1 shows a basic NETL element as it was originally conceived. Commands from the control computer are received over the common party-line bus. The applicability of any command to a given element depends on the element’s unique serial number, on the state of 16 write-once flag bits which indicate what type of node or link the element represents, and on 249 the state of 16 read-write marker bits which indicate the current state of the element. These marker bits represent the short-term memory in the system. Also present are some number (4 in this design) of distinct link wires, and a node terminal to which link wires from other elements can be connected. Commands typically specify that all elements with a certain bit-pattern should send a one-bit signal across incoming or outgoing link wires, and that any element receiving such a signal should set or clear certain marker bits. It is also possible to address a command to a specific element, or to get any element with a certain marker pattern to report its serial number over the common bus. Using these commands, it is possible to propagate markers through the network Quillian-style or to control the marker propagation in any number of more precise ways. For details, see [l], especially section 2.3 and appendix A.1. In a million-element design, then, we have 4 sets of lo6 link wires to connect to IO6 node terminals, each by a private line. A link wire is connected to only one node terminal, but a node terminal may have any number of link wires attached to it. Unlike the telephone system, this network must support all 4 million connections simultaneously; once a connection is made, it becomes part of the system’s long-term memory, and a connection is seldom, if ever, released. As the system learns new information, new links are wired up one at a time, and this must be done without disturbing the connections already in use. If the same network hardware is to be used for different knowledge bases at different times, it must be possible to drop one set of connections and replace them with a new set. A few additional constraints will help us to separate interesting designs from uninteresting ones. If we want to obtain a roughly human-like level of performance in our knowledge base, the transmission of a bit from one element to another can take as long as a few milliseconds. Since answering a simple question -- for example, determining whether an elephant can also be a cabbage -w takes something like 20 to 60 basic marker-propagation cycles, a propagation time of 5 milliseconds gives a response time of .l to .3 seconds. This figure is independent of the total size of the network. This means that some degree of parallelism is essential, but that with microsecond-speed technologies there is room for some time-sharing of the hardware as well. New connections are established individually, and setting them up can take somewhat more time than simple propagations: humans are able to add only a few items to long-term memory per second. If an attempt to create a new connection should fail occasionally, nothing disastrous occurs -- the system simply skips over the link it is trying to wire up and goes on to the next free one. 3. The NETL Machine Design As mentioned earlier, the key problem here is to build a switching network for connecting link-wires to nodes. Since the four wires of a link are used at different times, we can think of this switch as four separate sub-netowrks, each with lo6 inputs, lo6 outputs, and with all lo6 connections operating at once. This network must be, in the jargon of network theory, a seldom- blocking network. It must be possible to add new connections one by one, without disturbing any connections that are already present, but some small chance of failure in establishing new connections can be tolerated. Once established, a connection must work reliably, and must be able to transmit one-bit signals in either direction. Note that by “connection” here I mean a setting of the network switches that establishes a path from a given input to a given output; the pysical wiring of the network is of course not altered during use. The basic concept used in this design is to build a 960 x 960 seldom-blocking switching network, then to time-share this network 1024 ways. The number 960 arises from packaging considerations; this gives us a total of 983,040 virtual connections, close enough to one million for our purposes. The 1024 time slices roll by in a regular cycle; a different set of switch settings is used during each slice. There are four sets of 1024 switch settings, corresponding to the four link-wire sub-netowrks. The bits describing the 4096 settings for each switch are stored in random access memory chips between uses. The NETL elements are likewise implemented in a time-shared fashion: 960 element units are implemented in hardware (four to a chip with shared bus decoders), and each of these element devices is time shared 1024 ways. Each NETL element exists in hardware only during its assigned time-slice: most of the time, it exists only as 32 bits of state in a memory chip. Let us begin by considering the construction of a 960 x 960 hashnet without time-sharing. The basic unit of construction is the 15way selector cell shown in figure 2a. This cell connects its input to any of its 15 outputs, according to the contents of a four-bit state register. A value of 0 indicates that the cell is currently unused and that its input is not connected to any output. A single integrated circuit chip can easily hold 15 of these selector cells; the corresponding outputs from each cell are wired together internally, as shown in figure 2b. With assorted power and control lines, this 15 x 15 switching element requires a 48.pin package. To build a 960 x 960 seldom-blocking network out of these elements, we arrange them in four layers with 1920 selector cells (128 chips) in each. (See figure 3.) The outputs of each layer are wired to the inputs of the next layer with a fixed but randomly chosen pattern of wires. Each of the input terminals of the hashnet is wired to 2 selector cells in the first layer; each of the outputs of the hashnet is wired to 2 outputs lines from the last layer of cells. Initially all of the selector cells are in the non-busy slate. As paths through the network are set up, each one uses up one of the selector cells in each layer. Note, however, that half of the selector cells remain unused even when all 960 connections are in place; this spare capacity ensures that the network will have a low chance of blocking even for the last few connections. To set up a new connection from a given input to a given output, we first broadcast a marking signal through the network from the input to all of the selector cells and outputs that can be reached. Only non-busy cells play a role in this process. If this signal reaches the desired output, one of the marked paths is traced back toward the source, with the selector cells along the way being set up appropriately. These cells become busy and will not participate in any other connections. Since the inter-layer wiring of the network is random, and since we are using many fewer switches than are needed for a strictly non-blocking network, we cannot guarantee that a desired connection can be found. We can guarantee that the probability of being unable to find a desired connection is very small. In simulation tests of this design, 100 complete sets of connections were attempted -- 10,000 connections in all -- with only 2 failures. As noted earlier, an occasional failure to find a connection is not disastrous in the NETL application; we just try again with a different link. Instead of using four layers of selector cells, it is possible to get the same effect using only one layer. The output wires of this layer are randomly shuffled and fed back to the inputs; these wires also go to the network’s ouput terminals. Four sets of switch settings are used, corresponding to the four layers of the original network. The signal bits are read into this layer of cells and latched. Using the first-layer switch settings, they are sent out over the appropriate output wires and shuffled back to the inputs where these bits are latched again. Then the second layer setup is used and the signal bits are shuffled again. After four such shuffles, the bits are in the right place and are read from the network’s outputs. We have traded more time for fewer switches and wires; the same number of setup bits are used in either case. To go from a thousand connections to a million, we time-share this network 1024 ways. This requires two modifications to the network. First, instead of having only 4 bits of state for each selector cell (or 16 bits if the shuffling scheme is in use), we need a 25CI different set of bits for each time slice. These bits are read from external memory chips; 4 bits of setup are read in through each selector cell input, followed by the signal bit. Second, since the NETL elements are permanently tied to their assigned time-slices, we need some way to move the signal bits from one time-slice to another. This operation is carried out by time-slice shifter chips. Each of these devices is essentially a 1024-bit shift register. During one cycle of time-slices this shift register is loaded: during each slice, a signal bit is received along with a IO-bit address indicating the slice that the signal bit is to go out on. The address governs where in the shift register the signal bit is loaded. On the next cycle of slices, the bits are shifted out in their new order. An entire layer of 1920 time-shifters is needed (packed 5 to a chip), along with memory chips to hold the lo-bit time-slice addresses. The chance of blocking is minimized if these are placed in the center of the network, between the second and third layers of cells. Some additional chance of blocking is introduced by addition of the shifters to the network, but not too much. In our simulations of an almost-full network with time sharing, we encountered 37 blocked connections in 110,000 attempts. 4. Cost and Performance As can be seen from the parts list, most of the cost of the NETL machine is in the memory chips. In order to keep the number of chips below 10,000 -- a larger machine would be very hard to build and maintain in a university research environment -- 64K dynamic RAM chips have been used wherever possible in the design. The memory associated with the element chips must be either 2K x 8 static RAMS or fast 16K dynamic RAMS for timing reasons. In fact, the limited bit-transfer rate of the memory chips is the principal factor limiting the speed of the network; if 16K x 4 chips were available, the system could be redesigned to run four times as fast. As it currently stands, the system has a basic propagation time of about 5 milliseconds. This is the time required to accept lo6 bits and to steer each of these to its independent destination. This assumes the use of 2K x 8 static RAMS for the element memories and allows for a rather conservative page-mode access time of 200 nanoseconds for the 64K RAM chips. (In page mode, half of the address bits remain constant from one reference to the next.) The 256K element version, using 16K RAM chips, should have a propagation time of 1.25 milliseconds. Since the million-element machine contains 4000 64K RAM chips, the parts cost of the machine is tied closely to the price of these chips. Currently, if you can get 64K RAMS at all, they cost over $100, but the price is likely to drop rapidly in the next two years. If we assume a price of $20 for the 64K RAMS and $10 for the 2Kx8 RAMS, we get a total of roughly $100,000 for the memory chips in the machine. It is also hard to know what price to assign to the three custom chips in the design. An initial layout cost of $150,000 for all three would probably be reasonable, but this cost would occur only once, and the chips themselves should be easy to fabricate. We have not yet thought hard about board-level layout, but rough calculations suggest that the machine would fit onto about 52 super-hex boards of two types. Two standard equipment racks ought to be sufficient for the whole machine. For comparison, it might be worthwhile to calculate the cost of storing the same information in a Planner-style data base in the memory of a serial machine. To make the comparison fair, let us assume that the entire data base is to be kept in main memory, and not paged out onto disk. To store the information in the network itself, in the simplest possible form, would require 80 million bits of memory: 4 million pointers of 20 bits each. This would require 1250 64K RAM chips. A simple table of pointers, of course, would be very slow to use. If we add back-pointers, the figure doubles. If we add even the most limited sort of indexing structure, or store the entries in a hash table or linked list, the arnount of memory doubles again. At this point, we have reached 4000 64K RAM chips, the same number used in the NETL machine. The moral of the story would seem to be that the greater power and flexibility of the NETL organization can be had at little, if any, extra cost. Acknowledgements Ed Frank and Hank Walker provided me with much of the technical information required by this design. The network simulations were programmed and run by Leonard Zubkoff. References ill Fahlman, S. E. NETL: A System for Representing and Using Real-World Knowledge. MIT Press, Cambridge, Mass., 1979. PI Fahlman, S. E. The Hashnet Interconnection Scheme. Technical Report, Carnegie-Mellon University, Department of Computer Science, 1980. [31 Fahlman, S. E. Preliminary Design for a Million-Element NETL Machine. Technical Report, Carnegie-Mellon University, Department of Computer Science, 1980. (forthcoming). Gable 1: Parts List Device Type Custom 15 x 15 Selector (48 pin) Custom 5x 1 K Time Shifter (16 pin) Custom 4x NETL Element Chip (48 pin) 64K Dynamic RAM (for selectors) 64K Dynamic RAM (for shifters) 2Kx8 Static RAM (for elements) Total Device Count Number 128 240 2176 1920 2160 7008 251 I I C%Ol Computer INPUT SELiCT BITS Marker bits (16) Node Terminal Figure 1: NETL Element IN 1 IN 2 IN 3 IN 4 OUT 1 OUT 2 OUT 3 OUT 4 Figure 2A: The Selector Cell Figure 2B: Selectors on 15x15 Chip (Drawn as 4x4 chip.) Link Wires Node Terminals Figure 3: The Basic Washnet A rrangement (simplified) 252

1980

25

18

PERCEPTUAL REASONING IN A HOSTILE ENVIRONMENT Thomas D. Garvey and Martin A. Fischler Artificial Intelligence Center SRI International, Menlo Park, CA 94025 ABSTRACT The thesis of this paper is that perception requires reasoning mechanisms beyond those typically employed in deductive systems. We briefly present some arguments to support this contention, and then offer a framework for a system capable of perceptual reasoning, using sensor- derived information, to survive in a hostile environment. Some of these ideas have been incorporated in a computer program and tested in a simulated environment; a summary of this work and current results are included. I INTRODUCTION Living organisms routinely satisfy critical needs such as recognizing threats, potential mates, food sources, and navigable areas, by extracting relevant information from huge quantities of data assimilated by their senses. How are such "relevant" data detected? We suggest that a reasoning approach that capitalizes on the goal-oriented nature of perception is necessary to define and recognize relevant data. Perception can be characterized as imposing an interpretation on sensory data, within a context defined by a set of loosely specified models. The ability to select appropriate models and match them to physical situations appears to require capabilities beyond those provided by such "standard" paradigms as logical deduction or probabilistic reasoning. The need for extended reasoning techniques for perception is due to certain critical aspects of the problem, several of which we summarize here: * The validity of a perceptual inference (interpretation) is determined solely by the adequacy of the interpretation for successfully carrying out some desired interaction with the environment (as opposed to verification within a "closed" formal axiomatic system). -------- * This work was supported by the Defense Advanced Research Projects Agency under Contracts MDAgOT-79- C-0588, F33615-77-C-1250, and F33615-80-C-1110. Since it is impossible to abstractly model the complete physical environment, the degree to which purely abstract reasoning will be satisfactory is limited. Instead, perception requires tight interaction between modeling/hypothesizing, experimenting (accessing information from the environment), and reasoning/verifying. Reasoning processes that embody concepts from physics, geometry, topology, causation, and temporal and spatial ordering are critical components of any attempt to "understand" an ongoing physical situation. Explicit representations appropriate to these concepts are necessary for a perceptual system that must provide this understanding. These representations are incommensurate and it is not reasonable to attempt to force them into a single monolithic model. There is typically no single, absolutely correct interpretation for sensory data. What is necessary is a "maximally consistent" interpretation, leading to the concept of perception as an optimization problem [l, 21 rather than a deductive problem. Research in perception and image processing at SRI and elsewhere has addressed many of these issues. An early effort focused upon the goal- directed aspect of perception to develop a program capable of planning and executing special-purpose strategies for locating objects in office scenes [31* Research addressing interpretation as an optimization problem includes [l, 2, 41. Current research on an expert system for image interpretation [ 5] h as also considered the strategy-related aspects of determining location in situations involving uncertainty. The most recent work (at SRI) on perceptual reasoning has addressed the problem of assessing the status of a hostile air-defense environment on the basis of information received from a variety of controllable sensors [6]. This work led us to attempt to formulate a theory of perceptual reasoning that highlighted explicit reasoning processes and that dealt with those aspects of perception just described. In the following section, we will use this work as a vehicle to illustrate a paradigm for perceptual reasoning. 253 II PERCEPTUAL REASONING IN A SURVIVAL SITUATION The specific problem addressed was to design a system able to interpret the disposition and operation (i.e., the order of battle or OB) of hostile air-defense units, based on information supplied by sensors carried aboard a penetrating aircraft [6]. The situation may be summarized as follows. A friendly aircraft is faced with the task of penetrating hostile airspace en route to a target behind enemy lines. Along the way, the aircraft will be threatened by a dense network of surface-to-air missiles (SAMs) and antiaircraft artillery (AAAs). The likelihood of safe penetration and return is directly related to the quality of acquired or deduced information about the defense systems. Partial information is furnished by an initial OB, listing known threats at, say, one hour before the flight. Additional knowledge is available in the form of descriptions of enemy equipment, typical deployments, and standard operating procedures. Since the prior OB will not be completely accurate, the information must be augmented with real-time sensory data. The OB forms the starting point for this augmentation. The explicit goal of the overall system is to produce and maintain an accurate OB, detecting and identifying each threat prior to entering its lethal envelope. The density of threats means that this goal will result in conflicting subgoals, from which selection must then be made to ensure that critical data will be received. This must be accomplished by integrating data from imperfect sensors with prior knowledge. The paradigm that was developed for this task is summarized below: (1) Available knowledge is used to create an hypothesized OB that anticipates the developing situation. (2) A plan that attempts to allocate sensors to detect or verify the presence of threats, in an optimal way, is constructed. Sensors are then allocated and operated. (3) Information returned from the sensors is interpreted in the context established by the anticipated situation. Interpretations modify the current OB, and the process is iterated. We will briefly discuss each of these steps. B. Experimentation (Accessing Information from the Environment) The goal of this step is to access information needed to detect or verify the presence of threats inferred in the anticipation step, but not available in the "internal" knowledge base of the system. In general, it might be necessary to define and execute one or more experiments to extract this needed information from the environment. In the more limited context of model instantiation by "passive" sensing, the problem reduces to that of allocating sensor resources to maximize the overall utility of the system; sensing is a specific instance of the more general process of experimentation. First the list of data-acquisition goals is ordered, based on the current state of information about each threat and its lethality. The allocator attempts to assign (a time sliced segment of) a sensor to satisfy each request based on the expected performance of the sensor for that task. Sensor detection capabilities are modeled by a matrix of conditional probabilities. These represent the likelihood that the sensor will correctly identify each threat type, given that at least one instance thereof is in the sensor's field of view. This matrix represents performance under optimal environmental conditions (for the sensor) and is modified for suboptimal conditions by means of a specialized procedure. This representation is compact and circumvents the need to store complete, explicit models describing sensor operation in all possible situations. Similar models describe each sensor's identification and location capabilities. The sensor models are used to compute the utility of allocating each sensor to each of the highest priority threats. These utilities form the basis for the final allocation, which is carried out by a straightforward optimization routine. At the same time, the program determines how the sensor should be directed (for example, by pointing or tuning). Appropriate control commands are then sent to the simulated sensors. C. Interpretation (Hypothesis Validation; Model Instantiation) In this phase, the program attempts to interpret sensor data in the context of threats that were anticipated earlier. It first tries to 254 determine whether sensor data are consistent with specifically anticipated threats, then with general weapon types expected in the area. Since sensor data are inherently ambiguous (particularly if environmental conditions are suboptimal), this step attempts to determine the most likely interpretation. Inference techniques used for interpretation include production rule procedures, probabilistic computations, and geometric reasoning. Production rules are used to infer probable weapon operation (e-g., target tracking, missile guidance), on the basis of such information as past status, environmental conditions, and distance from the aircraft. Probabilistic updating of identification likelihoods is based on the consistency of actual sensor data with expected data, and on agreement (or disagreement) among sensors with overlapping coverage. Geometric reasoning introduces a concept of global consistency to improve identification by comparing inferred identifications and locations of threat system components with geometric models of typical, known system deployments. The interpretation phase brings a great deal of a priori knowledge to bear on the problem of determining the most likely threats the sensors are responding to. This results in much better identifications than those produced by the sensors alone. Confident identifications are entered into the OB and the entire process is continued. D. Performance An experimental test of the system, using a simulated threat environment, allowed a comparison between two modes of operation--an "undirected" mode and one based on perceptual reasoning. A scoring technique that measured the effectiveness with which the system detected, identified, and located hostile systems in a timely fashion was used to grade performance. The ability of the perceptual reasoning system to use external knowledge sources effectively, and to integrate information from multiple sensors, produced superior capabilities under this measure. These capabilities showed themselves even more prominently in situations where environmental conditions tended to degrade sensor performance, rendering it critical that attention be focused sharply. III DISCUSSION Our approach to perceptual reasoning suggests_ that the problem of perception actually involves the solution of a variety of distinct types of subproblems, rather than repeated instances of the same general problem. The system we described utilizes a nonmonolithic collection of representations and reasoning techniques, tailored to specific subproblems. These techniques include both logical deduction and probabilistic reasoning approaches, as well as procedures capable of geometric reasoning and subjective inference. We have discussed several key aspects of the general problem of perceptual reasoning, including the assertion that perception is goal oriented, and inductive and interpretative rather than deductive and descriptive; that because complete modeling of the physical world is not practical, nexperimentationn is a critical aspect of perception; and finally, that multiple representations and corresponding reasoning techniques, rather than a single monolithic approach, are required. The specific system discussed above constitutes an attempt to address the reasoning requirements of perception in a systematic way and, to our knowledge, represents one of the few attempts to do so. While systems that truly interact with the physical world in an intelligent manner will certainly assume a variety of forms, we believe they will all ultimately have to resolve those aspects of the problem that have been described here. REFERENCES 1 . 2. 3. 4. 6. M. A. Fischler and R. A. Elschlager, "The Representation and Matching of Pictorial Structures," IEEE Transactions on Computers, Vol. C-22, No. 1, pp. 67-92 (January 1973) H. G. Barrow and J. M. Tenenbaum, "MSYS: A System for Reasoning About Scenes," Artificial Intelligence Center Technical Note No. 121, SRI International, Menlo Park, California (1976) T. D. Garvey, "Perceptual Strategies for Purposive Vision," Artificial Intelligence Center Technical Note No. 117, SRI International, Menlo Park, California (1976) A. Rosenfeld, "Iterative Methods in Image Analysis," Pattern Recognition, Vol. 10, No. 4, pp- 181-187 (1978) R. C. Bolles et al., "The SRI Road Expert: Image-to-Database Correspondence," Proc. DARPA Image Understanding Workshop, Scienr- Applications, Inc., Report No. SAI-79-814-WA, pp. 163-174 (November 1978) T. D. Garvey and M. A. Fischler, "Machine- Intelligence-Based Multisensor ESM System," ' Technical Report AFAL-TR-79-1162, Air Force Avionics Laboratory, Wright-Patterson AFB, OH 45433 (Final Report DARPA Project 3383) (October 1979) 255

1980

26

19

OVERVIEW OF AN EXAMPLE GENERATION SYSTE?l Edwina L. Rissland Elliot M. Soloway Department of Computer and Information Science University of Massachusetts Amherst, MA 01003 ABSTRACT This paper addresses the process of generating examples which meet specified criteria: we call this activity CONSTRAINED EXAMPLE GENERATION (CEG). We present the motivation for and architecture of an existing example generator which solves CEG problems in several domains of mathematics and computer science, e.g., the generation of LISP test data, simple recursive programs, and piecewise linear functions. II THE CEG MODEL --- From protocol analyses of experts and novices working CEG problems in mathematics and computer science, we developed the following description of the CEG task [lOI. When an example is called for, one can search through one's storehouse of known examples for an example that can be JUDGEd to satisfy the desiderata. If a satisfactory match is found, then the problem has been solved through SEARCH and RETRIEVAL. If a match is not found, one can MODIFY an existing example, judged to be close or having the potential for fulfilling the desiderata. If generation through modification fails, one often switches to another mode in which one CONSTRUCTS an example by instantiating certain models and principals or combining more elementary exemplars. The CEti model we use thus consists of processes that RETRIEVE, JUDGE, MODIFY and CONSTRUCT examples. III THE CEG -- SYSTEM The CEG system described here is written in LISP and runs on a VAX 11/780. In addition to solving CEG problems concerning data and simple programs in the LISP domain [I1 1, it is being used to solve CEG problems in a number of other domains 1121: the generation of descriptions of scenes for use in conjunction with the VISIONS scene interpretation system II21; the generation of action sequences in games; and the generation of piecewise linear functions. The flow of control in the Exanple Generator is directed by an EXECUTIVE process. In addition, there is: (1) the RETRIEVER which searches and retrieves examples from a data base of examples; (2) the MODIFIER which applies modification techniques to an example; (3) the CONS'ER which instantiates general "model" examples, such as code 'lte.mplates" (C91, 1161)~ (4) the JUDGE which determines how well an example satisfies the problem desiderata; (5) the AGENDA-KEEPER which maintains an agenda of examples to be modified, based for instance on the degree to which they meet the desiderata or possess flepistemological" attributes (C81, 191). The components use a common knowledge base consisting of two parts: a "permanent" knowledge base which has an Examples-space 191 containing known examples, and a temporary knowledge base which -contains information gathered in the solution of a specific CEG problem. In the Examples-space, an example is represented by a frame, which contains information describing various attributes of that example, e.g., epistemological class, worth-rating. Examples are linked together by the relation of nconstructional derivation," i.e., Example 1 ---> Example2 means that Exanplel is used in the construction of Example2. The temporary knowledge 256 base contains such,information as evaluation data generated by the JUDGE and proposed candidate examples created from known examples by the MODIFIER. IV AN EXAMPLE OF A CEG PROBLEM ----VP In the context of examples of LISP data elements, an example of a simple CEG problem would be the followirg: Give an example of a LISP list of length 3 with the depth of its first atom equal to 2. (Examples teaching.) such as this are needed when debugging and Suppose the permanent knowledge base only contains the lists RETRIEVAL phas;hasf;ils), the system enters the MODIFICATION (A B C), (0 11, (A), ( ); Since the list (A B C) the first two lists are satisfies two of the'three constraints, and thus has standard ttreferencelt examples, the third, a "start-up" example, and the fourth, a known "counter example," i.e., an a higher "constraint-satisfaction-count" example which often than the is handled other examples, it is incorrectly by programs. Since the knowledge base does not contain an example placed which completely satisfies the desiderata (i.e., the as the top-ranking candidate for MODIFICATION by the AGENDA-KEEPER and modifications are tried on it first. existing example is modified to create a new example that is no longer deficient with respect to the constraints. For example, if there were more than one unsatisfied constraint, one could attempt to rectify a candidate example along each of the unsatisfied dimensions. Consider the following problem: Give an example of a list of O’s and l's, which is longer than 2, and which has at least one element deeper than depth 1. One could modify the list (0 1) to meet the unsatisfied second and third constraints by adding at least one more element, say another 1 to give the list (0 1 11, and then modify this new list by adding parens around 0r.e of the elements. Alternatively, one could add parens and then fix the length, or even do both 'at once' by appending on an element such as (1) to the original list (0 1). In this example, there are many ways to modify the original list to meet the constraints, and the order of modification does not much matter. One can 'ldivide-and-conquer'l the work to be done in a GPS-like manner of difference-assessment followed by However, difference-reduction in other cases the order of difference reduction since matters greatly. For instance, consider the constraints are independent [51. the problem: Give an example of a list of length 5 with an embedded sublist of length 2. The candidate example, (A B C) is analysed and found to be lacking in one respect, namely, the depth of its first atom, A, must be made deeper by 1. The sys tern accomplishes this by adding parentheses around the atom A to create the new example ( (A) B C). This list meets all the constraints specified; as it is a solution to the problem, it is entered into the Examples-space as a new example constructionally derived from (A B C) . Thus * if the following problem is asked of the system, Give an example of a list of length 3, the depth of the first atom is 2, and the depth of the last is 3. the system can use the newly constructed its attempt to satisfy the new problem. example in Suppose one is working with the list (A B C>. If one first rectifies the length by adding two more elements, such as 1 and 2, to create the list (A B C 1 2 ), and then modifies this list to have the needed embedded list by “making-a-group” of length 2 around any of the first four elements, say to arrive atthelist(ABC(12) ),one has also modified the length as a side-effect of the grouping modification, i.e., one has messed up the length constraint. In other circumstances, it possible to set up totally contradictory constraints where satisfaction of one precludes satisfaction of the other. Thus a purely GPS approach is not sufficient to handle the complexity of constraint interaction. We are currently investigating the constraint interaction problem as well as issues concerning maintenance of agendas. For example, we are looking to employ planning-type control mechanisms for dealing with the constraint interaction problem. cc 131, C31) VI THE LARGER PICTURE --__I_ V THE HANDLING OF CONSTRAINTS - An application of our CEG system is in an The hand 1 ing of constraints is especially intelligent computer-assisted instruction tutoring important in the MODIFICATION phase where an environment . We are currently building a tutor to 257 teach students about programming languages such as LISP and PASCAL (C143, C151). In this context, CEC will serve the tutor in two ways: (1) it will generate examples to the specifications set by the tutor; and (2) it will evaluate student generated examples for the tutor. The same JUDGE used by CEG can be used to evaluate a student's example and help track down his misconceptions and bugs through analyses of differences between what the tutor requested and what the student acutually generated. In the future, we also plan to incorporate adaptation into the system. For example, the system can keep track of the performance of the various example ordering functions and choose the one that has the best performance. Also, we plan to apply hill-climbing techniques to the modifying processes themselves. That is, since there are alternative ways to massage and modify an example, those routines which lead to the most succeses should eventually be selected first. Adaptation on the modification techniques will be particularly important if the system is to be able to improve its performance, and thus "learn" from its own experience. The current implementation is only a “first-pass” and does not capture the richness of the CEG model. Nonetheless, we feel that it has demonstrated the utility of this model and we feel that subsequent implementations incorporating the characteristics of additional task domains should provide us with a rich environment to continue our investigation into the important process of example generation. Lll [21 c31 c43 c51 REFERENCES Bledsoe, W. (1977) A Maximal Method for Set Variables in Automatic Theorem Proving, Univ. of Texas at Austin, Math Dept. Memo ATP-33. Hanson, A. and Riseman E. (1978) VISIONS: A Computer System for Interpreting Scenes, in Computer Vision Systems, Hanson and Riseman, Eds. t Academic Press, New York. Hayes-Roth, B., ar,d F. Hayes-Roth (1978) Cognitive Process in Planning, Rand Report R-2366-ONR, The Rand Corporation, CA. Lakatos, I. ( 1963) Proofs and Refutations, - - British Journal for the Philosophy of Science, Vol. 19, May 1963. Also published by Cambridge University Press, London, 1976. Newell, A., Shaw, J., and Simon, H. (1959) Report on a General Problem-Solving Program. Proc. of the International Conference on Information Processing. UNESCO House, Paris. 163 Polya, G . (1973) now To Solve St, Second Edition, Princeton UXJ. Press, N.J, E7l Polya, G. (1968) Mathematics and Plausible Reasoning, Volunes I and II, S&?&d Edition, Princeton Univ. Press, N.J. C83 Rissland (Michener) , E. (1978a) Understanding Understanding Mathematics, Cognitive Science, Vol. 2, No. 4. [91 Rissland (Michener), E. (1978b) The Structure of Mathematical Knowledge, Technical Report No. 472, M.1.T Artificial Intelligence Lab, Cambridge. Cl01 Rissland, E. (1979) Protocols of Ex mple Generation 1 internal report, M.I.T., Cambr idge. [ill Rissland, E. and E. Soloway (1980) Generating Examples in LISP: Data and Programs, COINS Technical Report 80-07, Univ. of Mass, (submitted for publication). Cl21 Rissland, E., Soloway, E. , O’Connor, S., Waisbrot, S., Wall, R. , Wesley, L., and T. Weymouth (1980) Examples of Exaple Generation using the CEG Architecture. COINS Technical Report, in preparation. cl31 Sacerdoti, E. (1975) The Nonlinear Nature of Plans, Proc. 4th. Int. Joint Conf. Artificial Intelligence, Tbilisi, USSR. Cl41 Soloway. E. ( 1980) The Development and Ev aluafiion of Instructional Strategies for an Intelligent Computer-Assisted Instruction System, COINS Technical Report 80-04, Univ. of Mass., Amherst. El51 Soloway, E., and E. Rissland (1980) The Representation ar,d Organization of a Knowledge Base About LISP Programming for an ICAI System, COINS Technical Report 80-08, Univ of Mass., in preparation. [161 Soloway, E., and Woolf, B. ( 1980) Problems, Plans and Programs, Proc. of the ACM Eleventh SIGCSE Technical Symposium, Kansas City. 258

1980

27

20

STRUCTURE COMPARISON AND SEMANTIC INTERPRETATION OF DIFFERENCES* Wellington Yu Chiu USC Information Sciences Institute 4676 Admiralty Way Marina de1 Roy, California 90291 ABSTRACT Frequently situations are encountered where the ability to differentiate between objects is necessary. The typical situation is one in which one is in a current state and wishes to achieve a goal state. Abstractly, the problem we shall address is that of comparing two data structures between the two structures. work well in determining all of addressing the semantic syntactic issues. and determining all Current comparison differences techniques differences but fall short We address this gap by applying Al techniques that yield semantic interpretations of differences. I INTRODUCTION One frequently encounters situations where the ability to differentiate between objects is necessary. The typical situation is one in which one is in a current state and wishes to achieve a goal state. Such situations are encountered under several guises in our transformation based programming system research [I, 2, 31. A simple case is one in which we are given two program states and need to discover the changes from one to the other. Another case is one in which a transformation we wish to apply to effect a desired change does not match in the current state and one wishes to identify the differences. An extension of this second case is the situation where a sequence of transformations, called a development, is to be applied to (replayed on) a slightly different problem. Abstractly, the problem we shall address is that of comparing two data structures and determining all differences between the two structures. Current comparison techniques determining all syntactic differences but fall short the semantic issues. situations, comparisons address this gap by applying interpretations of differences. work well in of addressing In the replay and state differencing must be more semantically oriented. We This paper describes a semantic based differencer Al techniques that yield semantic part of my thesis: the design and its ongoing implementation. of a * This work woa rupportrd by Nitional Scknce the viowr cwpremed are thaw of thr author. Foundation Grant MCS 7683880. II AN EXAMPLE -- The following example is presented to show the types of information used to infer and deduce the semantic differences. Below are the before and after program states from a transformational development [ 1 J. BEFORE : uhile there exists character in text do 1 if character is linefeed 2 then replace character in text by space:3 uhile there exists character in text do i f Pfcharacter) : then remove character from text: 6 AFTER: while there exists character in text do begin ;: if character is Iinefeed then replace character in text by space: : if character in text then e i f P (character) f then remove character from text end: h9 The after state above was produced from the before state via application of a transformation but the following explanations of differences were generated without knowledge of the transformation. - The current syntactic differencing techniques [4, 5, 6, 7, 8, 9, lo] typically explain differences in the following terms: For BEFORE: Delete while in line 4 Delete there in line 4 For AFT&I:’ ’ Insert if in line f Insert then in line f . * . - A higher level syntactic explanation is achieved by generalizing and combining the techniques for syntactic differencing to explain differences in terms of embeds, extracts, swaps and associations, in addition to inserts and deletes, and by incorporating syntactic information about the structures being compared [3] The second loop is coerced into a conditional. 259 The conditional is embedded into the first loop, trees for 23, 5,6 and c,d,f,g are loop generators. - The proposed explanation of the semantic difference is: Loops merged. The following is the derivation of the syntactic explanation. It - Infer the similarity of 5,6 to e,f,g from the syntactic equivalence of 5,6 to f,g. 5,6 embedded in a test for generator consistency inferred from semantic knowledge of loop merging. - Conclusion: Loops merged. is presented to show the mechanisms upon which the semantic differencer will be based. - Syntactically, 2,3 (first loop body of the before state) is equivalent to c,d (part of the loop body of the after state) and 5,6 is equivalent to f,g. - Infer composite structure 1,2,3 similar to composite structure a,b,c,d,e,f,g,h based on 2,3 being equivalent to c,d. - tnfer an embed of composite structure 4,5,6 into the composite structure 1,2,3 to produce a,b,c,d,e,f,g. The support for this inference comes from 5,6 being equivalent to f,g, the adjacency of 1,2,3 to 4,5,6, and the adjacency of c,d to f,g. - Infer coercion of loop 4,5,6 to conditional e,f,g based on 5,6 being equivalent to f,g, and the similarity of the loop generator to the conditional predicate. - Conclude second loop embedded in the first loop. Our current syntactic differencer produces this type of difference explanation. It exte_nds the techniques currently The explanation generated by our syntactic clifferencer is not plausible because i? doesn’t make sense to transform a loop into a conditional only to embed this new conditional into a similar loop. The following is the desired explanation: The body of the decond loop is embedded in a conditional that tests for loop generator consistency. This is done without changing the functionality of the body. The two adjacent loops can now be merged subject to any side effects, caused by the first loop body, that will not be caught by the loop generator consistency check around the second loop body. Ill DESIGN OF THE SEMANTIC BASED DIFFERENCER -e--p We start by defining relations (profiles) on objects, where objects are the substructures of the structures we are comparing (see Appendix A). The information provided by this profile consists of: - Sequence of nonterminals from the left hand side of productions used in generating the context tree of the substructure. A context tree (i.e. super tree) is that part of the parse tree with the substructure part deleted. available by imposing structure on the text strings being advance, the explanations fall short of the desired semantics of compared, thereby making use of structural constraints and the the changes. In the SCENARIO section below, an explanation extra structural dimensions in its analysis [3]. Despite this - The sequence of productions used in the derivation of the substructure, given the context tree above as the starting point. Positional information. from the proposed semantic based differencer is presented. The domain we are dealing with is one where changes are made for the following reasons: 1. To optimize code Our syntactic differencer makes use of the information provided by the object profile to determine base matches. The techniques used in the differencer to determine base matches are combinations of the common unique, common residual and common super tree techniques described in [3, 4, 6, 7, lo]. 2. TO prepare code for optimization. Below is a brief description of the common unique and common residual techniques for linear structures. Within such a domain, we can use the constraints on the semantics of changes to derive the semantic explanation “loop merged” and at the same time rule out the syntactic explanation. - Longest Common Subsequence (LCS): The main coniern with LCS is that .of finding the minimal edit distance between two linear structures. The only edit ooerations allowed are: delete a substructure, Building on the mechanisms that generated the derivation above, or insert a substructure [6, 81. For the above the following is a proposed derivation of the comparison that example, the LCS is: 1,2,3,5,6 matching a,c,d,f,g. yield the semantic interpretation “loops merged”. - Common Unique KU): The key to this technique the - Syntactically, 2,3 (the first loop body) is equivalent to c,d and 5,6 (the second loop body) is equivalent to f,g. The context trees (i.e. super trees) containing 2,3 and 5,6 and the context tree for c,d,f,g are the same. Infer FACTORING of context trees with supports being the 2 to 1 mapping of substructures and the equivalence of context trees. use of substructures that are common to both structures and unique within each as anchors [4]. For the above example, the common unique substructures are: linefeed, replace, space, P, remove. We then build on these base matches by inferring matches of - Infer loop merge from the fact that the context substructures that contain substructures that are matches. An example of this is inferring that 1,2,3 is similar to a,b,c,d,e,f,g,h 260 from the assertion that 2,3 matches c,d. There are two types Of inferred matches: those without syntactic boundary conflicts and those with conflicts. Syntactic boundary conflicts result from embedding, extracting or associating substructures. The third type of profile is one that describes the relationship between substructures within a given structure. Considerations here are: adjacency, containment, and relative positioning. There are several semantic rules that describe a majority of structure changes. Some are: factoring, distributing, commuting, associating, extracting, embedding, folding, and unfolding. A oartial description of the factor semantics can be found in Appendix A. Factors currently considered by our semantic rules are: support for matches, the generating grammar, object profiles, and relations between substructures of the same structure. With each set of examples tried, we add to our set of semantic rules, and our intuitive guess, given our domain of changes due to optimization or preparation for optimization, is that this set will be fairly small when compared to the set of transformations needed in a transformation based programming system. IV A SCENARIO -- Our syntactic differencer makes use of structural information. For LISP programs it know about S-expressions. For programs written in our specification language, differencing is performed on the parse trees. The differencer first tries to isolate differences into the smallest composite substructure containing all changes. With this reduced scope, the differencer uses the common unique technique to assert relations on content base matches. In our example, substructure 2,3 is equivalent to c,d and substructure 5,6 is equivalent to f,g. Once all of these possible assertions have been made, we use them as anchors to assert relations based on positional constraint and content matches (residual structure matches). This residual technique works well as a relaxation of the uniqueness condition in the common unique requirement, and acts as a backup in case no substructures are common to both structures and unique within each. The super tree technique is used as a default when neither of the above techniques applies. The intuitive explanation for this third technique has to do with both objects occupying the same position with respect to a common super tree. With the super tree technique, content equivalence is relaxed. With the two asserted relations regarding substructures 2,3 being equivalent to c,d and 5,6 being equivalent to f,g, we now infer that 4,5,6 is similar to e,f,g because 5,6 is common unique with f,g and without conflicting evidence (i.e. boundary violations) the assertion is made. Once made, further analysis of this reduced scope shows relationships between the loop generator 4 and the conditional predicate e. Since we are given, via the super tree technique, that 1,2,3,4,5,6 matches a,b,c,d,e,f,g,h we assert the inference 2,3 to c,d even though conflicts due to boundary violations arise. The boundary violation in this instance is the mapping of two substructures into one (i.e. the segmental or n-m problem). Given that we want to produce all plausible explanations, we assert that 1,2,3 is similar to a,b,c,d,e,f,g,h because 2,3 and c,d are common unique matches. With this assertion, we could with our Embed Semantics say that the second loop is embedded in the first. But our knowledge about optimizations makes more plausible the Factor Semantics that is triggered by the segmental matching. When the Factor Semantics is triggered, the relationships within a given structure, such as adjacency and relative positioning, are asserted (see Appendix A). All the requirements except for body2 being equivalent to body4 are met. But once the cases are considered, we discover that the operator being factored is indeed a loop generator and that we can relax the requirement that body2 be equivalent to body4 to that of similarity of the two bodies. This follows from the support for Final analysis the relationship between body2 and body4. reveals that body2 is embedded in body4. V CONCLUSION Given the small example above, we see that the derivation of the semantic difference involves syntactic and semantic knowledge of the structure domain as well as techniques for managing and applying the knowledge. We present a design that addresses the issue of managing and applying both syntactic and semantic knowledge of the structures being compared so as to provide a semantic interpretation of changes. This allows us to bridge the gap that exists between the information provided by current differencers and the information needed by current differencing tasks. VI APPENDIX A: -- TEMPLATES FROM THE SEMANTIC DIFFERENCER ----I__ Every substructure of the structures being compared has associated with it an Object Profile that is an record with the following role names as fields: Content: Value of the substructure. Type Context: Sequence of nonterminals of productions, of the grammar, used in the derivation of the substructure. Posi t ional Position in parse tree. (i.e. a Context: sequence of directions in reaching the substructure from the root of the parse tree. Abetraction: If a grammar is used to generate the substructure, this refers to the sequence of productions used to generate the substructure itself. 261 A Relation Profile describes the relationships between two substructures, one from each of the structures being compared. The role names of this record are: Base Content (i.e. common unique) Matches: Context (i.e. positional determined from context trees) Positional constraint and Content (i.e. from the largest common residual substructure technique where uniqueness of context matches is relaxed). Inferred Conflict free (i.e. no syntactic Matches: boundary violations) With conflicts (inferences depends on heuristics regarding current substructure abstraction and weights associated with substructure matches). Happ i ngs: l-l substructure matches 2-1 substructure matches n-m substructure matches A second relation profile is one between substructures within a given structure. Some of the considerations here are: adjacency of two substructures, containment, and relative positioning. There are several semantic rules that describe a majority of structure changes. Some are factoring, distributing, commuting, associating, extracting, embedding, folding and unfolding. Below is a partial description of the Factor Semantics used in generating the semantic interpretation above: FACTOR SEllANTICS: FORM: LHS: opl body1 op2 body2 RI-IS: op3 body3 body4 KEY: Segmental matching REDUIREMENTS: requirements opl=opZ requirements opl =op3 requirements bodyl=body3 requirements bodyZ=body4 requirement9 relative positioning of body1 to body2 holds for body3 to body4 requirements adjacency of body1 to body2 holds for body3 to body4 CASES : opl is a loop generator relaxations body2=body4 relaxed to body2 similar to body4 RELAXATIONS: relaxations adjacency requirement relaxations relative positioning relaxations equivalence (=I to similar VII REFERENCES 1. Baiter, R., Goldman, N., and Wile, D., “On the Transformational Implementation Approach to Programming,” in Second International Conference on Software Engineering, pp. 337-344, IEEE, October 1976. 2. Balzer, R., Transformational Implementation: An Example, Information Sciences Institute, Research Report 79-79, September 1979. 3. Chiu, W., Structure Comparison, 1979. Presented at the Second International Transformational Implementation Workshop in Boston, September, 1979. 4. Heckel, P., “A Technique for Isolating Differences Between Files,” Communications of the ACM 21, (41, April 1978, 264-268. 5. Hirschberg, D., The Longest Common Subsequence Problem, Ph.D. thesis, Princeton University, August 1975. 6. Hirschberg, D., “Algorithms for the Longest Common Subsequence Problem,” journal of the ACM 24, (41, October 1977, 664-675. 7. Hunt, J., An Algorithm for Differential file Comparison, Bell Laboratories, Computer Science Technical Report 41, 1976. 8. Hunt, J., “A Fast Algorithm for Computing Longest Common Subsequences,” Communications of the ACM 20, (51, May 1977,350-353. 9. Tai, K., Syntactic Error Correction in Programming Languages, Ph.D. thesis, Cornell University, January 1977. 10. Tai, K., “The Tree-to-Tree Correction Problem,” Journal of the ACM 26, (3), July 1979, 422-433. 262

1980

28

21

Performing Inferences over Recursive Data Bases Shamim A. Naqvi and Lawrence J. Henschen Dept. of Electrical Engineering and Computer Science Northwestern University Evanston, Illinois 60201 Abstract The research reported in this paper presents a solution to an open problem which arises in sys- tems that use recursive production rules to re- present knowledge. The problem can be stated as follows: "Given a recursive definition, how can we derive an equivalent non-recursive program with well-defined termination conditions". Our solution uses connection graphs to first detect occurrences of recursive definitions and then synthesizes a non-recursive program from such a definition. I. Introduction In recent years, attention has focused on adding inferential capability to Codd's relational model of data (Codd 1970). This usually takes the form of defining new relations in terms of existing relations in the data base. The defined relations constituting the Intensional Data Base describe general rules about the data whereas explicit facts stored in the data base as base relations comprise the Extensional Data Base. This paper is concerned with the problem of find- ing a finite inference mechanism for a defined relation. Reiter (1977) suggests that for non-recursive data bases the essentially logical operations in- volved in unifying and resolving intensional lit- erals can be taken care of, i.e. "compiled", be- fore receiving queries, leaving only those opera- tions specifically related to information retrie- val from the extensional data base. We propose to extend this idea to the general case by analyzing what resolutions are possible that can lead to answers for a particular kind of query. In the case of recursive axioms this in- volves finding a pattern of data base retrievals instead of just a single data retrieval as in Reiter (1977). II. Problem Representation We shall view a data base as the ground unit clauses of a first order theory without function signs. The words literal and relation will be used interchangeably and all variables are as- sumed to be universally quantified. We propose to solve the problem of recursive def- initions by using connection graphs like those of Sickel (1976) in which nodes represent intensional axioms and edges connect unifiable literals of opposite signs. A loop is a Potential Recursive Loop (PRL) if the substitutions around the loop are such that the two literals at both ends of the loop are not the same literal and are not in the same instance of the clause partition. Figure 1 shows an example PRL in which E and F are base relations and letters at the end of the alphabet denote variables whereas letters at the start of the alphabet denote constants. In this case, starting from A(a,x,z,p) and resolv- ing around the loop (separating variables as we go) we eventually come back to clause 1 yielding an ultimate resolvent -El(x,y) A(a,y,z,b) B(y,y') lEl(y,y') in which the literal at the end of the cycle, A(a,y,z,b), is a different literal than the one we started the loop with. Two features of this loop traversal are noteworthy. First, the literal E causes data base accesses which provide possibly new values for y. Second, these values of y instantiate x for the next traversal around the loop and also cause data base accesses for F which provides answers to the query. 4 Flgure 1. Example PRL 263 III. Derivation of the Iterative Program Since non-atomic queries can be decomposed into equivalent atomic queries we shall only consider atomic queries in this paper. Before describing our method of deriving an iterative program for a recursive definition we notice that two kinds of edges exist in a PRL. A cycle edge contributes to the PRL whereas an exit edge provides an exit from the PRL. Extensional literals reached by traversing exit edges are called Exit Extensional Literals and those reached by traversing cycle edges are called Cycle Extensional Literals. For example in Figure 1 edges 2, 3 and 4 are cycle edges and edges 5 and 6 are exit edges; 1El(x,y) is a Cycle Extensional Literal whereas 1E3(u,v) and 1E4(q,r) are Exit Extensional Literals. We make the following observations about a PRL. Observation 1 A PRL must have an exit edge, which corresponds to the presence of a basis case for a recursive definition, in order for its clauses to contribute an answer to a query. In figure l the basis case is A(a,q,r,b) 1F(q,r). Notice that a literal having an exit edge has a non-exit edge which contributes to the cycle. Observation 2 In Horn data bases, if a PRL exists for a literal Q, then a literal- must exist which provides the closing edge for the PRL. We represent the defined relations as a connection graph and in a preprocessing step identify all PRLs. A set of answer expressions corresponding to a PRL is derived as follows: We note that the exit edges of Observation 1 above must be connect- ed to cycle literals. Starting from the intension- al axiom from which we expect to get a query, we first delete the literal which would resolve against the query. We then resolve around the cycle until we come to an exit edge. At this point the exit literal represents an expression which can be con- sidered as a call to a procedure. This procedure provides one way of obtaining some of the answers. Paving derived the expression for the first exit we proceed to successive exits in the same manner. These expressions are called answer expressions. In Figure 1 the answer expressions are -tEl(x,y) OR 1E3(y,z) and 1El(x,y) OR -rE4(y,z). A loop residue is obtained by resolving around the loop, starting from the intensional axiom from which we expect to get a query, and traversing only the cycle edges of the PRL. The ultimate resolvent is of the form E:= 1(El(argl,..) & . . . SC Ei(argl,..)) where the Ei (ihO) are base or defined relations. This expression is called the loop residue. In Figure 1 the loop residue is 1El(x,y). In order to derive a program from a PRL we use an algorithm given in Naqvi (1980). In this section we shall illustrate the working of this algorithm by considering two similar definitions of the an- cestor relation. Consider the first definition given below and the corresponding connection graph is shown in Figure 2. 7. 8. -'ANCESTOR(x,y) lFATHER(y,z) ANCESTOR(x,z) lMOTHER(u,v) ANCESTOR(u,v) It is straight-forward to show that with the +NCESTOR(w,a) we can generate the resolvent query 9. tiNCESTOR(w,y') lFATHER(y',y")....lFATHER(y,a) which corresponds to a left recursive definition of the ancestor relation. In this case the basis statement is used in the end to get the expression -+iOTHER(w,y') 1FATHER(y',y")...lFATHER(y,a). The data retrieval pattern is to find successive fathers of 'a' and then find a mother. In terms of the connection graph this corresponds to tra- versing the loop a certain number of times and then taking the exit edge. Examining the PRL we find that z, which is the variable that is expect- ed to be the driver, is replaced in the loop by y. Moreover, z determines y through the extensional evaluation of Father(y,z). This determination occurs within the loop without recourse to the basis statement. Our algorithm does the above kind of analysis and uses the answer expression derived from the loop and the substitutions from the closing edge of the loop to derive a program for the PRL. For this example it derives the following program fragment: 2: =a ENQUE(q,z) /* q is a queue */ while (q 1= empty) do z:= DEQUE(q) x:= 1(MOTHER(x,y) & FATHER(y,z)) ENQUE (4, Y 1 od Now, consider the second definition of the ancestor relation given below and the corresponding connec- tion graph shown in Figure 3. 11. lFATHER(x,y) -'ANCESTOR(y,z) ANCESTOR(x,z) 12. -+lOTHER(u,v) ANCESTOR(u,v) Once again, the query -&NCESTOR(w,a) and 11 can be shown to generate the resolvent 13. 1FATHER(x,y) 1FATHER(y,y)...7ANCESTOR(y",a) In this case, our first answers come from resolv- ing (12) and (13) which corresponds in figure 3 to taking the exit edge to the basis case. Sub- sequent answers are derived by finding the succes- sive fathers which corresponds to going around the loop a certain number of times. Examining the PRL we find that the values of y derive the next set of answers, x. The expected driver vari- able z does not participate in this process. Our algorithm uses the resolvent of the basis case and the query to start the loop. The substitutions at the closing edge of the PRL identify the cor- rect variables which drive the loop and serve as place holders for answers. The loop residue de- rives all the subsequent answers. The program is as shown below. .- x.-w t'=lOTHER(w,a) ENQm (q ,w> while (q 1 = empty) do y:= DEQUE(q) 1 FATHER(x,y) ENQUE (4, x> od YlZ i ANCESTOR(u,v) LMOTHER(J Figure 2. First Definition of Ancestor Relation -rFATHER(x,y) -,ANCESTOR(y,z) ANCESTOR(x,z) (y’u*z’y &?j z) (w/u,a/v) Figure 3. Second Definition of Ancestor Relation To review then, our algorithm analyzes the PRL oE a recursive definition to determine the loop resi- due, the answer expressions, the resolvent of the query and the basis case and whether the defini- tion is left or right recursive. It then derives a program whose structure corresponds to one of the two program structures outlined above. It now remains to discuss the termination condi- tions of the derived programs. Our termination conditions are designed for data bases without function signs. Briefly, we use a queue to store all unique values of the variables, indicated by the loop analysis, during each iteration. Each new iteration dequeues a previously enqueued value to generate some possibly new answers. Since the domain of discourse is finite in the absence of function signs the number of unique answers in the data base is finite. Thus the program queue will ultimately become empty. It should be noted that our technique for the detection of and generating programs for recursive definitions works in the presence of function signs. However, the termina- tion condition does not guarantee finite computa- tions in this case. v. Summary and Conclusions We have outlined an algorithm which derives itera- tive programs for recursively defined relations. The case where a defined relation is mutually recursive with some other definition (e.g. X & R -> R and R & Y -b X) leads to the derivation of mutually recursive programs. Transitive recursive axioms (e.g. ancestor of an ancestor is an ancestor) lead to the derivation of recursive programs. These situations require a fairly complicated con- trol mechanism for execution time invocation of the derived programs. This is discussed in detail in Naqvi (1980) and the algorithJz for deriving the programs is also given there. We can show the finiteness and completeness of our method (Naqvi 1980). Although we have considered a first order theory without function signs the method is ap- plicable to data bases containing function signs. The termination condition, however, may not be rigorous in this case. This is an obvious area for further research. References Chang, C. L., (1979) "On Evaluation of Queries Containing Derived Relations in a Relational Data Base", Workshop on Formal Bases for Data Bases, Toulouse, France. Codd, E.F., (1970) "A Relational Model of Data for Large Shared Data Banks", CACM 13, 6, 377-387. Naqvi, S. (1980) "Deductive Question-Answering in Recursive Rule-Based Systems", Ph.D. Diss., Northwestern University, (in preparation). Reiter, R., (1977) "An Approach to Deductive Question Answering", BBN Tech. Report no. 3649. Sickel, S. (1976) "A Search Technique for Clause Interconnectivity Graphs", IEEE Trans. on Comput- ers, Vol. C-25, No. 8. 265

1980

29

22

Automatic Goal-Directed Progrem ‘bvnsformetion Stephen Fickas USC/Information Sciences Institute* Marina del Rey, CA 90291 1. INTRODUCTION This paper focuses on a major problem faced by the user of a semi-automatic, transformation-based program-development system: management of low level details. I will argue that it is feasible to take some of this burden off of the user by automating portions of the development sequence. A prototype system is introduced which employs knowledge of the transformation domain in achieving a given program goal state. It is assumed that such a system will run in a real environment containing a large library of both generalized low level and specialized high level transformations. 2. THE TI APPROACH TO PROGRAM DEVELOPMENT The research discussed here is part of a larger context of program development through Transformational Implementation (or TI) [I, 21. Briefly, the TI approach to programming involves refining and optimizing a program specification written in a high level specification language (Currently, the GIST program specification language [8] is being used for this purpose) to a particular base language leg. LISP). Refinement and optimization are carried out by applying transformations to program fragments. This process is semi-automatic in that a programmer must both choose the transformation to apply and the context in which to apply it; the TI system ensures that the left hand side (LHS) of the transformation is applicable and actually applies the transformation. The TI system provides a facility for reverting to some previous point in the development sequence from which point the programmer can explore various alternative lines of reasoning. 3. CONCEPTUAL TRANSFORMATIONS AND JITTERING TRANSFORMATIONS In using the TI system, programmers generally employ only a small number of high level “conceptual” transformations, ones that produce a large refinement or optimization. Examples are changing a control structure from iterative to recursive, merging a number of loops into one, maintaining a set incrementally, or making non-determinism explicit. Typically these transformations have complex effects on the program; they may even have to interact with the user. Although only a relatively small number of conceptual transformations are employed in a typical TI development, the final sequence is generally quite lengthy. Because the applicability conditions of a conceptual transformation may be 9; This research was supported by Defense Advaked Research Projects Agency contract DAHCIS 72 C 0308 Views and conclusions contained in this document are fhooe of the authors and should not be interpreted as repreeenting the official opinion or policy of DARPA, the U.S. Government, or any other person or agency connected wilh them. quite specialized, usually with a number of properties to prove, much of the development sequence is made up of lower level transformations which massage the program into states where the set of conceptual transformations can be applied. I call these preparatory steps “jittering” steps. Examples of quite low level jittering steps include commuting two adjacent statements or unfolding nested blocks. More difficult jittering steps include moving a statement within a block from the kth position to the first position, merging two statements (eg. conditionals,loops) into one, or making two loop generators equivalent. 4. AN AUTOMATIC JITTERING SYSTEM Requiring the programmer to carry out the Jittering process detracls from his performance in several ways: it consumes a large portion of his time and effort; it disrupts his high level planning by forcing him to attend to a myriad of details. There is then strong motivation for automating some or all of the Jittering process. The following sections will discuss the types of mechanisms used to actually implement such a system thenceforth known as the Jitterer). 4.1. BLACK BOX DESCRIPTION The Jitterer is initially invoked whenever the TI system is unable to match the LHS of a transformation Tk selected by the user. The Jitterer’s inputs are 1) the current program state C, 2) a goal state G corresponding to the mismatched LHS of Tk, and 3) a library of transformations L to use in the jittering process. The Jltterer’s return value is either a failure message or a program state S matching G and a sequence of instantiated transformations from L which when started in C will result in S. If the Jltterer is successful, then Tk will be removed from its suspended state and applied as specified in state S. If G is a conjunction of sub-goals then currently a simple STRIPS like approach is employed in solving each in some determined order. This approach can be both inefficient and unbale to find solwtlons for certain jittering problems. Improving this process, possibly using some of the techniques proposed by problem solvers in other domains (see Sacerdoti [9] for a survey), remains a high priority item for future research. 4.2. THE GOAL LANGUAGE Contained in the TI system is a subsystem called the Differencer [53. The Differencer takes as input a current state pattern and a goal pattern, and returns a list of difference-descriptions between the two (an empty list for an exact match). Each element of the list is a description taken at a particular level of detail, and is written in a goal language I call GLl. For example, 68 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. suppose that the Differencer was passed “if P then FOO” as a current pattern, and “if Q then SA” as a goal pattern (The notation SX stands for a variable pattern matching a single statement). The output of the Differencer would be the following three descriptions in GLl: CHANGE-PREDICATE(P 0); CHANGE-ACTION(“if P then FOO” “if Q then SA”), ie. change from one conditional to another without going out of the current context; PRODUCE-ACTION-AT-POSITION(“if Q then $A” J), ie. use any means necessary, including looking at the surrounding context, to produce a conditional matching “if Q then $A” at position J, J being bound by the Differencer. The above three descriptions form a disjunction of goals, each of which gives a slightly wider perspective. Currently the Jitterer attempts to solve the most narrow goal first. If it is successful, then control returns to TI. If not, it attempts to solve the next higher goal and so on until the list is exhausted. De fault-bindings: none t various tactics 1 We now must define the individual jittering plans or Tactics which will achieve the CHANGE-PREDICATE goal. Each Toceic construct is composed of a set of constraints which further limit its applicability and an Actron for achieving the matching goal. Our three plans become formally STRATEGY change-predicate-from-a-to-b - _ - 1 . . Tactic (1) enlbed Applicabhty-condition: none Action: POST (EMBED (Predl Pred21) Taceic(2) extract Applicability-condition: NOT (VARIABLE (Predl I 1 Action: POST (EXTRACT tPred2 Predl) 1 Other goals of GLl not included in the above example include EMBED, EXTRACT, DELETE, ADD, COMMUTE, DISTRIBUTE, PROVE-PROPERTY, FOLD and UNFOLD. Because GLl is the language used to both describe pattern differences by the Dlfferencer and describe desired goal states by the Jitterer, it acts as a common interface between the two. 5. JiTTERIf’& PLANS A small number of jittering plans are capable of solving many of the goals of GLl. Suppose we take for example the goal of the previous section CHANGE-PREDICATE(P Q). This is a specific instance of the more general goal CHANGE-PREDICATE(pattern1 pattern2). There are three basic plans for solving this more general goal: 1) embed pattern1 in patternil, 2) extract pattern1 from pattern2, or 3) first embed pattern1 in something containing pattern2, and then extract pattern2. Since in our case each pattern is a simple variable, we can rule out the second plan. Similar plans exist for the other two goals of the example. In general, jittering plans can be defined for all but the most detailed goals of GLl. 5.1. STRATEGY AND TACTICS The plans available for solving a particular goal have been organized into a construct I call a STRATEGY. Each of the plans is known as a Tactic. Each STRATEGY construct takes the following form (see [4], [7] for similar planning constructs in other domains.): STRATEGY name Relevant-goal: goal the strategy matches Applicability-condition: used to narrow strategy’s scope Default-bindings: bind any unbound goal parameter 6 Tactic(l) . . . Tactic(n) . . . To illustrate, let us package the three plans described informally for solving the CHANGE-PREDICATE goal into a STRATEGY Tactic (31 embed-and-then-extract Applicabihy-condition: none Action: SEQ (POST (EMBED (Predl (#ANY-PRED fr Pred211) POST (EXTRACT tPred2 (#ANY-PRED Predl Pred21 Here, POST(G) says mark the goal G as a sub-goal to be achieved. *ANY-PRED will match to either AND or OR. SEQUENCE(A1 A2 . . . An) says execute the list of actions scqucntlally; if Ai is of the form POST(G), then do not move on to Ai+l until G has been achieved. There exist similar functions for executing a sequence of actions in parallel and in disjunctive form. In general, an Action can specify an arbitrarily ordered list of actions to take to achieve the STRATEGY’S goal. 5.2. BACKWARD CHAINING The transformations available for jittering, along with the Tactics introduced through the STRATEGY construct, define the methods available for solving a particular goal. Transformations are applied in backward chaining fashion: once a transformation’s RHS matches a goal G, variables are bound and the LHS is instantiated. The Differencer is then called in to see if a match exists between the current state and the instantiated LHS. If so, then G is marked as achieved by the application of the transformation. if there is a mismatch, then the disjunction of sub-goals produced by the Differencer will be marked as awaiting achievement. 6. THE JITTERlNG SCHEDULER Whenever a new goal is posted (marked to be achieved), the Jitterer attaches to it a list of methods (transformations and Tactics) which might solve it. It is at this point that the Scheduler is called in. The Scheduler first must decide among all active goals (ones with at least one untried method) which to construcl. Suppose that we wanted to rule out working on goals work on next. Once a goal IS chosen, it must decide which of the that attempt to change True to False or False to True, and that untried methods to employ. A set of domain dependent metrics we expected both of CHANGE-PREDICATE’s parameters to be which help the Scheduler make an intelligent decision in both bound. We would get the following STRATEGY header: cases has been identified. STRATEGY change-predicate-from-a-to-b Relevant-god: CHANGE-PREDICATE (Predl Pred2) .&n!:cability-conditions: NOT (Predl=True h Pred2=Fa I se) A NOT (Predl-Fa I se A Pred2=True) 6.1. Choosing among competing goals Lcngfh of path: the types of problems presented to the .litterer by Ti do not generally involve a large (over 10) number of transformation steps. Clcnce, as a path (sequence of 69 transformations) grows continuing it decreases. over a fixed threshold, the desirability of Conflict with high level motivation: if the Scheduler is able to determine the user’s current high level goal, it may be able to rule out certain goal states as un-productive. For example, if it is known that the user is trying to optimize his control structure by merging a number of loops into one, it may be unwise to try to achieve a sub-goal which produces still another loop. Such a sub-goal would be given low priority. Ease of achieving: A rough estimate is made of the cost to continue the goal by taking the minimum cost of the untried methods attached to it. This is a rough estimate because there is no easy way to compute exactly the final cost associated with any method, or in fact that any method will lead to a solution. 6.2. Choosing among competing methods Ease of application: a rough static estimate of how difficult the method may be to apply. In the case of a transformation, how complex is the LHS (eg, how many properties must be proven)? In the case of a Tactic, how many sub-goals must be achieved? User assistance: some methods call for the user to supply needed information. The preference is to avoid bothering the user as much as possible. Side effects: What undesirable actions will a method take besides the desired one (a qualitative judgment)? For instance, a method which unfolds a large procedure body to produce a certain type of goal pattern is seen as having a large side effect, ie. it tends to “flatten” the functional structure of the program. On the other hand, a method which simply changes a current statement into something matching a goal pattern has very little side-effect; it does nothing to disturb the surrounding context. Prefer small side effects over large ones. Tactic ordering rules: a STRATEGY writer may provide rules for ordering Tactics. These rules take the form “if Condition then Ordering”, where Condition can refer to any piece of knowledge known to the STRATEGY at match time and Ordering takes the form “Try Tactic(J) before Tactic(K)” or “Try TactdN) last”. The Scheduler makes use of this information when choosing among competing tactics for a particular goal. 7. CONCLUSION The purpose of this paper has been to present a prototype Jitterer with enough domain knowledge to deal competently with the types of jittering problems typically encountered in a TI environment. The prototype Jitterer described is currently being implemented in the Hearsay-ill knowledge representation language [3] and represents a preliminary system. Future systems will deal much more with performance issues (see for example the next section). 8. FURTHER RESEARCH There are generally many ways of achieving a given jittering goal. The metrics of section 5 give some help in ordering them. Even so, in some cases the Jitterer will produce a solution not acceptable to the user. A simple minded approach (and the one currently employed) is to keep presenting solutions until the user ik satisfied. A better approach is to allow the user to specify what he didn’t like about a particular solution and allow lhts information to guide the search for subsequent solutions. in fact, the user may not want to wait until an incorrect solution has been presented, but give jittering guidance when the Jitterer is initially invoked (Feather [6] proposes such a guidance mechanism for a fold/unfold type transformation system). Still another approach may be to “delay” jittering until more high level contextual information can be obtained. Both user guidance and delayed reasoning are being actively studied for inclusion in future jittering systems. Acknowledgments I would like to thank Bob Balzer, Martin Feather, Phil London and Dave Wile for their contributions to this work. Lee Erman and Neil Goldman have provided willing and able assistance to the Hearsay Ill implementation effort. 9, REFERENCES 1. Balzer, R., Goldman, N., and Wile, D. On the Transformational Implementation Approach to programming. Second International Conference on Software Engineering, October, 1976. 2. Balzer, R. TI: An example. Research Report RR-79-79, Information Sciences Institute, 1979. 3. Aalzcr, R., Erman, L., London, P., Williams, C. Hearsay-Ill: A Domain-Independent Framework for Expert Systems. First National Conference on Artificial Intelligence, 1980. 4. Bulrss-Rozas, J. GOAL: A Coal Oriented Command language for Int ractue Proff Constructlorr. Ph.D. Th., Computer Science Dept., Standford University, 1979. 5. Chiu, W. Structure Comparison and Semantic Interpretation of Differences. First National Conference on Artificial Intelligence, , 1980. 6. Feather, M. A System For Developing Programs by Transformatron. Ph.D. Th., Dept. of Artificial Intelligence, University of Edinburgh, 1979. 7. Friedland, P. Knowledge-based Hierarchical Planning in Molecular Genetics. Ph.D. Th., Computer Science Dept., Stanford University, 1979. 8. Goldman, N., and Wile, D. A Data Base specification. International Conference on the Kntity-Relational Approach to Systems Analysis and Design, UCLA, 1979. 9. Sacerdoti, E. Problem Solving Tactics. Technical Note 189, SRI, July 1979. 70

1980

3

23

PIAGET AND ARTIFICIAL INTELLIGENCE Jarrett K. Rosenberg Department of Psychology University of California, Bcrkelcy 94720 ASSTRACT Piaget’s Genetic Epistemology and Artl$cial Intelligence can be of great use to each other, since the strengths of each approach complement the weaknesses of the other. Two ways of bringing the two approaches together are suggested: elaborating the parallels between them (such as the concept of schemata), and building AI models based directly on Piaget’s theory. How this mighl benejit AI is illustrated by examining how Piagerian research on problem solving can suggest new ways of building programs that learn. I. INTRODUCTION. My thesis in this paper is that two such superficially disparate areas as Piaget’s Genetic Epistemology and Artificial Intelligence can inform each other immensely. This fact has already been noticed, as long ago as Seymour Papert’s work at Geneva [l] and as recently as Margaret Boden’s writings on Piaget [2, 31. Nevertheless, I think it still needs to be pointed out how the two fields both parallel and complement each other, and how one might go about trying to connect them. Piaget’s work is immense, stretching over half a century and dozens of volumes, and so only a glimpse of it can be given here, but hopefully it will be enough to stimulate some AI people (and maybe psychologists too). II. WHY SHOULD PIAGET AND RELEVANT? To address the problem of artificial intelligence is to presuppose what natural intelligence is, but AI work in general doesn’t really treat the latter in any comprehensive way; it’s assumed that playing chess is intelligent, or planning, or ordering a hamburger. But can we really hope to understand intelligent behavior, natural or artificial, without deciding what criteria we use for intelligence? If the answer is No. then it seems natural to look to psychology for a theory of intelligence. The only theory I find satisfying, and I suspect most AI people will as well, is Piaget’s. Piaget’s theory is relevant to AI because it is a fully developed, motivated theory of intelligence which is general enough to subsume not only human and animal intelligence, but purely artificial intelligence as well. By motivated I mean that the form of Piaget’s theory of knowledge is based on both the logical requirements of epistemology, and on the biological requirements of behavioral evolution. Hence there are good reasons for accepting his approach to intelligence, just as there are few, if any, compelling reasons to believe the (often implicit) definitions given in AI (or psychology, for that matter). Part of my argument, then, will be that AI rcscarchcrs might be able to use this foundation as a basis For their own work. Piaget, on the other hand, can learn 3 lot from AI, since it has developed ways of looking at the detailed structures and processes of complex behavior, thus suggesting ways of elaborating Piaget’s theory into one with a more “psychological” flavor. Piaget’s theory has always been more epistemologically than psychologically oriented (in the sense of information-processing psychology), and it has only been recently that Genevan researchers have addressed the question of the procedures involved in the acquisition and use cf knowledge- structures. Table 1 gives a general (and biased!) comparison of the strengths and weaknesses of Piaget’s theory and AI work taken as a whole. The important thing to note is that the strengths and weaknesses of the two approaches complement each other nicely. Let me discuss them briefly. The advantages of Piaget’s theory are: b It is both epistemologically and biologically motivated, that is. it starts from the most central issues of a theory of knowledge (e.g., how space, time, and causality are constructed), and places this analysis within the biological context of human evolution. This contrasts with most AI work, which considers intelligence outside of both its philosophical and biological context. b It is highly theoretical and comprehensive. As a biologist with philosophical interests, Piaget resolved to study the “natural history of knowledge”, and so developed a general theoretical framework which he then applied to a number of areas, particularly those of physical and logico-mathematical knowledge. This contrasts with the general lack of interest in AI for large-scale theories or applications. b It has a large amount of empirical validation: virtually everything he has proposed has been extensively replicated and consistently extended. b It proposes a “fundamental unit” for cognition, the scheme, from which all knowledge-structures are constructed. This supplies a uniform analytic framework for all research in cognition. b It has a formalism for the structures of knowledge. This formalism is an algebraic, rather than a computational one, and is only in its beginnings, but is nevertheless valuable. b It is a developnmtal theory of knowled&, promising a way to -- understand how knowledge-structures can radically change and yet prcscrve their integrity. It can thus provide an additional, developmental, set of constraints on AI theories: both structural ones (certain logical properties of cognition, with a logical and devclopmcntal ordering on them), and l%nctional/proccdural ones on the way those structures are constructed and used (e.g., forms of coordination). This should be one of its greatest attractions for AI. The disadvantages of Piaget’s theory are: b It is not very detailed about how knowledge is represented, or how it is processed; to someone looking for a theory at that level, Piagct’s seems mostly descriptive. b It says little about the proc&zrres used in cognition; only recently has this issue been addressed at Geneva. b Its emphasis is on physical and logico-mathematical knowledge, rather than other kinds such as social or linguistic. Here the field of AI has to offer prccirely what Piagctian research lacks, and a glance at the table shows th&t I consider the converse to be true. III. HOW CAN WE PUT THEM TOGETHER? Given that we want to bring these two areas into contact, two possible ways of doing it come immediately to mind. First, there are a number of parallels between Piagetian concepts and ones introduced in AI, the most obvious being that between Piaget’s notion of a scheme and the schema/frame/script ideas in vogue the past few years in AI. The most important such parallels are: b Schemata. For Piaget, the fundamental unit of knowledge is the scheme, a generalizable action-pattern. b Schematically-based processing As a consequence of his scheme- based theory, Piaget views cognitive processes as primarily top-down. b Knowlcdgc as action. In Piagct’s epistemology, no knowledge is --- gained without acting and its conscqucnt transfclmaticn of experience. These parallels can be used not only to suggest new ways of developing the AI notions, but as ways of formulating the Piagetian concepts in more detail. Second, we can attempt to create AI models of Piaget’s theory (not tasks, as has sometimes been done). This will introduce another paradigm into AI research, no worse than the others and hopefully better. Simultaneously, such work can serve to rigorously develop the processing aspects of Piaget’s theory, using all the insights of AI research. Let me give a brief example. One of the ways that Piagctian theory is relevant to AI is that it can give both theoretical and empirical suggestions for how to design programs that learn, i.e., develop new cognitive structures in order to solve problems. Current AI learning programs work by assuming some initial body of knowledge and learning techniques, and then showing how additional knowledge and techniques could be learned. However, this creates a regress: how the starting knowledge and techniques are acquired is left for someone else to explain. In addition, the initial knowledge used is often much more advanced than what is to be acquired; for example, Sussman’s HACKER uses fairly sophisticated debugging techniques to learn something children acquire by age two. Piaget’s theory, on the other hand, tries to account for all of cognitive development starting with only the simplest of structures. Piagetian research reveals a number of important characteristics of problem solving that are not tilly achieved until adolescence, being gradually constructed starting around age five (ignoring the sensori- motor precursors). They are: b The ability to construct and overgcncralize theories, and as a corollary, to interpret phenomena in terms of confirmation or disconfirmation of those theories. b The ability to construct refifations as tests, as well as confirmations. b The presence of articulated part-whole relations, e.g., as reflected in an appropriate decomposition of the problem. Table 1. A Comparison of Piagetian Theory and Artificial Intelligence. Piaget Al Strengths 1. Motivated (biologically & epistemologically) 2. Comprehensive theory 3. Lots of empirical validation 4. Fundamental unit of knowledge S..Formal model of knowledge structures 6. Developmental theory Weaknesses 1. Not detailed 2. Not procedural Strengths Weaknesses -- 1. Precise/formal 1. Unmotivated 2. Process-oriented 2. Not theoretical 3. Detailed 3. Little empirical 4. Body of validation techniques 4. Little agreement 5. Narrow 6. No model of development 267 ) The ability to reflect on procedures in order to switch or modify them. ) The ability to coordinate goals and subgoals. Piagetian studies of how ,these abilities are constructed can provide useful ideas for AI. Consider, for example, how children learn to balance blocks (see [4]). For very young children, actions of objects are more or less random unless assimilated to their own actions and intentions. Their attempts to balance blocks ignore completely the relevant factors in favor of pushing down on the block to maintain a balance. Somewhat older children, after much trial and error, recognize a regularity in the outcomes of their efforts, namely, that most things balance at their geometrical centers. Once this idea (a primitive theory) is grasped, it is rapidly gcncralized, to the point where exceptions to it are dismissed as chance. The eventual overgeneralization of this notion of regularity, and consequent rcpeatcd failures to confirm it, give rise to the idea that the exceptions to the theory may have regularities themselves. At this point, children will retain their original theory where it works, but construct an additional one for the counterexamples (balance based on weight distribution). As their powers of conceptual coordination grow, they can combine the two theories into a single, more general, one. Then finally the idea of a retitation test usually appears (the preference for confirmations stemming from the same source as that of overgeneralizing in the first place). What does this suggest for AI work on learning and problem solving? If Piaget is correct, then in order to be generally successful, AI programs should have the abilitieamentioned, and they should be able to develop ,them in the manner given above. In particular, we can imagine a class of AI programs that approach problem solving in the following way. Faced with a new problem they would: ) engage extcnsivcly in trial-and-error behavior, ) strive to find regularities in the outcomes (including regularities in the procedures used), ) use the regularities to construct and then overgeneralize a theory, ) construct confirmation tests of the theory and note the exceptions, ) construct a theory for the exceptions, if they are numerous enough, ) try to coordinate the two theories into a larger one, ) test this resultant theory by confirmation and refutation tests. Besides theories of problem situations, the programs would also construct theories of the problem-solving procedures themselves, i.e., they would study the relationships of procedures among themselves and with respect to their outcomes, rather than just theories of the relationships among states of the world. Some of the above is partially captured in various AI programs (Sussman’s HACKER, Winston’s arch builder, erc.), but none of them are configured in exactly this way, and all of them assume fairly sophisticated techniques and knowledge as a basis for learning. The proposal above is obviously not the last word in learning and problem solving, but at least *provides a useful place to start. In addition, the vast amount of empirical data that Piagetians have collestcd can be invaluable as an empirical testing ground for AI models. Up to now, AI rcscarchers have only attempted to model performance on Pidgetian tasks, *vithout rcalbing that it is both the empirical and the theoretical constraint5 that would make Piagetian work of use to Al: were a model to simulate performance on a Piagetian task. and do so in a way fully consistent with the theory, it would have met a very strong test of its validity. Moreover, the tasks themselves have a high degree of ecological validity and cpistcmological significance (for example, conservation of matter and number). Thus, by making use of the existing parallels between the two approaches, and by recasting Piagctian concepts in AI terms, the two might produce a hybrid that would have all the advantages of a Piagetian foundation as well as all the benefits of Al’s hard-won insights into knowledge-representation and information-processing. For this to happen, however, it will be necessary for both Piagctians and Al researchers to learn more about each other’s work; as mentioned above, previous AI work has ncglccted Piagct’s theory in favor of his tasks, while Piagetians have only the vaguest notion of what a computational model is all about. And of course, Piaget’s theory is not some sort of panacea-psychology’s gift to AI. On the contrary, it needs to be much more developed, along the lines that AI work is pursuing. It is precisely because each could profit from the other that I’ve presented these points here. ACKNOWLEDGEMENTS I would like to thank Jonas Langer, John Sccly Brown, and Tom Moran for their comments. No endorsements implied. REFERENCES [l] Papert, S. “Etude comparee de l’intelligence chez l’enfant et chez le robot.” in La ZYiafion des Structures. Etudes d’Epistemologie G&nCtique. vol. 15. Paris: P.U.F. 1963. [2] Boden, M. “Artificial Intelligence and Piagetian Theory.” Synlhese. 38: 389-414. 1978. 131 -- Piagel. New York: Viking. 1979. [4] Karmiloff-Smith, A. and B. Inhcldcr. “If you want to get ahead, get a theory.” Cognirion. 3: 195-212. 1975. 268

1980

30

24

RI: an Expert in the Computer Systems Domain’ John McDermott Department of Computer Science Carnegie-Mellon University Pittsburgh, Pennsylvania 15213 INTRODUCTION. Rl* is a rule-based system that has much in common with other domain-specific systems that have been developed over the past several years [l, 81. It differs from these systeins primarily in its use of Match rather than Generate-and-Test as its central problem solving method [2]; rather than exploring several hypotheses until an acceptable one is found, it exploits its knowledge of its task domain to generate a single acceptable solution. Rl’s domain of expertise is configuring Digital Equipment Corporation’s VAX-l l/780 systems. Its input is a customer’s order and its output is a set of diagrams displaying the spatial relationships among the components on the order; these diagrams are used by the technician who physically assembles the system. Since an order frequently lacks one or more components required for system functionality, a major part of Rl’s task is to notice what components are missing and add them to the order. Rl is currently being used on a regular basis by DEC’s manufacturing organization.3 THE DOMAIN. The VAX-11/780 is the first implementation of DEC’s VAX-1 1 architecture. The VAX-1 l/780 uses a high speed synchronous bus, the sbi, as its primary interconnect; the central processor, one or two memory control units, up to four massbus interfaces, and up to four unibus interfaces can be connected to the sbi. The massbuses and particularly the unibuses can support a wide variety of peripheral devices. A typical system contains about 90 components; these include cabinets, periperal devices, drivers for the devices, and cables. There are a large number of rules that constrain the ways in which these components may be associated. RI’S DEFINING CHARACTEKISTICS. Rl is implemented in OPS4, a production system language developed at Carnegie-Mellon University [3, 71. An OPS4 production system ‘This paper describes Rl as it exists in June of 1980; it is a highly condensed version of [5]. 2 Four years ago I couldn’t even say *‘knowledge engineer”, now I . . . 3The development of Ri was supported by Digital Equipment Corporation. The research that led to the development of OPS4, the language in which Rl is written, was sponsored by the Defense Advanced Research Projects Agency. (DOD), ARPA Order No. 3597, and monitored by the Air Force Avionics Laboratory under Contract F33615-78-C-1151. The views and conclusions contained in this document are those of the author and should not be interpreted as representing the official policies, either expressed or implied, of Digital Equipment Corporation, the Defense Advanced Research Projects Agency, or the U.S. Government. VAX, PDP-11, UNIBUS, and MASSBUS are trademarks of Digital Equipment Corporation. consists of a set of productions held in production memory and a set of data elements (eg, state descriptions) held in working memory. A production is a rule composed of conditions and actions: Pi (Cl C2 . . . Cn --> A, A2 . . . Am) Conditions are forms that are instantiated by memory elements. Actions add elements to working memory or modify existing elements. The recognize-act cycle repeatedly finds all production instantiations and executes one of them.4 Rl exploits this recognition match. Its rules have conditions that recognize situations in which a particular type of extension to a particular type of partial configuration is permissable or required; the actions then effect that extension. OPS4’s two memories have been augmented,. for this application, with a third. This memory, the data base, contains descriptions of each of the 420 components currently supported for the VAX. Each data base entry consists of the name of a component and a set of eight or so attribute/value pairs that indicate the properties of the component that are relevant for the configuration task. As Rl begins to configure an order, it retrieves the relevant component descriptions. As the configuration is generated, working memory grows to contain descriptions of partial configurations, results of various computations, and context symbols that identify the current subtask.’ Production memory contains all of Rl’s permanent knowledge about how to configure VAX systems. Rl currently has 772 rules that enable it to perform the task.5 These rules can be viewed as state transition operators. The conditional part of each rule describes features that a state must possess in order for the rule to be applied. The action part of the rule indicates what features of that state have to be modified or what features have to be added in order for a new state that is on a solution path to be generated. Each rule is a more or less autonomous piece of knowledge that watches for a state that it recognizes to be generated. Whenever 40PS4’s cycle time, though it is essentially independent of the size of both production memory and working memory [4], depends on particular features of the production system (eg, the number and complexity of the conditions and actions in each production); the average cycle time for OPS4 interpreting Rl is about 150 milliseconds. CPS4 is implemented in MACLISP; Rl is run on a PDP- 10 (model KL) and loads in 4 12 pages Of core. 5 Only 480 of these ru!es are “configuration rules”; the remainder contain more general (non domain-specific) knowledge that Rl needs in order to use the configuration rules. 269 that happens, it can effect a state transition. If all goes well, this new state will, in turn, be recognized by one or more rules; one of these rules will effect another state transition, and so on until the system is configured. English translations of two sample rules are shown in Figure 1. ASSlGN-UB-MODULES-EXCEPT-THOSE-CONNECTlNG-TO-PANELS-4 IF: THE CURRENT CONTEXT IS ASSIGNING DEVICES TO UNIBUS MODULES AND THERE IS AN UNASSIGNED DUAL PORT DISK DRIVE AND THE TY PE OF CONTROLLER IT REQUIRES IS KNOWN AND THERE ARE TWO SUCH CONTROLLERS NEITHER OF WHICH HAS ANY DEVICES ASSIGNED TO IT AND THE NUMBER OF DEVICES THAT THESE CONTROLLERS CAN SUPPORT IS KNOWN THEN: ASSIGN THE DISK DRIVE TO EACH OF THE CONTROLLERS AND NOTE THAT THE TWO CONTROLLERS HAVE BEEN ASSOCIATED AND THAT EACH SUPPORTS ONE DEVICE degree of conditionality in the configuration task. The fan-in of a rule is the number of distinct rules that could fire immediately before that rule; the fan-out is the number of distinct rules that could fire immediately after the rule. The average fan-in and fan-out of Rl’s rules is 3. The graph of possible rule firing sequences, then, has 666 nodes, one for each of the rules (excluding the 106 output generation rules); each of these nodes has, on the average, three edges coming into it and three going out. It should be clear that unless the selection of which edge to follow can be highly constrained, the cost (in nodes visited) of finding an adequate configuration (an appropriate path through the rules) will be enormous. It is in this context that the power of the Match method used by Rl becomes apparent. When Rl can configure a system without backtracking, it finds a single path that consists, on the average, of about 800 nodes. When Rl must backtrack, it visits an additional N nodes, where N is the product of the number of unsuccessful unibus module sequences it tries PUT-UB-MODULE-6 IF: THE CURRENT CONTEXT IS PUTTING UNIBUS MODULES IN BACKPLANES IN SOME 80X AND IT HAS BEEN DETERMINED WHICH MODULE TO TRY TO PUT IN A BACKPLANE AND THAT MODULE IS A MULTIPLEXER TERMINAL INTERFACE AND IT HAS NOT BEEN ASSOCIATED WITH ANY PANEL SPACE AND THE TY PE AND NUMBER OF BACKPLANE SLOTS IT REQUIRES IS KNOWN AND THERE ARE AT LEASTTHAT MANY SLOTS AVAILABLE IN A BACKPLANE OFTHE APPROPRIATETYPE AND THE CURRENT UNIBUS LOAD’ON THAT BACKPLANE IS KNOWN AND THE POSITION OFTHE BACKPLANE IN THE BOX IS KNOWN THEN: ENTER THE CONTEXT OF VERIFYING PANEL SPACE FOR A MULTIPLEXER Figure 1: Two Sample Rules (which is rarely more than 2) and the number of nodes that must be expanded to generate a candidate unibus module configuration (which is rarely more than 300). Rl ‘S EVOLUTION. In a period of less than a year, Rl went from an idea, to a demonstration system that had most of the basic knowledge required in the domain but lacked the ability to deal with complex orders, to a system that possesses true expertise. Its development parallels, in many respects, the development of the several domain-specific systems engineered by Stanford University’s Heuristic Programming Project [2]. Rl’s implementation history divides quite naturally into two stages. During the first stage, which began in December of 1978 and lasted for about four months, I spent five or six days being tutored in the basics of VAX system configuration, read and reread the two manuals that describe many of the VAX configuration constraints, It is usual to distinguish the matching of forms and data from search: for example, in discussing the amount of search occurring in a resolution theorem prover, the unification of clauses is considered to be part of the elementary search step. But Match is also a method for doing search in a state space [6]; it is analogous to methods such as Hill Climbing or Means-ends Analysis, though much more powerful. The characteristic that distinguishes Match from other Heuristic Search methods is that in the case of Match the conditions (tests) associated with each state are sufficient to guarantee that if a state transition is permissible, then the new state will be on a solution path (if there is a solution path). Thus with Match, false paths are never generated, and so backtracking is never required. Match is well suited for the configuration task because, with a single exception, the knowledge that is available at each step is sufficient to distinguish between acceptable and unacceptable paths. The subtask that cannot always be done with Match alone is placing modules on the unibus in an acceptable sequence; to perform this subtask, Rl must occassionally generate several candidate sequences. and implemented an initial version of Rl (consisting of fewer than 200 domain rules) that could configure the simplest of orders correctly, but made numerous mistakes when it tried to tackle more complex orders.” The second stage, which lasted for another four months, was spent in asking people who were expert in the VAX configuration task to examine Rl’s output, point out Rl’s mistakes, and indicate what knowledge Rl was lacking. RI was sufficiently ignorant that finding mistakes was no problem. Given a criticism of some aspect of the configuration by an expert, all that was necessary in order to refine Rl’s knowledge was to find the offending rule, ask the expert to point out the problem with the condition elements in the rule, and then either modify the rule or split it into two rules that would discriminate between two previously undifferentiated states. During this stage, Rl’s domain knowledge almost tripled. VALIDATION. During October and November of 1979, Rl was involved in a formal validation procedure. Over the two month period, Rl was given 50 orders to configure. A team of six experts The fan-in and fan-out of Rl’s rules provide a measure of the ‘During this first stage, Rl’s name was XCON. 270 examined RI’s output, spending from one to two hours on each REFERENCES order. In the course of examining the configurations, 12 pieces of errorful knowledge were uncovered. The rules responsible for the errors were modified and the orders were resubmitted to RI and were all configured correctly. Each of these 50 orders contained, ‘. Amarel, S. et al. Reports of panel on applications of artificial intelligence. Proceedings of the Fifth International Joint Conference on Artificial Intelligence, MIT, 1977, pp. 994-1006. on the average, 90 components; RI fired an average of 1056 rules 2. and used an average of 2.5 minutes of cpu time in configuring Feigenbaum, E. A. The art of artificial intelligence. Proceedings of the Fifth International Joint Conference on each order. Since January of 1980, RI has configured over 500 orders. It is now integrated into DEC’s manufacturing organization. It has also begun to be used by DEC’s sales organization to configure orders on the day they are booked. Artificial Intelligence, MIT, 1977, pp. 1014-1029. to enlarge its domain so that it can become a more helpful system. CONCLUDING REMARKS. RI has proven itself to be a highly competent configurer of VAX-l l/780 systems. The configurations that it produces are consistently adequate, and the information Work has already begun on augmenting RI’s knowledge to enable that it makes available to the technicians who physically assemble systems is far more detailed than that produced by the humans who do the task. There are, however, some obvious ways in which 3. Forgy, C. L. and J. McDermott. OPS, A domain-independent production system language. Proceedings of the Fifth International Joint Conference on Artificial Intelligence, MIT, 1977, pp. 933-939. 4. Forgy, C. L. RETE: A fast algorithm for the many pattern/many object pattern match problem. Carnegie-Mellon University, Department of Computer Science, I 980. 5. systems. Carnegie-Mellon University, Department of Computer McDermott, J. RI : a rule-based configurer of computer Science, 1980. it to configure other computer systems manufactured by DEC. In those capabilities. Ultimately we hope to develop a salesperson’s addition, we plan to augment its knowledge so that it will be able to assistant, an Rl that can held a customer identify the system that help with the scheduling of system delivery dates. We al,so plan to augment RI’s knowledge so that it will be able to provide best suits his needs. interactive assistance to a customer or salesperson that will allow him, if he wishes, to specify some of the capabilities of the system he wants and let RI select the set of components that will provide 6. Newell, A. Heuristic programming: ill-structured problems. ln Progress in Operations Research, Aronofsky, J. S.,’ Ed.,John Wiley and Sons, 1969, pp. 361-414. 7. Newell, A. Knowledge representation aspects of production systems. Proceedings of the Fifth International Joint Confeience on Artificial Intelligence, MIT, 1977, pp. 987-988. 8. Waterman, D. A. and F. Hayes-Roth. Pattern-Directed Inference Systems. Academic Press, 1978. ACKNOWLEDGEMENTS. Many people have provided help in various forms. Jon Bentley, Scott Fahlman, Charles Forgy, Betsy Herk, Jill Larkin, Allen Newell, Paul Rosenbloom, and Mike Rychener gave me much encouragement and many valuable ideas. Dave Barstow, Bruce Buchanan, Bob Englemore, Penny Nii, Ted Shortliffe, and Mark Stefik contributed their knowledge engineering expertise. Finally, Jim Baratz, Alan Belancik, Dick Caruso, Sam Fuller, Linda Marshall, Kent McNaughton, Vaidis Mongirdas, Dennis O’Connor, and Mike Powell, all of whom are at DEG, assisted in bringing RI up to industry standards. 271

1980

31

25

RULE-BASED MODELS OF LEGAL EXPERTISE D. A. Waterman and Mark Peterson The Rand Corporation 1700 Main Street Santa Monica, California ABSTRACT This paper describes a rule-based legal de- cisonmaking system (LDS) that embodies the skills and knowledge of an expert in product liability law. The system is being used to study the effect of changes in legal doctrine on settlement stra- tegies and practices. LDS is implemented in ROSIE, a rule-oriented language designed to facilitate the development of large expert systems. The ROSIE language is briefly described and our approach to modeling legal expertise using a prototype version of LDS is presented. L. INTRODUCTION We are currently engaged in designing and building rule-based models of legal expertise. A rule-based model of expertise is a computer program organized as a collection of antecedent-consequent rules [l] that embodies the skills and knowledge of an expert in some domain. The primary goal of our work is to develop rule-based models of the de- cisionmaking processes of attorneys and claims ad- justers involved in product liability litigation. We will use these models to study the effect of changes in legal doctrine on settlement strategies and practices. Some progress has already been made in developing computer systems to perform legal analysis. The LEGOL Project [2] has been working for a number of years on the construction of a language for expressing legislation. In addition, systems have been developed for analyzing cases on the basis of legal doctrine 13,41, investigating the tax consequences of corporate transactions [S], automating the assembly of form legal documents if313 and performing knowledge-based legal informa- tion retrieval [7]. Our legal decisionmaking system (LDS) is being implemented in ROSIE, a rule-oriented language designed to facilitate the development of large ex- pert systems. In section II the ROSIE language is briefly described. Section III discusses our ap- proach to modeling legal expertise and describes the operation of our prototype version of IDS. The conclusions are presented in section IV. II. METHODOLOGY - A rule-oriented system for implementing exper- tise (ROSIE) is currently under development-to pro- vide a tool for building expert systems in complex domains 181. ROSIE is a direct descendant of RITA [9] and more distantly MYCIN [lo] in that the models created are rule-based with data-directed control [ll], and are expressed in an English-like syntax. In addition, the models use special language primitives and pattern matching routines that facilitate interaction with external computer systems. not found The ROSIE design also includes features in these successor systems, such as a hierarchical data structure capable of supporting abstraction and inheritance in a general way, par- titioned rulesets that can be called as subroutines or functions, a clearer differentiation between rule antecedent matching and iterative control by permitting actions that involve looping through the data base, and a user support environment with ex- tended facilities for editing and explanation. In the latest version of ROSIE, directed modules data structure are divided knowledge or rules and the declarative knowledge or facts. Both rules and facts are represented as the pattern- used to examine and modify the antecedent-consequent into the imperative where the consequent is either an action to be executed (for rules) or a statement to be deduced (for facts). Rules operate via forward chaining and are of two basic types: existence-driven (IF-THEN) as in RITA, and event- driven (WHEN-THEN) as in ARS [12]. Facts, on the other hand, operate via backward chaining and are represented only as IF-THEN pairs. The facts in ROSIE are similar to RITA goals, but are more gen- eral since they are implicitly referenced by the rules and automatically executed whenever the rules need information the facts can supply. In effect, the information that can be inferred from the facts is a %irtual data base" or extension to the stan- dard ROSIE data base. The current ROSIE syntax is more English-like than that of RITA or earlier versions of ROSIE. It is intended to facilitate model creation, modifica- tion and explanation. This syntax is illustrated in Figure 1, which shows our definition of strict liability in the product liability domain. 272 IF: THEN: assert the defendant is liable theory of strict-liability. under the FIGURE 1. Definition of Strict Liability in ROSIE III. LEGAL MODEL - -- The model of legal decisionmaking we are building will contain five basic types of rules: those based on formal doctrine, informal princi- pies, strategies, subjective considerations and secondary effects (see Figure 2). The formal doc- trine evolves from court decisions and statutes, while the informal principles, strategies, etc. are shaped by example and experience. Sources for these rules include legal literature, case his- tories and interviews with experts. By separating the rules as described we can study both the relevant inference mechanisms and the influence of each type of knowledge on the decisionmaking pro- cess. We are using our model of legal decisionmaking to systematically describe how legal practitioners reach settlement decisions and to test the effect of changes in the legal system on these decisions. Individual cases are analyzed by comparing the chains of reasoning (the chains of rules) that lead to the outcomes to similar chains in prototypical cases. This helps clarify the relationships exist- FORMAL DOCTRINE: rules used as the basis for legal judgements such as legislation and common law. INFORMAL PRINCIPLES: rules that don't carry the weight of formal law but are generally agreed upon by legal practitioners. This includes ambiguous concepts (e.g., reasonable and proper) and generally accepted practices (e.g., pain and suffering = 3 J; medical expenses). STRATEGIES: methods used by legal practitioners to accomplish a goal, e.g., proving a product defective. SUBJECTIVE CONSIDERATIONS: rules that anticipate the subjective responses of people involved in legal interactions, e.g., the effect of plaintiff attractiveness on the amount of money awarded, or the effects of extreme injuries on liability decisions. SECONDARY EFFECTS: rules that describe the interactions between rules, e.g., a change in the law from contributory negligence to comparative negligence may change other rules such as strategies for settlement or anticipated behavior of juries. FIGURE 2. Components of Legal Decisionmaking ing between the formal doctrine, informal practices and strategies used in the decisionmaking. We are examining the effects of changes in legal doctrine, procedures and strategies on the processing of cases by modifying appropriate rules in the model and noting the effect on the operation of the model when applied to a body of selected test cases. This can provide insights that will suggest useful changes in legal doctrine and practices. Our current implementation of LDS is a small prototype model of legal decisionmaking containing rules representing negligence and liability laws. This prototype contains rules describing formal doctrine and informal principles in product liabil- ity. Future versions of the system will incor- porate the other rule types shown in Figure 2. The model consists of approximately 90 rules, half of which represent legal doctrine and principles. Given a description of a product liability case the model attempts to determine what theory of liabili- ty applies, whether or not the defendant is liable, how much the case is worth, and what an equitable value for settlement would be. Once a decision is reached the user may ask for an explanation in terms of the rules used to reach the decision. We will now describe the use of LDS to test the effect of a legislative change on a case out- come. The case is briefly summarized in Figure 3, while the operation of the model on this case is illustrated in Figure 4. The system was first ap- plied using the definition of strict liability given in Figure 1. It was determined that the de- fendant was partially liable for damages under the theory of comparative negligence, with the amount of liability lying somewhere between $21,000 and 273 The plaintiff was cleaning a bathtub drain with a liquid cleaner when the cleaner exploded out of the drain causing severe burns and permanent scarring to his left arm. Medical expenses for the plaintiff were $6000, and he was unable to work for 200 working days, during which time his rate of pay was $47 per day. The cleaner was manufactured and sold by the defendant, Stanway Chemical Company. The contents of the product were judged not to be defective by experts retained by the defendant. The product's label warned of potentially explosive chemical reactions from improper use of the product, but did not give a satisfactory description of means to avoid chemical reactions. The plaintiff was familiar with the product but did not flush out the drain before using the cleaner. The amount of the claim was $40,000. FIGURE 3. Description of Drain cleaner Case (Note: the model actually used a much more detailed description of of the case than is shown here-.) use was reasonable QI IU PI U)JW $29,000. The case was valued between $35,000 and $41,000. After the definition of strict liability was modified to state that the product must be un- reasonably dangerous for strict liability to apply, the defendant was found to be not liable. In this prototype implementation of LDS a somewhat more restrictive ROSIE rule syntax was used than is shown in Figure 1. v. CONCLUSIONS Our preliminary work with LDS has demonstrated the feasibility of applying rule-based modeling techniques to the product liability area. In spite of the inherent complexity of product liability law, the number of basic concepts manipulated by the rules is easily handled (in the hundreds), while the number of rules required to adequately represent legal doctrine and strategies is manage- able (in the thousands). The rules that represent legal doctrine in this area are basically declarative in nature. use was * rl lforeseeable no strict Droduct was defective defendant manufactured product - product not unreasonably dangerous victim’s responsibility = .4 victim was not a minor victim knew of hazard \- liability y - - - - defenclant -la/, Y “9 . . _ ,iabi.ity total amount of loss is between $35,000 and $41,000 location not dangerous z r6 medical expenses were $6136 lost 228 working days __IIc base pay of $47 per day / victim iparative :ligence pa I defendant’s liabilitv = .6 FIGURE 4. Inference Process for Drain Cleaner Case (Crosshatched area shows inference before law change) 274 Most of them are easily represented as definitions with complex antecedents and simple consequents that name the concept being defined. Rules of this sort can be organized as relatively unordered sets that are processed with a simple control scheme. Most of the action takes place in calls to other rule sets representing definitions of terms used by the initial set. This simple control structure fa- cilitates rule modification and explanation. In this application area improved methods are needed for dealing with vague or ambiguous concepts used in the rules. It is sometimes difficult to decide whether or not these concepts are applicable in a particular case, e.g., whether the use of the product was actually "reasonable and proper." Pos- sibilities include gradual refinement: a query scheme involving presenting the user with increas- ingly specific sets of questions, each of which may have ambiguous terms that will be further refined by even more specific query lists, and analogy: displaying case histories involving similar proto- typical concepts and having the user select the one closest to the term in question. ACKNOWLEDGMENTS This work has been supported by a grant from the Civil Justice Institute at the Rand Corpora- tion, Santa Monica, California. [71 [al [91 (101 [Ill [=I Process: A Computer That Uses Regulations and Statutes to Draft Legal Documents." American Bar Foundation Research Journal, No. 1, (1979) 3-81. -- Hafner, C.D., "Representation of Knowledge in a Legal Information Retrieval System" In Research and Development in Information Re- trieval. Proceedings of the Third Annual SI- GIR Conference, Cambridge, England, 1980. Waterman, D.A., Anderson, R.H., Hayes-Roth, F s:;. , Klahr, P., Martins, G., and Rosenschein, Design of 2 Rule-Oriented System for Implementing Expertise. - N-~~s~-ARPA, The Rand Corporation, Santa Monica, California, 1979. Anderson, R.H., and Gillogly, J.J., Rand In- - - telligent Terminal Agent (RITA): Design Phi- losophy. R-1809-ARPA, The Rand Corporation, Santa Monica, California, 1976. Shortliffe, E.H., Computer-Based Medical Con- --- sultations: MYCIN. New York: American El- sevier, 1976. Waterman, D.A., and Hayes-Roth, F. Pattern- Directed Inference Systems. New York: Academic Press, 1978. Sussman, G.J. "Electrical Circuit Design: A Problem for Artificial 1977 Intelligence Research." In Proceedings of the 5th Inter- --- - national Joint Conference on Artificial In- telligence, Cambridge, Massachusetts, 1977, 894-900. REFERENCES [II [21 [31 [41 [51 [61 Waterman, D. A., "User-Oriented Systems for Capturing Expertise: A Rule-Based Approach." In D. Michie (ed.) Expert Systems in the Mi- -- - cro Electronic &. Press, 1979. Edinburgh University Jones, S., Mason P.J., & Stamper, R.K. "LEGOL-2.0: A Relational Specification Language for Complex Rules." Information Sys- terns, 4:4, (1979). Meldman, J. A. (( A Structural Model for Computer-aided Legal Analysis." Journal of Computers and Law, 6 (1977) 27-71. - - -- Popp, W.G., and Schlink, B., "JUDITH: A Com- puter Program to Advise Lawyers in Reasoning a Case." Jurimetrics Journal, 15:4 (1975) 303-314. McCarty, L.T., and Sridharan, N.S., "The Representation of an Evolving System of Legal Concepts, Part One: Logical Templates" In Proceedings of the Third National Conference - _-- of the Canadian -- Society for Computational Studies of Intelligence', Columbia,T980. Victoria, British Sprowl, J.A., "Automating the Legal Reasoning 275

1980

32

26

EXPLOITING A DOMAIN MODEL IN AN EXPERT SPECTRAL ANALYSIS PROGRAM David R. Barstow Schlumberger-Doll Research, Ridgefield, Connecticut 06877 ABSTRACT Gamma ray activation spectra are used by nuclear physicists to identify the elemental composition of unknown substances. Neutron bombardment causes some of the atoms of a sample to change into unstable isotopes, which then decay, emitting gamma radiation at characteristic energies and intensities, By identifying these isotopes, the composition of the original substance can be determined. GAMMA is an expert system for performing this interpretation task. It has a detailed model of its domain and can exploit this model for a variety of purposes, including ratings for individual isotopes and elements, ratings based on multiple spectra, complete interpretations, and even calibration. GAMMA's performance is generally quite good when compared with human performance. 1. INTRODUCTION Gamma ray spectra are commonly used by nuclear physicists to identify the elemental composition of a substance. One kind of gamma ray spectrum (an "activation spectrum") is produced by bombarding a sample of the substance with neutrons. This causes certain changes in some of the atoms in the sample, many of which result in unstable isotopes that then begin to decay. As they decay, the unstable isotopes emit gamma rays at characteristic energies and intensities. By measuring these, the unstable isotopes (and from these, the elements of the original sample) can be identified. For example, Figure 1 shows such a spectrum, with peaks identified and labeled by a physicist. In this case, the peaks were produced by emissions from the isotopes Na-24, Cl-37, and S-37; the original substance was a sample of salt. An expert system, called GAMMA, has been developed to perform this task, and GAMMA's performance compares well with human interpreters. The basic strategy employed in developing GAMMA was to develop a detailed model of the domain and then to exploit this model for a variety of tasks and situations. Early work on GAMMA was discussed in another paper[ll; in this paper, recent progress will be discussed. 2. The Domain Model GAMMA's domain model was described in detail in the earlier paper and will only be summarized here. Basically, the process that produces gamma ray activation spectra can be seen at six different levels as follows: Figure 1: Gamma Ray Activation Spectrum 276 (1) elements in the original sample 3.1. Isotopic and Elemental Ratings (2) isotopes in the original sample (3) isotopes after bombardment by neutrons (4) decays of unstable isotopes (5) gamma ray emissions during decay (6) g-a ray detections during decay Level 6 represents the actual spectrum and level 1 represents the ultimate goal of the interpretation. Level 3 is a convenient intermediate level used by most practicing nuclear physicists. GAMMA’s use of this domain model involves considering hypotheses at each of the levels. A hypothesis consists of a set of triples, each triple consisting of an object appropriate to the level (e.g., naturally occurring isotope for level 2; gamma ray emissions at a particular energy for level 5), an estimated concentration for the object (e.g., the number of decays of a given unstable isotope for level 4), and a label which encodes the path from level 1 to the triple (e.g., “NA-23/NG/NA-24,’ denoting that the unstable isotope Na-24 was produced when the naturally occurring isotope Na-23 underwent an N-y transition). Relationships between levels can be expressed in terms of several formulae that have been derived from both theoretical considerations and empirical observations. These formulae involve such parameters as the likelihood of particular isotopic transitions during neutron bombardment, the half-lives of unstable isotopes, and the characteristic gamma ray energies and intensities for different isotopes. Nuclear physicists consult published reports when they need such information; the data from one such source[21 has been converted into a LISP data base for GAMMA’s use. Further details of the formulae and data base are available elsewherec 11. In GAMMA’s case, the formulae are all used predictively; that is, given an hypothesis at one level, the appropriate formula can be used to predict hypotheses at the next level down. By chaining such predictions together, GAMMA can go from hypothetical interpretations at levels 1 or 3 down to predicted gamma ray detections that can be compared against spectral data. 3. Applications of the Doglain Model The accuracy with which predictions can be made and the high resolution of this particular detector enable GAMMA to exploit the domain model in a variety of tasks and situations. Some of these were discussed earlierElI, and will be mentioned here only briefly. GAMMA was first used to “rate,, the likelihood that any particular unstable isotope was present after neutron bombardment. This was done by hypothesizing decays of that isotope at level 4, predicting detections at level 6, and using an evaluation function to compare the predictions with the spectral data. The evaluation function was designed to take positive and negative evidence into account, allowing both for background radiation and for noise and errors in the prediction and detection processes. The peaks were individually rated for both energy and intensity, and the final rating was the average of the individual ratings, weighted by intensity (i.e., stronger peaks were more important). When a predicted peak had a corresponding detected peak, a positive rating was given; when no peak was detected, a negative rating was assessed, unless the predicted intensity was low enough that the peak could have been obscured by background radiation. Noise and errors were taken into account by using what we call the trapezoidal rule. For example, the trapezoidal rule for peak energies is shown in Figure 2. If a peak was predicted at energy E, then a detected peak within the range (E-6 E+b > was considered a perfect match, peaks out&e thJ range (E- ’ all, and peaks in 8; E?a!$e~e~~-~t ?Za!?:edazdt (E+& E+h2)were scaled to provide 2a’ con&nuous fun&on. Such trapezoidal rules were used throughout GAMMA’s evaluation function, and has proved quite adequate. GAMMA’s performance at the isotopic rating task was moderately good compared with that of human experts: although it gave high ratings to isotopes identified by experts, it also occasionally gave such ratings to isotopes considered implausible by the experts. GAMMA’s second task was to do a similar rating for elements in the original sample (i.e., hypotheses at level 1). The same predict-and-match technique was used, and GAMMA, s performance was again moderately good, although not quite as good as in the isotopic case: fewer implausible elements were rated but some elements identified by the human experts received low ratings. This was due largely to certain simplifying assumptions in the formulae relating levels 2 through 4. Further details of GAMMA’s rating scheme are given elsewhere[ll. Recently, GAMMA’s repertoire has been expanded to include several other tasks, and its performance seems to have improved with age. rating 1.0 -1 .o -I -Gy+ A-A-+- energy 62 blE% b2 Figure 2: Trapezoidal Rule for Peak Energies 277 3.2. Ratings for Multiple Spectra GAMMA's next major task was to do similar ratings for individual isotopes and elements, but to do so on the basis of multiple spectra: in a typical experimental situation, not one but several spectra are recorded, each for a different time interval. Generally, the first few spectra are for comparatively short time periods (10 to 30 seconds), and the later spectra may be for periods as long as several hours. The primary advantage of multiple spectra is that they permit greater use of half-life information: unstable isotopes with short half-lives will appear on the earlier spectra but not on the later ones; isotopes with longer half-lives emit ganrma rays at roughly constant rates, so they appear most distinctly on the later spectra (for which the detection time is longer). The technique used by GAMMA is to find the isotopic (or elemental) ratings for the individual spectra, use these to hypothesize an initial concentration for the isotope, redo the predictions based on this concentration, and finally combine the ratings for these predictions into a single overall rating. The hypothetical concentration for an isotope is determined by considering all spectra for which the isotope's rating is sufficiently high, taking the concentrations (byproducts of the original prediction-and-match rating) that agree sufficiently well (within one order of magnitude), and then computing the average. This technique is designed to ignore those results which, for any of several reasons, deviate from the norm and in practice seems to work quite well. Given this hypothesized concentration, the prediction-and-match rating is again computed for each individual spectrum. These ratings are then averaged to determine the overall ,,multiple spectra,, rating for the isotope. In our first attempt to average these ratings, we weighted them by the total predicted intensity for a spectrum, as was done for ratings within individual spectra. But this seemed to attach too much weight intensities, to spectra with high predicted so on our second attempt, we took the simple average of the ratings for all spectra for which the evidence was significant (either positive or negative), and the results were much better. (Interestingly, the first symptom of this problem was due to an INTERLISP error: under certain circumstances, INTERLISP ignores floating point overflow and underflow, thereby producing a very large number when multiplying two very small ones. With simple averaging, such isolated erroneous computations no longer have much overall effect. In fact, we now take this as a maxim: no numeric rating scheme should depend too heavily on any single data point.) GAMMA's performance on multiple spectra is generally much better than on individual spectra, primarily because of the value of half-life information. GAMMA's ratings generally compare well with those of human experts, and implausible isotopes (Or elements) are only rarely given high ratings. 3.3. Producing a Complete Interpretation The major problem with the tasks described so far is that the ratings are given to isotopes and elements as if they were totally independent of each other. The fact that the same peak may be caused by emissions from two different isotopes does not detract from the rating of either one. The ultimate interpretation of spectral data should not be ratings for individual elements, but rather a set of elements (and concentrations) which, taken together, explain the data well. A first pass at coming up with such a complete interpretation might be to take all and only those elements with sufficiently high ratings, but that does not take into account the interaction between the elements, and is simply inadequate. GAMMA's solution to this problem is essentially a hill-climbing algorithm designed to maximize an "interpretation measure,,. For this algorithm, a complete interpretation is defined to be a set of <element, concentration> pairs, and a mapping of detected peaks to sets of labels. (The labels describe the path from one of the <element, concentration> pairs to the detected peak. Under this definition, the same detected peak may have several different labels, a situation which actually occurs in the spectra under consideration.) The interpretation measure that GAMMA currently uses is based on two different considerations. First, the individual spectra are rated in terms of (1) how many peaks have no labels (i.e., are there peaks which are not explained by the interpretation?), (2) how many labels have no peaks (i.e., are there predictions which do not appear in the detected spectra?), and (3) how well the peaks and associated labels match (i.e., do the energies and intensities of the detected peaks match well with the energies and intensities predicted for the associated labels?). The second consideration is that the relative concentrations of the elements be plausible. This is used only as negative evidence: if the concentration of an element is high (relative to the concentrations of the other elements), but the rating for that element is low, then the interpretation is suspect, since the detector and model can be expected to be quite accurate for relatively pure substances; if the concentration is below a certain threshold, then the interpretation is also suspect, since the detector simply cannot be expected to find elements in such smali concentrations: The task is thus to find the set of <element, concentration> pairs that maximizes this measure. GAMMA uses the following hill-climbing algorithm: INTERPRETATION := 0; CANDIDATES := {<element, concentration> I rating is above a threshold); consider all interpretations formed by moving one element from CANDIDATES to INTERPRETATION or from INTERPRETATION to CANDIDATES; if no such interpretation increases the measure then quit else select that which maximizes the measure; update INTERPRETATION and CANDIDATES; repeat. 278 While we have no theoretical basis for claiming that this algorithm does, indeed, find the subset of candidates with maximal measure, our experience indicates that it performs very well, and the interpretations that GAMMA produces are quite good. 5. References Cl1 Barstow, D.R. Knowledge engineering in nuclear physics. In Sixth International Joint Conference on Artificial Intelligence, pages 34-36.- Stanford Computer Science Department, 1979 3.4. Calibration GAMMA’s final task is to calibrate the spectra. Up to this point, it has been assumed that the input spectra have already been calibrated (spectral channels have been associated with gamma ray energies), and this is a task which has hitherto been performed by physicists before the data are given to GAMMA. We have not yet completed the attempts at solving this problem, but our first results are encouraging : we have developed a technique for recalibrating a spectrum more precisely, given an initial approximate calibration. The basic technique is to use a set of good calibration priori by physicists) isotopes (as identified a and look for any which satisfy two criteria: (1) there is exactly one match within a given range of the initial calibration; (2) that match gets a fairly high rating. A linear least squares fit of these points gives the recalibration. Experience with this approach is quite good, and it is currently used to correct for calibration differences among the individual spectra within a set of spectra. c21 Erdimann , G . Neutron Activation Tables. WChemie , New York, 1976 . c31 Shortliffe, E.H. MYCIN: Computer-Based Medical Consultations. American Elsevie-w-1975 . Our plan for finding the initial calibration is to take a sum spectrun for the entire set (thereby increasing the data/noise ratio) and apply a similar strategy: find any ,,calibration,, isotope which has exactly one match anywhere on the sum spectrum and for which the match rating is quite high. We hope to know soon whether this approach will succeed. 4. Shallow and Deep Domain Models GAMMA’s success is due largely to its use of a relatively detailed model of its domain. This may be compared with systems such as MYCINC31 whose success is due largely to shallow models that encode, in a sense, an expert’s compiled version of a more detailed model that the expert may or may not know explicitly. In comparing these two approaches, several observations can be made. First, while a deep model can be put to great use (as it was in GAMMA), there are several circumstances in which a shallow domain model is necessary: (1) a deep model doesn’t exist (e.g., there is no computational theory of the entire hunan body) ; and (2) a deep model is not computationally feasible (e.g., one cannot hope to do weather prediction based on the fundamental properties of gases). Second, although shallow models will often suffice, it seems likely that future expert systems based on shallow models will require access to deep models for difficult cases. Third, “deep,, and ,,shallow,, are obviously relative terms : GAMMA’s ,deep model is shallow when viewed as a model of subatomic physics. The relationship between deep and shallow models seems to be an important topic for future work on expert systems. 279

1980

33

27

Project EPISTLE: A system for the Automatic Analysis of Business Correspondence Lance A. Miller IBM Thomas J. Watson Research Center Yorktown Heights, New York 1 0598 ABSTRACT: The developing system described here is plan- ned to provide the business executive with useful applications for the computer processing of correspondence in the office environment. Applications will include the synopsis and abstraction of incoming mail and a variety of critiques of newly-generated letters, all based upon the capability of understanding the natural language text at least to a level corresponding to customary business communication. Succes- sive sections of the paper describe the Background and Prior Work, the planned System Output, and Implementation. I. BACKGROUND AND PRIOR WORK We conclude from these behavioral findings that there are indeed extensive regularities in the characteristics of business letters, determined primarily by the purpose objec- tives. It is these constraints that most strongly indicate to us the feasibility of developing automatic means for recognizing content-themes and purposes from the letter text (as well as the converse, generating letter text from information about purposes). Other analyses have been undertaken to estimate the linguistic complexity and regularities of the texts. The aver- age letter appears to contain 8 sentences, with an average of 18 words each; in the 400 letter-bodies there are roughly 57900 words and 4500 unique words totaL An ongoing hand analysis of the syntactic structure of sentences in a 50-letter sample reveals a relatively high frequency of subject-verb inversions (about 1 per letter) and complex lengthy complementizers (l-4 per letter). These features, along with very frequent noun phrase and sentence coordina- tion, accompanied by a wide variety of grammatical but un- systematic structure deletions, indicate an exceptionally high level of grammatical complexity of our texts. With respect to overall text syntax we have analyzed 10 letters for text cohe- sion, using a modification of Halliday and Hasan’s coding scheme E41; 82 percent of the instances of cohesion detect- ed were accounted for by 4 categories: lexical repetitions (29%), pronouns (28%), nominal substitutions (9%, e.g., “one”, “same”), and lexical collocations (words related via their semantics, 16%). In an extension of this discourse structure analysis we are analyzing 50 letters, coding all occurrences of functional nouns in terms of (1) the grammat- ical case function served and (2) the cohesive relation to prior nouns. Preliminary results indicate consistent patterns of case-shift and type of cohesion as a function of the prag- matic and content themes. The results of these linguistic analyses will help determine the strategy ultimately adopted for selecting surface parses and meaning interpretations. II. SYSTEM OUTPUT The planned system will provide the following for each letter in our database: (1) surface syntactic parses for each sen- tence; (2) meaning interpretations for each sentence, adjust- ed to the context of prior sentences; (3) a condensed synop- sis of the overall meaning content of the letter; (4) a critique of each letter’s spelling, punctuation, and grammaticality; (5) a mapping of the meaning content onto common business communication themes; and (6) some characterization of the author’s style and tone. In addition to the above, we plan to develop a limited facility to generate short letters of a certain type (e.g., information requests) conforming to a particular author’s normal “style” and “tone”. III. IMPL.EMENTATION COMPONENTS A. Semantic Representations: Many of the lexical items in our texts appear to have two or more literal word-sense usages (as well as occasional non-literal ones); it also appears that the discriminating semantic features among highly- related lexical items cannot be ignored if the intended letter nuances are to be preserved. We therefore do not expect much reduction in cardinality when mapping from the lexical to the concept space; we also anticipate that our representa- tions will have to be unusually rich, in terms of both a large number of features distinguishing the concepts underlying lexical items and the capability to relate different concepts together. Among the most important anticipated semantic features are those describing the prcxmdhns and cmse- quences of ACTIONS and those characterizing the internal states of ACTORS (e.g., their intentions, expectations, and reactions). B. Parsing: We will employ a system called NLP as the basic “operating system” for our application development. This system has been used for other understanding projects and provides all of the general components required, including a word-stem dictionary, a parser, knowledge representation, and natural language generation I: 25 1. In particular, the parser proceeds left to right, character by character, in proc- essing a sentence, generating all possible descriptions of text segments in a bottom-up fashion by application of rules from an augmented phrase structure grammar -- essentially a set of context-free rules augmented with arbitrary conditions and structure-building actions. In writing the grammar we are attempting to keep to an absolute minimum the use of se- mantic information, to increase the applicability of the parser over a variety of semantic domains. Our general strategy to minimize the number of surface parses (and increase parsing efficiency) is to attach lower syntactic constituents (e.g., post-nominal prepositional phrases) to the highest-possible unit (e.g., to the verb-phrase rather than the noun-phrase), with the final decision as to the most appropriate attachment to be resolved by the interpretation components. C. Meaning-Assignment: Our planned strategy for choosing the meaning to assign to a sentence is basically to find that action-concept whose case-relations are most completely “satisfied” by all of the concepts implied by the sentence. In the expected frequent case of only partial “fits” to several action-concepts, prefer- ence among these will be based on several factors, including: (1) the number of case relations of a concept “filled” or “unfilled” by elements of the present (or prior) text, and the relative importance in the intensional definition of each of these; (2) the “directness” of the mappings of text segments to underlying concepts; and (3) the syntactic structure of the sentence (e.g., syntactically “higher” components usually will be preferred to “lower” ones). D. Text-Interpretation: We propose to keep separate lists of each action-concept and entity-concept encountered in the text. Following the meaning-assignment to a sentence, the sentence will be re- examined to determine if it supplies qualification information for any prior-mentioned action or entity; if so, these separate representations will be so augmented, a process called “updating”. Statistics of each such updating of information will be kept for each sentence for subsequent characteriza- tion of. style. Next, these separate entity/action representa- tions will be examined directly to determine whether they can be combined as elements of some broader concept. By this process we will therefore be able to update and condense our representations as we go along, facilitating eventual synopsis and abstraction of content-themes. In addition to the above semantic interpretations for the complete text, we will also build up a composite representa- tion of the syntactic structure of the text. We, first, are hopeful of being able to discover a relatively small number of schema for characterizing syntactic structures within sen- tences:wetllenheJieYethatthe~Qfleuer~canbe accounted for these schemas. in terms of E. Adaptation: frequently occurring patterns of to As a unique feature, we plan to implement the capability dynamically modify or adapt our system so as to change the manner in which word-senses are selected or meanings assigned as a function of the system’s experience with various kinds of texts. This would be accomplished by assigning “weight” attributes to each lexical item and to each underly- ing concept (and its attribute-values); the weight-values along the path finally selected for mapping text to concepts would then be incremented by some amount. Given that preference ordering of text to concept paths is determined by such overall path-weights, the system could thus achieve self-adaptation to word-usages of particular application do- mains. This facility could also be employed to characterize individual authors’ styles. F. Abstraction and Critique of - Letters: Concerning Content-Themes and Purposes, we plan to map the system’s meaning interpretations onto a set of com- mon business content-themes and communication purposes, and we are presently conducting behavioral and analytical studies to determine these. With respect to Grammaticality, we anticipate being able to detect incomplete sentences, subject-verb disagreements, and inappropriate shifts in verb tenses; in addition, we will be able to identify ambiguities and some instances of clearly “awkward” syntax. Spelling errors of the “non-word” type are easily caught, and certain spelling errors in which the misspelled word is in the diction- ary may also be caught if they contain sufficient syntactic information. In addition, some fraction of “spelling” errors involving semantically inappropriate words should be detecta- ble. Finally, we may be able to discover a number of Punctuation errors. The last aspect of critiques is that of style and tone. We are aware of the several “indices” for measuring various aspects of these but consider them to be at best very crude indicators C 61. As a starting point we have identified five dimensions for each concept, and we will implement the capability to assess texts on these dimensions until we are better informed. For S~J&, defined as “the organizational strategy for conveying content”, the dimensions are: sentence precision, sentence readability, reference clarity, information- value, and cohesion. Tone, defined as “the connotations of interpersonal attitudes”, is to be rated on the dimensions of: personal-ness, positive-ness, informal-ness, concrete-ness, and strength. We plan to output and highlight those text segments which fall below a certain level of acceptability on these measures. REFERENCES Cl1 Miller, L. A. “Behavioral studies of the programming process. ” IBM Research Report RC 7367, 1978. c 21 Heidorn, G. E. “Augmented phrase structure gram- mars”. In Nash-Webber, B.L. and Schank, R. C. (Eds.), Theoretical Issues in Natural Language --- - - -- Processing. Association for Computational Linguis- t.ic& .hle+ 1975. 281 c31 Miller, L. A. and Daiute, C. “A taxonomic analysis of c51 Heidorn, G. E. “Automatic programming through natu- business letters”. IBM Research Report, in prepara- ral language dialogue: A survey”. IBM Journal of --- tion, 1980. Research and Development, 1976, 20 302-3 13. -- - 141 Halliday, M. A. K. and Hasan, R. Cohesion in English. -- London: Longman Group Ltd., 1976. C61 Dyer, F. C. Executive’s Guide to Effective Speaking --- and Writing. Englewood Cliffs, N. J.: Prentice-Hall, Inc., 1962. 282

1980

34

28

A KNOWLEDGE BASED DESIGN SYSTEM FOR DIGITAL ELECTRONICS* Milton R. Grinberg Department of Computer Science University of Maryland College Park, Maryland 20742 1. Overview and Goals of the SADD System v----p the goal of a human ex e;ia;; such as "build a digits P display interface into a This is a problem solving activity in which the roblem expert's general purpose R solving abilities interact with his rich nowledge of the digital world. initial ideas By translatin through successively more !! the re ined "sketches", the human expert gradually arrives at a chip-level digital schematic that realizes the initial goals, and which may have bugs that will only be discovered after the circuit has been simulated or built and tested. The Semi-Automatic Digital Designer (SADD) is an experimental interactive knowledge-based design s stem, II whose domain of expertise is digital e ectronics. First, I want The SADD project has two goals. to provide an intermediate digital design problem solver for which a human expert can interactively provide the high-level functional descriptions-of-a circuit and which can take . high-level circuit descri;tion and refine it in?: ciycuit Second, schematic that performs the required task. corn uter guman iii I am ;EaEyptmg to discover and express in that_.knowledge esig;t; an expert in.digital that makesT;F; desi n. * includes f eneric information K or each hi~h-le~~l d;g,f;zdsfun;;;c!;d(e.g., counter, clock) transform these functional descriptions into r%izable circuits. As a first case study in relative1 desi n, sophisticated TV video % I adopted a called 4: isplay circuit t e Screensplitter. This is a real circuit of moderate overall complexity. 2. Digital Design -- How does a digital desi designing a circuit? The % ner go about the task of esigner generally starts with a well-defined goal for a circuit with only an ill-defined solution for accomplis$;ng Using his design experience, that goal. designer can pinpoint some of the required and hence pivotal components needed in the circuit. The circuit is slowly sketched around these The pivotal sketching process allows components. specify where the inputs come from the designer to and where the outputs go. concerning The designer of:;: makes notes characteristics of I$f;;;g this sketch phase, components. the expert discovers that components are needed and adds them to the design. each Eventually the designer starts to refine component by selecting and interconnecting chips to implement the component. What are the "primitives" that a designer uses? At the im leme$afion*(i.$., final circui;idlevel R $h;;; a;zdt ree prim+tives : (2)(l) a chip its wire and si nal. (3) a f* The"%!;tisPin?&rction those &tputs are de ined by its *current state ;~~,,a;;,,current A wire i?!iE:siefining anizleFtrica1 equivalence. A ST:: physical f is the electrical information present on a wire. At the sketch ad level "primitives": (17 a there are two additional f unction and (2) a connection. *This research is sup orted b the Office of Naval Research under Grant 800014-765-0477. A function performs some desi nated digital task. I have concentrated on nine 5ifferent functions (i.e., counter, shifter, memory, combinational d divider, selector, and clockj in the Screensplitter. These are components that a designer uses in the A connection is an information path between- two functions. It identifies where information flows and when it is allowed to flow. 3. SADD Organization There are three distinct design phases in the In the specification acquistion phase, the user describes the functionatsstructure offEk;e;ircuit in English. During . phase, are introduced relevant information about the frames is filled in, and the interrelationship among the frames established. From builds a model of this description, SADD the semantic net. circuit, expressed as a component is selected using the conceptual function &zracteristics (which may-have to be deduced from characteristics provided by the designer and the functions relationship to the other functions in the model). the circuit schematic is implemented using '%: selected and conceptual function description. strategy In the circuit simulation phase, the correctness of the circuit is determined by simulating the circuit on a conceptual simulator. circuit proves not to If the designed fulfill the goal of the desi ner during the simulation phase, it can be modi led and redesigned. 3. 3.1. Screensplitter Circuit In order to develop SADD, the Screensplitter was chosen as a benchmark. Fig. 1 illustrates one of DCCBR READ DM ADDRESS DCCBR LOAD STCOL Fig. 1. Schematic for the DCC Counter the 12 logical components (the Display Character 283 Counter) from the original Screensplitter circuit schematic. the circuit schematic design scenerYii%s develoned a verbal that descsibes the funcrional components, characteristics of these functional components and the interconnections. Fig. 2 shows the relevant portion of the input 1. 2. 3. 4. 5. THE DISPLAY CHARACTER COUNTER (DCC) COUNTS FROM . _ 0 TO 3519. WHEN THE COUNT OF THE SLC EQUALS 4, THE COUNT OF THE DCC CAPTURED-IN THE LOAD OF THE DCC-BASE REGISTER (DCCBR) WHEN THE HORIZONTAL BLANK (HBLANK) BEGINS. _ THE DCC CAPTURED-FROM THE DCC-BASE REGISTER (DCCBR) WHEN HORIZONTAL BLANK (HBLANK) BEGINS. EACH-TIME THE COUNT OF THE PIXEL COUNTER (PC) EQUALS 5 ENDS, THE DCC INCREMENTS. THE COUNT OF THE DCC CAPTURED-IN THE ADDRESS OF THE DISPLAY MEMORY (DM). Fig. 2. Scenerio for the DCC Counter description for the Display Character Counter. The complete scenerio is 41 sentences. 3.2. Parser A phrase-keyword parser using procedurally encod ed case-framewo rks for the verbs was develoned to interpret the input. Phrase-keyword means that the parser is $way;r;;z;ng to build up one of seven digitalt$gE?gn wor P that are common in the d. identify It uz;; the keyworcl; tot:::: phrases. the beginnings endings After a sentence has been Darsed into Dhrases. the procedure associated w!ith the verb' in the sentence is verifies applied. This processing first that the sentence is acceptable (that the phrase semantically for the verb types Then those are legitimate currently in circuit objects not t h g circuit model are introduced into the model, and the verb's manipulations of the model are processed. The parser uses 5 directives to manipulate the model. These directives either add new information to the model or interrelate existing parts of the model. The directives are: (a> (b) cc> (d) (e> 3.3. _ To Specify a function - a function (e.g., clock, counter) is introduced into the model. Assi n a value ~0 a function's.aspect -. one i?i o,fvt,l;efunction s characteristics 1s assigned Define i conceptual si assumed to be a port nal - a global name is o B either a known or unknown function. Define the source of a conceptual signal - the source function of a global signal is identified. Make a connection - identify an information path between two functions and a condition gating the information flow. Example - -- Specification Acquisition illustrate the specification acquistion g* hase, the five sentences shown above for the isplay Character Counter are reduced to their effects on the model. The effects for each sentence are preceded by a letter which references the corresponding directive type from the above list of directives. 1. THE DISPLAY CHARACTER COUNTER (DCC) COUNTS FROM . . ?aT"Iz?k%uce a COUNTER function named DISPLAY CHARACTER COUNTER (DCC). (b) Assi n a value to the Aspect of (b 3519). COUNT-SEQUENCE 2. WHEN THE COUNT OF THE SLC EOUALS 4. THE COUNT OF THE DCC CAPTURED-IN THE LOAD OF zTHE DCC-BASE REGISTER t (DCCBR) WHEN THE HORIZONTAL BLANK HBLANK) BEGINS. c) Define a conceptual signal named HORIZONTAL ?IANK (HBLANK). 3. 4. 5. (e) Make a connection between (DCCBR LOAD) and (DCC COUNT) under the condition (AND (HIGH SLC DS~) (RISING UNKNOWN HBLANK)). THE DCC CAPTURED-FROM THE DCC-BASE REGISTER DCCBR WHEN HORIZONTAL BLANK (HBLANK) BEGINS. [e) EPIC; cpon;;ti;;de:etween (DCC LOAD) and the condition (RISING UNKNOWN HBLANK). EACH-TIME THE COUNT OF THE PIXEL COUNTER (PC) . _ EQUALS 5 ENDS, THE DCC INCREMENTS. (a) Introduce COUNTER function named PIXEL COUNTER (PC). (e) Make a connection between (DCC COUNT-UP) and T under the condition (FALLING PC DSS). THE COUNT OF THE DCC CAPTURED-IN THE ADDRESS OF THE DISPLAY MEMORY (DM). (a) Introduce MEMORY function named DISPLAY . ~ MEMORY (DM). (e) Make a connection between (DM ADDRESS) and (DCC COUNT). The model of the circuit after just these five sentences are processed is shown in Fig. 3. After (I I - HBLANK Fig. 3. State of the Model - (1, 2, 3, 4, and 5) this description, the specified and its DCC has been completely the next section. implementation is discussed in 3.4. Function Each function type has an associated frame structure that provides the prototypical knowledge about that function. There are two corn onents the function frame: the ASPECT and the 8 to ONPT. The ASPECT identifies the im ortant characteristics of the function which mig R t in the input scenerio. be mentioned by the user The CONPT identifies the ports (the input and output lines) that are associated with the function. The width of the CONPT is also identified aseeither 1 or the value of one of the ASPECTS of function. When a particoulfar function is introduced into the model, a co Y x the protot pe Eg. instantiated wit K is copied intFhkhe database the name of function. 4 is the instantiation of the DCC counter as represented in the model after sentences 1, 2, 3, 4, and 5 have been processed. The asterisked items are those processing. that were modified during the sentence The ex ert component E has a conce tual K view einer described. T e function of the frame with the- associated filled-;;easpect val;:; re,pL;ese:;," that conceptual view component. - exnert I 4. Designing a Circuit for a Function ---_ There are two phases involved ' the implementation of a function. First a method for implementing the function must be selected. Then the selected method must be processed to produce the chip-level design. 284 (COMPONENT DCC SADD COUNTER) * ‘TYPE DCC DEVICE (DISPLAY CHARACTER COUNTER)) * DESC-VAL WIDTH DCC 10-2) DESC-VAL COUNT-SE UENCE DCC (0 3519)) ASPECT WIDTH DCC 8 IL) ASPECT OUTPUT-CODING DCC NIL) ASPECT COUNT-SEQUENCE DCC NIL) ASPECT COUNT-DIRECTION DCC NIL) ASPECT DISTINGUISHED-STATES DCC NIL) ASPECT LOADABLE DCC NIL) ASPECT RESETABLE DCC NIL CONPT LOAD DCC WIDTH Gl 0)) 4 CONPT READ DCC WIDTH G265 G171)) CONPT COUNT-UP DCC 1 1 CONPT COUNT-DOWN DCC 1 NIL CONPT LOAD-LINE DCC 1 NIL) CONPT CARRY DCC 1 NIL) where IO-2 G150 G265 = 12 - (CONNECT 1 ~~~~~~ t - (CONNECT (DC t % (DM c Mom) (DCCBR C COUNT-UP)(:P CBR LOAD) (DCC ADDRESS) (Dd! READ) UNKNOWN HBLANK) 'D,g;TpC DSS)) IGH COMB1 OUTPUT COUNT) T) Fig. 4. DCC Counter Function - (1, 2, 3, 4, and 5) 4.1. Implementation Method The first ste F in the selection E recess is to deduce values or with the many of the ASP CTs associated function and to verify that there are no inconsistencies ' deductions made flZr t'k ASPECT values DCC are that it The is loadable, and that the count resetable, has a binary out ut, direction * K Te set of implementation methods aiiocii!ld with the type of function is then considered. An implementation method has 4 components: Prere uisites, Fig. 2 Eliminators, Traits, and Procedure. illustrates two im lementation methods for a counter, one based on a 7 e 161 loadable counter and the other on a 7493 counter. The Prerequisites are (STRATEGY BIN7493 COUNTER (PREREQUISITES I COUNT-DIRECTION UP OUTPUT-CODING BIN) (ELIMINATORS ) I (LOADABLE YES) mms (RESETABLE YES)) > PROCEDURE $BIN7493) Fig. 5. Example Counter Implementation Methods a list of ASPECTS and their values that must be valid for the function in order to select the method. In the BIN74161 function being im lementation.method, the implemente 8 must output coding that only counts up. re ;i;e a bi~~;~ 9 1s true for the BIN7493 implementation method. The Eliminators are a list of ASPECTS and their values which if matched by the function under im lementation, P eliminate that method from being se ected to implement the function. a list of ports that can be explicitly available if the method is selected. The Procedure is the recipe used to build the circuit. It is re resented procedurally in LISP code and is !i re erenced by name in Fig. 5. only Each method that can be used to implement a function is processed in the selection ph;s;liz checking the Prerequisites and Eliminators. of all acceptable methods is selection. compiled during the If there is only one acceptable method it is used to implement the function. If more than one method is acceDtable then the Traits of all candidates are used method. to find the most appropriate If this does not narrow the list to a single method then one user chooses the method. randomly chosen or the this examnle for the DCC assuming that there are I 5 "ZZilk%e two implementation methods from Fig. im lement 51 a counter, the BIN74161 method is tlt$ on y acceptable method and is selected. The procedure associated with the selected method is then run. This procedure consults the frame associated with the function and the connections information involving the function and from this constructs a real diiital interconnecting- circuit, introducin and collective y implement the function. 7 chips that 4.2. Implemented Circuit After the first levels of two phases, there are three the DCC design. is the DCC function and the A:e:;e,e;op level there connections. At the bottom level are the chins and wires used to im lement in erfaces f %,e Dee* The -intermediate level other two levels bv Drovidine. interconnections between chip pins or wire's and th: connection points on the function. Hence the hierarchy of the components is maintained and. if necessary, any circuit can be altered without much effect on the fragment rest of the circuit. The im lementation BIN7416P for the DCC as designed b the implementation method is illustrate in i Fig. 6. DCCBR READ DM ADDRESS DCCBR LOAD vcc Fig. 6. DCC Implementation 5. Conclusion the SADD is a general purpose design system based on ideas of structured, an interactive modular circuit design via user interface. The current 41 sentences in the innut scenerio have been run successfully thro 3 h the parser creating a database of approximately 6 0 entries. counters in the The design of the 3 Screensplitter are czmpleted and circuits functional1 used in the P equivalent to those Screensp ac tuaktz itter have been designed. design of a symbolic simulator is currently in progress. This simulator will allow the desi ner to test and debug the circuits and will camp ete f the design S&i?%?1 envisioned for SADD. When completed, extensible E rovide a generaioFurpose and design. knowledge- ased system digital The Traits are a list of ASPECTS and their values that are true if the method is selected and 285

1980

35

29

THEORY DIRECTED READING DIAGNOSIS RESEARCH USING COMPUTER SIMULATION Christian C. Wagner John F. Vinsonhaler The Institute for Research on Teaching Michigan State University East Lansing, Michigan 48823 Abstract A five year project studying the diagnosis and remediation of children with reading problems is described. The discussion includes observational studies of reading diagnosticians at work, observations of diagnostician training programs and computer simulation of theories about decision making in reading diagnosis. The results of the observational studies are mentioned and the theories and systems for computer simulated diagnosis are described. I Introduction The Institute for Research on Teaching is a federally funded project whose purpose is to investigate teaching, where teaching is conceptualized as an information processing task. The Clinical Studies project of the Institute is just finishing its first five year plan studying teachers who diagnose and remediate children with reading problems. This paper should serve as a short introduction of this work to the AI community. The Clinical Studies project has been primarily concerned with understanding reading diagnosis and remediation - whether performed by classroom teacher, reading specialist, learning disability specialist or school psychologist. Theories or models have been developed to account for the significant behaviors that occur when one of the teaching professionals works with a child. These theories have been tested against 1)direct observational studies of the specialists working with cases, 2)training studies observing the instruction of new specialists, and 3)computer simulation studies observing the behaviors implied by the theory through simulation. Results of these studies have shown that an individual's decision making may be very unreliable - suggesting that individual behavior may not warrant simulation. Before turning to the computer simulation studies, consider briefly the theory and results of the other studies. II Theories and Models The content-independent theory that attempts to account for the problem solving behavior of the clinicians is termed the Inquiry Theory Cll. The reading clinician-child interaction is viewed by this theory as follows: 1)The case is considered to be an information processing system that must perform a certain set of tasks. Some information processing abilities are critical to the adequate performance of these tasks. An adequate diagnosis of a problematic case must contain a statement relevant to each such critical ability. There may exist prerequisites for these critical abilities. An adequate diagnosis will include a statement relevant to all prerequisites for each deficient critical ability. Finally, a good diagnosis will include a statement of the history of the case that has led to current deficiencies. 2)The clinician must diagnose a case as described above. This is accomplished by the application of elementary information processing tasks to an understanding of how a case performs its function. The elementary tasks include ones such as hypothesis generation, hypothesis testing, cue collection, cue interpretation, etcetera [21. 286 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. processing task, the MORAL describes other significant aspects of behavior required such as the allocation of information processing capacity and conditions of information overload. At the current time the MORAL describes the critical abilities for human reading. It further details what other factors might effect these abilities and cause deficiences and how correction of any deficiency might be attempted. To date it appears to be quite effective in determining the reading problems in a case. The MORAL does not at the current time make any attempt to describe how reading comprehension takes place. Instead the various types of reading or listening comprehension tasks which people must perform to be good readers are listed. Of course the MORAL is at best incomplete and possibly incorrect. But it does serve to direct research and diagnose cases of reading difficulty. One final note with respect to theory - it is the intersection of these two relatively independent models (a content independent model of clinical problem solving and a content dependent model of the reading process) that is our area of concern. The questions that arise include: how do the specialists diagnose and treat children, how should they do it, how can they be trained to do what they should, what impact does the model of reading have on the decision making process, etc. III Observational and ApDlication Studies The observational studies of teaching professionals and clinicians in reading have basically indicated one thing - the specialists are not reliable. Careful observation in seven interconnected studies of these professionals diagnosing a simulated case of reading difficulty has shown correlations generally not significantly different from zero, whether from the same clinician diagnosing the same case twice or different clinicians each diagnosing the same case once. This finding holds for professionals selected by their peers as the best; for classroom teachers, reading specialists, and learning disability specialists. After eliminating most counter explanations of low reliability through carefully designed replication studies, the reliabilities are still of borderline significance from zero (e.g., 0.10 - 0.15). c41 Training studies have indicated more optimistic possibilities - in thirty hours of instruction, the correlations may be raised from 0.1 to 0.4 - a value close to physician reliability. Close observation of the training process and its transfer to work settings will hopefully uncover means by which reliability may be raised further. IV Computer Simulation Studies With this backdrop, consider the contribution of computer simulation to this program of research. For this discussion we will ignore the use of case simulation which has been so vital for stimulus control in our experimental design and turn ins tead to clinician simulation theories described earlier. on the All of our simulations of reading specialists have been simulations based on the Inquiry Theory and the Model of Reading and Learning to read described earlier. In this way, the results of each simulation can be used to expand and refine a model that directs our reseach efforts in simulation, training and observation. All studies described here were run on an interpretive procedural language whose primitives were based on the Inquiry Theory. This system is entitled BMIS - the Basic Management Information System. Effectively, a system subroutine was created for each elementary information processing task described by the Inquiry Theory (e.g., hypothesis generation, cue collection, diagnostic determination, etcetera). Each subroutine could be called up by a command in an interpretive language. An initial hypothesis directed program was set up in which the hypotheses generated about a case direct the collection of information about a case and information interpretation, which generated more hypotheses, and so on. On the basis of any decision that was made (accept hypothesis X as part of the diagnosis, reject hypothese Y, etc.) sub procedures might be activated to handle the peculiarities of the particular decision. This system was designed for theory investigation and was not intended to be easy to use or flashy. Furthermore, there were many restrictions on its input to bypass the natural language communication problems. As time permitted, a new system was created to rectify these and other identified shortcomings of BMIS. The new production oriented system with similar types of primitives is entitled MOSES, the Multiple Object Simulated Encounter System. Both systems are available through TELENET on the Wayne State University Amdahl 470~6. *The SIMCLIN Modeling Study: The first simulation study was basically a modeling study. Given the framework provided by the Inquiry Theory, memory structures were created by systematic interview with a senior reading clinician. Such things as hypotheses, cues, observations, diagnoses, strategies, etc. were defined. The goal was the creation of a simulation that would closely emulate this specialist's problem solving behavior. Comparisons were drawn between the human specialist and the computer analog as they diagnosed the simulated cases of reading difficulties mentioned earlier. The results indicated that the simulation was a very effective model in terms of all measures used - the number and order of cues collected, the diagnosis and suggested remedial plan, etc. *The Pilot SIMCLIN Reliability Study: With the very low human clinician reliability, it became clear that modeling of individual people was a pointless procedure. Instead we directed our efforts to the simulation of behavior of 287 groups of clinicians; i.e., to the simulation of models of diagnosis agreed upon by clinicians. At this point, then, the emphasis turned to the creation of intelligence that would be reliable and valid with respect to group reading diagnosis and still be teachable to unaided human specialists. It was at this point that the development of the Model of Reading and Learning was begun - it would serve to define the content of clinician memory. This study examined the reliability of a computer diagnostic system that was based on the Inquiry Theory and the newly developed MORAL. The simulated clinician (SIMCLIN) was set up and asked to diagnose four simulated cases twice (no SIMCLIN memory of previous runs was allowed but different initial contact settings were used). These diagnoses were compared with respect to reliability with the diagnoses of human clinicians. The results were that the SIMCLIN had a reliability of 0.65 compared to human reliability of 0.10. Further, commonality scores - which indicate how an individual agrees with a criterion group diagnosis - indicated that the SIMCLIN included 80% of the categories agreed upon by the group of human clinicians while the mean for individual human clinicians was 50%. *The Pilot SIMCLIN Validity Study: Finally, a simulation study has been run to get a first measure of the validity of the SIMCLIN's diagnostic decisions when those decisions are directed by the Inquiry Theory and MORAL. Reading case records were taken from Michigan State University's reading clinic for SIMCLIN workup. Records were selected which indicated correct diagnosis and others that indicated poor diagnosis (as measured by the child's response to treatment). The areas of concern were the adequacy of the SIMCLIN as an embodiment of the theories, the reliability of the SIMCLIN diagnosis and the validity of the SIMCLIN diagnosis. It was hoped that the SIMCLIN would agree closely with the clinic's diagnosis for the correctly diagnosed case and not as closely for the poorly diagnosed one. The SIMCLIN did, in fact, behave as dictated by the MORAL - the simulation checked out the critical abilities of reading and the prerequisite factors and past history of those that were problematic. The reliability of the diagnostic decisions was essentially 1. Adherence to the MORAL almost guarentees this. With respect to the SIMCLIN diagnosis on the well and poorly diagnosed cases, the results were equivocal. The reason for this is that data required by the SIMCLIN was not present in the clinic files. Such things as classroom observation of engaged academic time, listening comprehension scores, and change scores over time were not available. In fact, indications are that these types of data are not routinely collected by reading clinicians, although the SIMCLIN considers them significant. The model and its simulation might well demonstrate inadequacies in the state of the art in reading diagnosis. V Conclusion In conclusion, the research paradigm described here has been quite effective. Models and theories direct and focus research designs. These designs - whether observational, training or simulation - reflect back to expand and refine the theories. Substantial data has shown that an individual's decisions may be very unreliable. Training in decision making models and content area theories can improve the reliability. But the key to effective problem solving seems to be the validity of the theories that are used to direct decision making. One effective means for examining the validity of such theories is through computer simulation. The next step will be the completion of a production oriented SIMCLIN that will be used as a preceptor during instruction of student clinicians and a decision aid by reading specialists in schools. The validity of the MORAL SIMCLIN will be checked by following its recommendations and watching the results for real children. The research will continue to be theory oriented. Further information on many aspects of this research program may be obtained by contacting the Institute for Research on Teaching at Michigan State University. References Cl1 c21 c31 c41 Wagner, C.C. and J.F. Vinsonhaler "The Inquiry Theory of Clinical Problem Solving: 1980", The Institute for Research on Teaching, Michigan State University, 1980. Elstein, A.S., L.S. Shulman, S. Sprafka, H. Jason, N. Kagan, L.K. Akkak, M.J. Gordon, M.J. Loupe, and R.D. Jordon. Medicd Problem Solving: An Analvsis of Clinical Reasoning Press, l&8. Cambridge: Harvard University Sherman, G.B., J.F. Vinsonhaler, and A.B. Weinshank "A Model of Reading and Learning", The Institute for Research on Teaching, Michigan State University, 1979. Vinsonhaler, J.F. "The Consistency of Reading Diagnosis", The Institute for Research University, 51:80. Teaching, Michigan State 288

1980

36

30

A WORD-FINDING ALGORITHM WITH A DYNAMIC LEXICAL- SEMANTIC MEMORY FOR PATIENTS WITH ANOMIA USING A SPEECH PROSTHESIS Kenneth Mark Colby, Daniel Christinaz, Santiago Graham, Roger C. Parkison The Neuropsychiatric Institute/UCLA 760 Westwood Plaza Los Angeles, California 90024 ABSTRACT Word-finding problems (anomia) are common in brain-damaged patients suffering from various types of aphasia. An algorithm is described which finds words for patients using a portable micro- processor-based speech prosthesis. The data- structures consist of a lexical-semantic memory which becomes reorganized over time depending on usage. The algorithm finds words based on partial information about them which is input by the user. I WORD RETRIEVAL PROBLEMS We are developing an "intelligent" speech prosthesis (ISP) for people with speech impairments Cll. An ISP consists of a small, portable computer programmed to serve a number of functions and interfaced with a speech synthesizer. ISPs can be operated by pressing keys on a keyboard or by eye- movements using a specially-designed pair of spectacles. How words are stored, organized, and re- trieved from human lexical memories constitutes a lively area of research in current psychology, computational linguistics, neurology, aphasiology and other cognitive sciences [2], [3], [4], [5], C61, 1171, 181. Words in a lexical memory can be associated to other words by means of several relations - synonymy, antonymy, part-whole, spatio-temporal contiguity, etc. [33. It can also be assumed that the process of word-finding for production begins with semantic concepts to which word-signs are connected. Once the word-represen- tation for the concept is found, it is subjected to phonological rules if it is to be spoken, and to graphological rules, if it is to be written. In the final stage of output, articulatory rules governing muscle movements for speech are utilized or rules for hand and finger movements for writing are applied. Impairment in word expression can be due to failures at any stage of this process from concept to utterance. Our interest here is in those instances in which the speaker has the word, or part of the word, or some information about the word, in consciousness but cannot produce the target word. Our efforts have been directed -_------ This research was supported by Grant #MCS78-09900 from the Intelligent Systems Program of the Mathematics and Computer Science Division of the National Science Foundation. towards writing and field-testing a computer program which can find words in a lexical-semantic memory, given some working information about the target word. It has long been known that some aphasic patients, who cannot produce a word, can indicate how many syllables the word contains by finger tapping or squeezing the examiner's hand [9]. Both Barton [lo] and Goodglass et. al. Lll] reported that aphasics know some generic proper- ties of the target word such as its first letter, last letter, etc. Our own experience with patients having anemic problems have confirmed and extended these observations. The word-expression disorders in which we are interested are commonly divided into two groups which are (weakly) correlated with the locations of brain lesions and accompanying signs and symptoms. The first group consists of patients with lesions in the anterior portion of dominant cerebral hemisphere. These patients have many concrete and picturable words in consciousness, and they perform well, although slowly, on naming tasks [12], [13]. The naming disruption is part of a more generalized disturbance which includes laborious articulation and often a disruption of productive syntax. The second group of disorders in this classification scheme involves lesions in the posterior region of the dominant hemisphere. Although these patients often fail to provide the correct name of an object, their substituted response can be related to the target word as described above in the studies of Rinnert and Whitaker [6]. Since the substitute word is so often systematically related to the target word, we thought it might be usable as a clue or pointer in a lexical search for the target word. II A WORD-FINDING ALGORITHM --- The first step involves constructing the data base, a lexicon of words stored on linked lists at various levels. The highest level is a "Topic Area" (TA), representing the idea being talked ahout. Currently we use 15 topic areas. The TA consists of list of words, each connected in a large lexicon of words, falling within that topic area. Each word within a particular topic area is also linked to a list of selected words, as in the 289 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. following example: Assuming that each slot in the pattern is (BODY (ACHE (HEAD HURT PAIN ( ANKLE (FOOT LEG) ) STOMACH) ) (WOUND (BLOOD CUT HURT ) ) ) The organization of the word lists changes dynamically over time according to their usage, as will be described below. In attempting to find a word, the program first asks the user to identify the topic-area from a list presented to him on a display. We start with an initial set of word-lists but change them in accordance with the individual user's environment. A topic-area specifies where the user is linguistically in the discourse, not where he is physically or socio-psychologically. He is then asked about certain properties of the target word, the questions appearing on the display as follows: (1) What is the topic area? (2) What is the first letter of the word? (3) What is the last letter of the word? (4) What letters are in the middle of the word? (5) What word does this word go with? Question (5) attempts to obtain a word associated to the target word however idiosyncratic it might be. It may even resemble the target word in sound. Our starting lexical memory of about 1800 words was based on known high-frequency discourse words, associations from Deese [81, word associa- tion norms cl41 and word-pairs from Rinnert and Whitaker [6] . Word-associations can be highly idiosyncratic to the individual. For example, if one asks 1,000 college students which word they associate to the word table, 691 says chair but the remainder of the responses are distributed over 32 other words including jewels like big, cards, and tennis [151. Hence, with each user, we add his particular associations and words that do not appear in our starting lexicon. This data is collected by a spouse, friend, research assistant or speech pathologist in conversations with the user. After the user has offered clues about the target word, we can represent the responses using a clue pattern of the form: CLUE = TA + L + L + STRING + GWW -t;rh-= TA = Topic area L = Letter STRING = One or more letters GWW = GOESWITH word correct and not null, the program first finds a short list of numbered candidate target words (zero to a maximum of 20 words) meeting the input criteria of the clue pattern. The search will be illustrated by cases. Case (1). Suppose the target word were steak. The clue pattern might be: CLUE = FOOD + S + K + A + meat. The word meat is looked up on the FOOD list and then a search is made on the list of words linked with the word meat which begin with S and end with K and have an A in between. If steak were on the list of words linked to meat, it would be display- ed visually and auditorily as: 1. STEAK When he sees or hears the target the user enters its number (in this case "1.") and the word is then inserted in the utterance of the ISP. (Some- times, on seeing or hearing the target, the user can utter the word himself.) In this first illustration of the program's operations, we have assumed full and correct entries have been made in the clue pattern and that steak was on the list of words which [GOESWITH] meat. But suppose steak is not on the meat list of FOOD. Case (2). If steak were not found on the meat list of FOOD, the "meat" part of the clue is ignored and all words under FOOD beginning with (S) , ending in (K), and with an (A) in between are retrieved. The word steak might appear in this group and if so, the program will add it to the meat list after the user signifies this is his target word. Thus the lexical memory becomes automatically reorganized over time. Case (3). If still no acceptable word is retrieved, the "FOOD" part of the clue is ignored and a search is made on all topic-area lists for (S + K + A) words. The word steak might appear in the ANIMAL topic-area associated with the word cow. After the user indicates this is the target word, steak is added to the meat list under FOOD. With repeated usage, steak becomes promoted to the top of the meat list. Case (4). If the target-word is still not retrieved bv an exhaustive search of all topic- areas, the word does not exist in the lexicon and the program ends. One might consider varying the constraints of (S + K + A) clues and searching further but in our experience this is rarely productive. Time is not a problem for the program since an exhaustive search requires only a few seconds. But a large number of candidate words are retrieved when most of the clues are ignored. And it is too time-consuming for the user to search through the list of candidates looking for the desired word. Case (5). It might be that the user cannot 290 answer completely the questions required by the clue pattern or the entries may be in error. For example, he may not know the first and last letters of steak, but he does know the topic-area is FOOD and the [GOESWITH] word is meat. If steak is on the meat list, the search will succeed. If not, there is no point to displaying all the words under FOOD because the candidate-set is too large. In our experience thus far we have found that at least 2-3 pieces of information are necessary to retrieve a target word. Further experience may indicate that some users can benefit from variations in the clue pattern. For example, with some patients, we have found it expedient to ask them first for the topic-area and the [GOESWITH] word. If this fails, the program then asks the letter questions. With other patients questions regard- ing word-size, number of syllables, and "what do you use it for?" may be helpful. We are currently field testing the program with a variety of anemic patients using keyboard or ocular control to study the value of utilizing different clue patterns.* In the meantime, others in the field of communica- tion disorders may wish to utilize and improve on the word-finding algorithm reported here. Such a program can also be used by speech pathologists as a therapeutic aid. The patient can be given linguistic tasks to practice on using the ISP at home. Repeated exercise and home- practice may facilitate the finding functions. patient's own III SUMMARY We have described a computer program for an intelligent speech prosthesis which can find words in a lexical-semantic memory given some informa- tion about the target words. The program is currently being tested with a variety of anemic patients who have word-finding difficulties. A future report will describe the results of this field-testing. REFERENCES [ll Colby, K.M., Christinaz, D., Graham, S. "A Computer-Driven Personal, Portable and Intelligent Speech Prosthesis." Computers and Biomedical Research 11 (1978) 337-343. [2] Caramazza, A., Berndt, R.S. "Semantic and Syntactic Processes in Aphasia: A Review of the Literature." Psychological Bulletin 85 (1978) 898-918. -------- *We are grateful to Carol Karp and patients of the Speech Pathology Division of Northridge Hospital Foundation, Northridge, California (Pam Schiffmacher, Director) for their collaboration. c31 c41 c51 [61 c71 c81 c91 Cl01 Cl11 Cl21 Cl31 Cl41 Cl51 Miller, G.A., Johnson-Laird, P.N. Language and Perception. Cambridge: Harvard University Press, 1976. Goodglass, Naming and Aphasia." 374. H ., Baker, E. "Semantic Field, Auditory Comprehension in Brain and Language 3 (1976) 359- Zurif, E., Caramazza, A., Myerson, R., Galvin, J. "Semantic Feature Representation for Normal and Aphasic Language. Brain and Language 1 (1974) 167-187. Rinnert, C , Whitaker, H.A. "Semantic Confusions by Aphasic Patients." Cortex, 9 (1973) 56-81. Geschwind, N. "The Varieties of Naming Errors." Cortex 3 (1967) 97-112. Deese, J. The Structure of Associations in Language and Thought. Baltimore: Johns Hopkins Press, 1965. Lichtheim, L. "On Aphasia." Brain 7 (1885) 433-484. Barton, M. "Recall of Generic Proper Words in Aphasic Patients." Cortex 7 73-82. ties of (1971) Goodglass, H., Kaplan, E., Weintraub, S., Ackerman, N. "The "Tip-of-the-Tongue" Phenomenon in Aphasia." Cortex 12 (1976) 145-153. Wepman, J.M., Bock, R.D., Jones, L.V., Van Pelt, D. "Psycholinguistic Study of the Concept of Anomia." Journal of Speech and Hearing Disorders 21 (1956) 468-477. Marshall, J.C., Newcome, F. "Syntactic and Semantic Errors in Paralexia." Neuro- psychologia 4 (1966) 169-176. Postman, L., Keppel, G. (Eds.). Norms of Word Association. New York: Academic Press, 1970. Palermo, D.S., Jenkins, J. Word Association Norms. Minneapolis: University of Minnesota Press, 1964. 291

1980

37

31

TROUBLE-SHOOTING BY PLAUSIBLE INFERENCE * Leonard Friedman Jet Propulsion Laboratory, California Institute of Technology Pasadena, CA 91103 ABSTRACT The PI system has been implemented with the ability to reason in both directions. This is combined with truth maintenance, dependency directed backtracking, and time-varying contexts to permit modelling dynamic situations. Credibility is propagated in a semantic network, and the belief transfer factors can be modified by the system, unlike previous systems for inexact reasoning. I TROUBLE-SHOOTING LOGIC The PI (for Plausible Inference) system enables a user to trouble-shoot physical systems in a very general way. The trouble-shooting process requires that our user first define a physical model which represents what takes place in normal operation with everything functioning correctly. If this physical model is translatable to the representation used by PI, it can be stored in computer memory and used to guide the search for the most likely failure. -In order to make the process clearer, we shall describe a few of the many methods of reasoning employed by human beings in their trouble-shooting. These methods are the ones we can at present imitate in the PI system. Suppose we have a desired goal state defined in our physical model, and this state depends on three conditions being true to attain the goal. If we execute the process and observe that the goal state was not attained, we conclude that at least one of the three conditions on which it depended must have been false, and all are possibly false. If we then perform some test designed to verify whether one of the conditions is actually true and the test shows that it is indeed true, we conclude that at least one of the remaining two untested conditions must be false. If all but one of the remaining conditions has been eliminated from consideration by further testing, we may conclude that the single condition remaining must be the guilty party. The process of elimination just described is the one normally employed by humans, and it is this process we have implemented on the computer. Of course, the three conditions may in turn have conditions on which they depend. In that case the method just described may be applied *This paper presents the results of one phase of research performed at the Jet Propulsion Laboratory, California Institute of Technology, sponsored by the National Aeronautics and Space Administration under Contract NAS 7-100. recursively to narrow the fault down further, at least to the granularity of the conditions employed in the representation. This method fails if there are conditions on which the goal state depends for realization and which are not explicitly represented in the model. Nevertheless, the exercise may serve as a valuable guide to help a user to focus attention on specific, possibly false, areas as likely sources of failure. Another difficulty with the method is the fact that either a test does not exist to determine whether a specific sub-goal was reached, or the sub-goal state in question was changed by a later event occurring in the model. In this case it is difficult to verify whether the changed sub-goal state was ever achieved. Only if there were long-lasting side-effects will it be possible to verify. Such difficulties plague human trouble-shooters as well. The present implementation can not reason about such "vanished" events, in a hypothetical mode, from a past context. II IMPLEMENTATION AND THEORY The PI system is part of a larger system called ERIS, named for the Greek goddess of argument. The basic module of the system, described in [ll, performs deduction and modeling in the Propositional Calculus. A planning module has been built by M. Creeger by augmenting the basic deduction module with many special features, including a "causal" connective that supplements the standard logical connectives (AND, OR, etc.). Similarly, the PI module has been built by augmenting Propositional Calculus with extra rules of inference, and another belief besides truth-value associated with each assertion. This additional belief, which we call credibility, is a subjective numerical measure of the confidence in the truth-value, with values between -1 and 1. The basic ERIS module generates a network of nodes linked by connectives as it reads in its knowledge base of assertions; this feature is retained in the other modules. The techniques used in ERIS make it possible to perform deduction without rewriting. Instead, "specialists" for each connective propagate the correct values of the beliefs to the assertions which they link. A theoretical foundation for this approach, aw lying 292 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. both to Propositional Calculus and First Order Predicate Calculus, is given by the method of Analytic Tableaux" [Zl. Because rewriting is totally avoided, inference, planning, model-revision, and dependency-directed backtracking can be performed in a single integrated system. Al(Q( PI 2) I :T; +.5 :F; -.5 Arcs link antecedent/consequent in implication Fig. 1 Transfer Factors in a Semantic Net Plausible Inference introduces two rules of inference besides Modus Ponens and Modus Tollens, called the Method of Confirmation and the Method of Denial. These extra modes permit the propagation of truth values, true or false, even in directions forbidden bY Propositional Calculus. Simultaneously, the associated credibilities are propagated through the net, employing all four modes as appropriate. A calculus of Credibility transfer between arbitrary logical expressions has been worked out to specify exactly the process of Credibility propagation through the net. The calculus is described in [31, and is based on equations employed in MYCIN [41. The basic quantities controlling propagation between antecedents and consequents in implication are transfer factors or DELTAl's, and there are four for each antecedent/consequent pair, one for each mode. (See Fig. 1) Both MYCIN and PROSPECTOR are limited to a single reasoning mode and transfer factor for each implication and use a static transfer factor structure that is specified by human "experts". The PI system, in trouble shooting, recalculates the appropriate transfer factors on the basis of incoming evidence. In addition to a dynamic transfer factor structure, PI also incorporates the use of default values for the transfer factors in trouble-shooting when the users do not have better information. III APPLICATION EXAMPLE Our example of automated trouble-shooting uses a toy case selected from the application domain, the mission control of spacecraft. It is a simplified representation of signal transmission from a space craft to Earth. The desired goal state is "Signal Received" from the SIC. Fig 2 shows a plan to accomplish this which has been generated by the ERIS planner. The user supplies three basic "action CAUSES state" relations: (1) "Ground Antenna Receiving" CAUSES "Signal Received", (2) "Point SIC" CAUSES "Pointed Correctly", and (3) "Transmit Signal" CAUSES "Signal Transmitted". The pre-conditions for the first relation are the end states achieved in the GROUND ANTENNA RECEIVING S/C POINTED CORRECTLY ) -“.“yT) -0.001 TRANSMIT SIGNAL RSS POWER ON -0.001 SIGNAL RECEIVED (3 77 CORRECT DATA MODE SIGNAL TRANSMITTED Fig. 2 Plan for Command Sequence Generation 293 second and third. In effect we are saying, "In order for an action to achieve the goal state certain pre-conditions must be true. In order to make those pre-conditions true certain actions will cause them which may also have pre-conditions to succeed". Thus complex sequences may be built up. The ERIS planner links the basic strategies in the manner shown, using the pre-conditions as hooks. It then collects the actions in a list, and supplies that list as the desired command sequence. The plan generated in this way is a descriptive model of the signal transmission process and constitutes our trouble-shooting knowledge base. The propagation of beliefs that takes place with the CAUSES connective is identical to the belief propagations of an implication as defined in [3], although the timing of belief propagation of CAUSES are different. We define the belief propagation-equivalent implication form of the CAUSES relation as (IMPLIES (AND action precondition1 . . preconditionN) goal-state). At the start, the assumption is made that all states are true. Suppose that the sequence is executed and the ground station fails to receive the signal. Then "Signal Received" is false, and this can be entered into the data base. The effects of this change of belief are propagated through the data base by a modified Modus Tollens, making all the events on which "Signal Received" depends Possibly-False (or PF). If a test is then performed by the human controllers like causing the s/c to roll while transmitting, and a signal is received during the roll, we may conclude that the action "Point space craft" worked, "Signal Transmitted" is true, and "Ground Antenna Receiving" has been verified. Inputting these facts into the data base causes the PI system to do two things: (1) Those preconditions required by "Signal Transmitted" are changed from PF to T by Modus Ponens. (Possibly False to True) (2) "Pointed Correctly" is changed to False, F, rather than PF. In addition, the PI system raises the credibility of failure for those events on which "Pointed Correctly" depends. Their truth value remains Possibly False because there are multiple possibilities. If one of these latter events is shown to be true by testing, the remaining one may be the only possibility left. For example, if the Sun Sensor and Canopus Sensor can be shown to work, and their truth status is input, the system will conclude that the Starmap must be at fault, even though the a priori credibility of such a mistake was extremely low. How the credibilities change at various stages of operation can be described now. At the start there are two possibilities: either a priori credibilities may be entered or default credibilities generated. Figure 2 shows a priori credibilities entered on the branching lines. These are subjective measures of the likelihood of failure of the respective events to which they lead given that the state which depends on them is false. Thus, for "Signal Received", "Pointed Correctly" has an a priori credibility associated with false of -.7, "Signal Transmitted" a value of -.3 and "Point space craft" a value of -.OOl . When we start, the assumption is that every state is true (at the appropriate time) with a credibility of 1.0 (equivalent to certainty). At the next stage, when all we know is that "Signal Received" is false, the a priori credibilities are assigned to all the states. If we used the default mode, credibility would be assigned equally; i. e., for two events, each would get .5, for three .33, for four .25, etc. Whenever an event is eliminated as true, the remaining credibilities are raised by an empirical formula that reflects a reasonable sharing of suspicion, based either on the a priori splits or an equal partition. Thus, in our example, "Starmap Accurate" went from true (cred 1.0) to Possibly False (cred -.007) to Possibly False (cred -.Ol> to False (cred -1.0). This, in a simplified way, describes the operation of PI in trouble-shooting using reasoning by Plausible Inference. Of course, humans employ many other methods in trouble shooting, such as analogy. For example, a person may say "This problem resembles one I encountered in another area. Maybe it has the same cause I deduced then." By such techniques, humans can often vector in on a problem, bypassing step-by-step elimination. We hope to implement some of these techniques eventually. References 111 Thompson A. M., "Logical Support in a Time-Varying Model", in Lecture Notes in Computer __I_-- Science, Springer-Verlag, Berlin, 1980. Fifth Conf. on Automated Deduction, July 8-11, Les Arcs, Savoie, France. [21 Smullyan R. M, First-Order Logic, Springer-Verlag, Berlin, 1968 [3] Friedman L., "Reasoning by Plausible Inference" in Lecture Notes * 580. Computer Science Springer-Verlag, Berlin, Proc. Fifth Conf. on Automated Deduction, July 8-11, Les Arcs, Savoie, France. [41 Shortliffe E. H. and Buchanan B. G., "A Model of Inexact Reasoning in Medicine" , Math. Biosci. 23, PP. 351-379, 1975. Also chapt. 4 of Computer-Based Medical Consultations: MYCIN, Elsevier, New York, 1976. 294

1980

38

32

AN APPLICATION OF THE PROSPECTOR SYSTEM TO DOE’S NATIONAL URANIUM RESOURCE EVALUATION* John Gaschnig Artificial Intelligence Center SRI International Menlo Park, CA 94025 Abstract A practical criterion for the success of a knowledge-based problem-solving system is its usefulness as a tool to those workilb in its specialized domain of expertise. Here we describe an application of the Prospector consultation system to the task of estimating the favorability of several test regions for occurrence of uranium deposits. This pilot study was conducted for the National Uranium Resource Estimate program of the U.S. Department of Energy. For credibility, the study was preceded by a performance evaluation of the relevant portion of Prospector’s knowledge base, which showed that Prospector’s conclusions agreed very closely with those of the model designer over a broad range of conditions and levels of detail. We comment on characteristics of the Prospector system that are relevant to the issue of inducing geologists to use the system. 1. Introduction This paper describes an evaluation and an application of a knowledgerbased system, the Prospector consultant for mineral exploration. Prospector is a rule-based judgmental reasoning system that evaluates the mineral potential of a site or region with respect to inference network models of specific classes of ore deposits. Here we describe one such model, for a class of “Western statesIr sandstone uranium deposits, and report the results of extensive quantitative tests measuring how faithfully it captures the reasoning of its designer across a set of specific sites (used as case studies in fine-tuning the model), and with respect to the detailed subconclusions of the model as well as its overall conclusions. Having so validated the performance of this model (called RWSSU), we then describe a pilot study performed in conjunction with the National Uranium Resource Evaluation (NURE) program of the U.S. Department of Energy. The pilot study applied the RwSSU model to evaluate and compare five target regions, using input data provided by DOE and USGS geologists (using the medium of a model-specific questionnaire generated by Prospector). The results of the experiment not only rank the test regions, but also measure the sensitivity of the conctusions to more certain or less certain variations in the input data. One interesting facet of this study is that several geologists provided input data independently about each *This research was supported by the U. S. Geological Survey under USGS Contract No. 14-08-0001-l 7227. Any opinions, findings, and conclusion or recommendations expressed in this report are those of the author and do not necessarily reflect the views of the U.S. Geological Survey. test region. Since input data about each region varies among the responding geologists, so do the conclusions; we demonstrate how Prospector is used to identify and resolve the disagreements about input data that are most significantly responsible for differences in the resulting overall conclusions. This paper is a condensation of portions of a larger report [4]. 2. Validation of the Model The practical usefulness of an expert system is limited if those working in its domain of expertise do not or will not use it. Before they will accept and use the system as a working tool, such people (we shall call them the “domain users”) usually expect some evidence that the performance of the system is adequate for their needs (e.g., see [8]). Accordingly, considerable effort has been devoted to evaluating the performance of the Prospector system and of its various models [2, 31. In the present case, we first needed to validate the performance of the uranium model to be used r- the pilot study for the U.S. Department of Energy. The methodology used to evaluate Prospector’s performance is discussed in detail elsewhere [2, 31. For brevity, here we outline a few relevant factors. The Prospector knowledge base contains a distinct inference network model for each of a number of different classes of ore deposits, and a separate performance evaluation is performed for each model. Here we are concerned with one such model, called the regional-scale “Western states” sandstone uranium model (RWSSU), designed by Mr. Ruffin Rackley. Since there exist no objective quantitative measures of the performance of human geologists against which to compare that of Prospector, we instead use a relative comparison of the conclusions of a Prospector model against those of the expert geologist who designed it. To do so, first a number of test regions are chosen, some being exemplars of the model and others having a poor or less good match against the model. For each such case, a questionnaire is completed detailing the observable characteristics that the model requests as inputs for its deliberation. Prospector evaluates each such data set and derives its conclusion for that test case, which is expressed 081 a scale from -5 to 5. As a basis of comparison, we also independently elicit the model designer’s conclusion about each test case, based on the same input data, and expressed on the same -5 to 5 scale. Then we compare Prospector’s predictions against the target values provided by the model designer. Table 1 compares the top-level conclusions of Prospector (using the RWSSU model) against those of the model designer for eight test regions. 295 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Table 1. Comparison of RWSSU model with Designer for Eight Test Cases Test Region Designer’s Prospector Di fference Target Score -------------------------------------------------- Black Hi Ils 3.50 4.33 -0.83 Crooks Gap 4.70 4.26 0.44 Gas Hills 4.90 4.37 0.53 Shirley Basin 4.95 4.13 0.82 Ambros i a Lake 5.00 4.39 0.61 Powder River 4.40 4.40 0.00 Fox Hi I I s 1.50 2.17 -0.67 Oi I Mountain 1.70 3.32 -1.62 ------------------------------------------------- Average: 0.69 Table 1 indicates that the average difference between the Prospector score and the corresponding target value for these eight cases is 0.69, which is 6.9% of the -5 to 5 scale. Besides the overall conclusions reported above, quite detailed information about Prospector’s conclusions was collected for each test case. One feature of the Prospector system is the ability to explain its conclusions at any desired level of detail. In its normal interactive mode, the user can interrogate Prospector’s conclusions by indicating which conclusions or subconclusions he wishes to see more information about. The same sort of information is presented in Table 2 (using the Gas Hills region as an example), in the form of Prospector’s overall evaluation, the major conclusions on which the overall evaluation is based, and the subconclusions that support each major conclusion. For brevity, each section of the RWSSU model represented in Table 2 is identified by its symbolic name, which is indented to show its place in the hierarchy of the model. For comparison, we first elicited from the model designer his target values for each section of the model listed in Table 2; these values are included in Table 2. Table 2. Detailed Comparison of RWSSU Model with Designer for Gas Hills Designer’s Prospector Difference Target Score ------------------------------------------------ RWSSU 4.90 4.37 .53 FTRC 4.80 4.64 .16 TECTON 4.50 4.50 .oo AHR 5.00 4.95 .05 FAVHOST 4.80 5.00 -. 20 S EDTECT 4.80 4.88 -.08 FAVSED 4.90 4.68 .22 FLUVSED 4.90 4.68 .22 KAlti’lESE:) -3.54 -2.03 -1.43 AEOLSED -2.50 -2.10 -.40 FMA 4.95 4.41 .54 RBZONE 5.00 4.60 .40 AIZONE 4.00 4.77 -.77 M I NZONE 5.00 5.00 l oo Average difference = 0.36 (Average of absolute values) The data in Table 2 indicate that Prospector not only reaches essentially the same numerical conclusions as its designer, but does so for similar reasons. This detailed comparison was repeated for each of the eight cases, resulting in 112 distinct comparisons between Prospector’s prediction and designer’s target value (i.e., 8 test regions times 14 sections of the model). The average difference between Prospector’s score and designer’s target value over these 112 cases was 0.70, or 7.0% of our standard lo-point scale.** Gaschnig [4] also reports sensitivity analysis experiments showing the models to be rather stable in their conclusions: for the RWSSU model, a 10% perturbation in the input certainties caused only a 1.2% change in the output certainties. 3. Results of the NURE Pilot Study Having established the credibility of the RWSSU model by the test results just discussed, we then undertook an evaluation of five test regions selected by the Department of Energy. For this purpose USGS and DOE geologists completed questionnaires for this model. As a sensitivity test, several geologists independently completed questionnaires for each test region. For comparison, the model designer, R. Rackley, also completed questionnaires for the five test regions. The overall results are reported in Table 3, in which the abbreviations M.H., P.B., MO., N.G., and W.R. denote the names of the test regions, namely Monument Hill, Pumpkin Buttes, Moorcroft, Northwest Gillette, and White River, respectively. Table 3. Overall Conclusions for Five Test Regions Geologist A B C D USGS Rackley Range team data ------------------------------------------------ M.H. 4.17 3.32 3.97 4.40 1.08 P.B. 4.20 3.30 4.19 4.40 1.10 MO. 3.92 3.88 4.00 0.12 N.G. 3.64 0.10 3.42 3.54 W.R. 0.13 0.01 0.12 The results in Table 3 indicate that the Monument Hill, Pumpkin Buttes, and Moorcroft regions are very favorable, and about equally favorable, for occurrence of “Western States” sandstone uranium deposits. Northwest Gillette is scored as moderately favorable, whereas White River is neutral (balanced positive and negative indicators). Note that each respondent has had different exposure to the target regions, in terms of both first-hand, on-site experience and familiarity with field data reported in the literature. These differences in experience are reflected in their answers on the questionnaires. Since different inputs yield different conclusions, one would expect a spread in the certainties about each region, reflecting the differences in input data provided by the various geologists. Inspection of Table 3 reveals, however, that the scores derived from different geologists’ input data about the same region agree rather closely for each region except Northwest Gillette (see the column labeled IrRange”). These generally close agreements reflect the capability of Prospector models to synthesize many diverse factors, mechanically ascertaining general commonalities without being unduly distracted by occasional disparities. In cases such as Northwest Gillette in which a large difference in conclusions occurs, it is easy to trace the source of the disagreement by comparing the individual conclusions for different sections of the model (representing different geological subconclusions), as in Table 4. Table 4. Comparison of Detailed Conclusions About Northwest Gi I lette Geologist C D Rackley Avg. data ----------------------------------------- RWSSU .lO 3.66 3.42 3.56 FTRC 4.67 3.80 4.63 4.37 TECTCN 4.90 4.50 4.50 4.63 AHR 4.95 1.03 4.94 3.64 FAVHQST 5.00 5.00 5.00 5.00 S EDTECT 4.98 4.33 4.78 4.69 FAVSED .04 3.92 4.79 2.92 FLUVSED .04 3.92 4.79 2.92 MAR I NESED -4.60 3.34 .02 -.41 AEOLSED -4.99 -2.10 -3.23 -3.44 FMA .27 2.45 1.33 2.18 RBZQNE 4.10 4.83 4.73 4.55 AIZQNE -3.29 2.40 0.00 -0.30 M I NZQNE .41 2.82 2.59 1.94 Inspection of Table 4 reveals that the conclusions agree fairly closely for the FTRC section of the model, and less closely for the FAVSED and FMA sections. Tracing the differences deeper, one sees that of the three factors on which FMA depends, there is fairly good agreement about RBZONE, but larger differences in the cases of the AIZONE and MINZONE sections. In some cases, such a detailed analysis can isolate the source of overall disagreement to a few key questions about rhich the respondents disagreed. These can then be resolved by the respondents without the need to be concerned with other disagreements in their questionnaire inputs that did not significantly affect the overall conclusions. Prospector has also been applied to several other practical tasks. One evaluated several regions on the Alaskan Peninsula for uranium potential [l 1, as one of the bases for deciding their ultimate disposition (e.g., wilderness status versus commercial exploitation). Another app!iccztian das concc’rnad with measuring quantitatively the economic value of a geological map, resulting in statistically significant results [ 71. 4. Discussion We have measured Prospector’s expertise explicitly and presented a practical application to a national project, demonstrating in particular how the Prospector approach deals effectively with the variabilities and uncertainties inherent in the task of resource assessment. This work illustrates that expert systems intenderr for actual practical use must accommodate the special characteristics of the domain of expertise. In the case of economic geology, it is not rare for field geologists to disagree to some extent about their observations at a given site. Accordingly, the use of various sorts of sensitivity analysis is stressed in Prospector to bound the impact of such disagreements and to isolate their sources. In so doing, we provide geologists with new quantitative techniques by which to address an important issue, thus adding to the attractiveness of Prospector as a working tool. Other domains of expertise will have their own peculiarities, which nust be accommodated by designers of expert systems for those domains. A more mundane, but nevertheless important, example concerns the use of a questionnaire as a medium for obtaining input data to Prospector from geologists. Most geologists have little or no experience with computers; furthermore, access to a central computer from a remote site may be problematic in practice. On the other hand, geologists seem to be quite comfortable with questionnaires. Our point is simply that issues ancillary to Al usually have to be addressed to ensure the practical success of knowledge-based Al systems. References 1. 2. 3. 4. 5. 6. 7. 8. Cox, D. P., D. E. Detra, and R. L. Detterman, “Mineral Resources of the Chignik and Sutwick Island Quadrangles, Alaska,” U.S. Geological Survey Map MF-1053K, 1980 in press. Duda, R.O., P.E. Hart, P. Barrett, J. Gaschnig, K. t\onolige, R. Reboh, and J. Slocum, “Development of the Prospector Consultation System for Mineral Exploration,” Final Report, SRI Projects 5821 and 6415, Artificial Intelligence Center, SRI International, Menlo Park, California, October 1978. Gaschnig, J. G., “Preliminary Performance Analysis of the Prospector Consultant System for Mineral Exploration,” Proc. Sixth International Joint Conference on Artificial Intelligence, Tokyo, August 1979. Gaschnig, J. G., “Development of Uranium Exploration Models for the Prospector Consultant System,” SRI Project 7856, Artificial Intelligence Center, SRI International, Menlo Park, California, March 1980. National Uranium Resource Evaluation, Interim Report, U.S. Department of Energy, Report GJO-111(79), Grand Junction, Colorado, June 1979. Roach, C. H., “Overview of NURE Progress Fiscal Year 1979,” Preprint of Proceedings of the Uranium Industry Seminar, U. S. Department of Energy, Grand Junction, Colorado, October 16-17, 1979. Shapiro, C., and W. Watson, “An Interim Report on the Value of Geologic Maps,” Preliminary Draft Report, Director’s Office, U.S. Geological Survey, Reston, Virginia, 1979. Yu, V.L., et al., “Evaluating the Performance of a Computer-Based Consultant,” Heuristic Programming Project Memo H PP-78-17, Dept. of Computer Science, Stanford University, September 1978. 297

1980

39

33

INCREMENTAL, INFORMAL PROGRAM ACQUISITION’ Brian P. McCune Advanced Information & Decision Systems 201 San Antonio Circle, Suite 286 Mountain View, California 94040 AbJrract. Program acquisition is the transformation of a program specification into an executable, but not necessarily efficient, program that meets the given specification. This paper presents a solution to one aspect of the program acquisition problem, the incremental construction of program models from informal descriptions [II, in the form of a framework that includes (1) a formal language for expressing program fragments that contain informalities, (2) a control structure for the incremental recognition and assimilation of such fragments, and (3) a knowledge base of rules for acquiring programs specified with informalities. 1. Introduction The paper describes a LISP based computer system called the Program Model Builder (abbreviated “PMB”), which receives informal program fragments incrementally and assembles them into a very high level program model that is complete, semantically consistent, unambiguous, and executable. The program specification comes in the form of partial program fragments that arrive in any order and may exhibit such informalities as inconsistencies and ambiguous references. The program fragment language used for specifications is a superset of the language in which program models are built. This program modelling language is a very high level programming language for symbolic processing that deals with such information structures as sets and mappings. 2. The Problem The two key problems faced by PMB come from processing fragments that specify programs incrementally and informally. The notion of incremental program specification means that the fragments specifying a program may be received in an arbitrary order and may contain an arbitrarily small amount of new information. The user thus has the most flexibility to provide new knowledge about any part of the program at any time. For example, a single fragment conveying a small number of pieces of information is the statement “A is a collection.” This identifies an information structure called A and defines it as a collection of objects. However, the fragment says nothing about the type of objects, their number, etc. These details are provided in program fragments occurring before or after this one. Informality means that fragments may be incomplete, semantically inconsistent, or ambiguous; may use generic operators; and may provide more than one equivalent way of expressing a program part. An incomplete program model part may be completed either by use of a default value, by inference by PMB, or frog later fragments from the user. ’ This paper describes research done at the Stanford Artificial Intelligence Laboratory and Systems Control, Inc., and supported by DARPA under Contracts MDA903-76-C-0206 and NOOOl4-79 C-0127, monitored by ONR. Fellowship support was provided by NSF, IBM, and the De Karman Trust. The views and conclusions contained in this paper are those of the author. Program model consistency is monitored at all times. PMB tries to resolve inconsistencies first; otherwise, it reports them to the user. For example, the membership test fragment x E A requires that either A have elements of the same type as x (whenever the types of R and x finally become known) or their types inferred to be the same. Because a fragment may possess ambiguities, its interpretation depends upon the model context. So PMB specializes a generic operator into the appropriate primitive operation, based upon the information structure used. For example, part-oflx,A) (a Boolean operation that checks if information structure x is somehow contained within A) becomes x E R, if A is a collection with elements of the same type as X, and an k.componenr if A is a plex (record structure). PMB is capable of canonization, the transformation of equivalent information and procedural structures into concise, high level, canonical forms. This allows subsequent automatic coding the greatest freedom in choosing implementations. Interesting patterns are detected by specific rules set up to watch for them. For example, expressions that are quantified over elements of a set are canonized to the corresponding expression in set notation. 3. Control Structure The model building problem is to acquire knowledge in the form of a program model. The control structure of PMB is based upon the “recognition” paradigm 121, in which a system watches for new information, recognizes the information based upon knowledge of the domain and the current situation, and then integrates the new knowledge into its knowledge base. PMB has one key feature: subgoals may be dealt with in an order chosen by the user, rather than dictated by the system. Subgoals are satisfied either externally or internally to PMB. The two cases are handled by the two kinds of data driven antecedent rules, response rules and demons, which are triggered respectively by the input of new fragments or changes in the partial model. When new information arrives in fragments, appropriate response rules are triggered to process the information, update the model being built, and perhaps create more subgoals and associated response rules. Each time a subgoal is generated, an associated “question” asking for new fragments containing a solution to the subgoal is sent out. This process continues until no further information is required to complete the model. To process subgoals that are completely internal to PMB, demon rules are created that delay execution until their prerequisite information in the model has been filled in by response rules or perhaps other demons. 4. Knowledge Base PMB has a knowledge base of rules for handling constructs of the program modelling language. processing informalities in fragments, monitoring consistency of the model, and doing limited forms of program canonization. Rules about the modelling language include facts about five different information structures, six control structures, and approximately twenty primitive operations. The 71 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. control structures are ones that are common to mose high level languages. The modelling language’s real power comes from its very high level operators for information structures such as sets, lists, mappings, records, and alternatives of these. Below are English paraphrases of three rules that exemplify the major types of rules used in PMB. Rule I is a response rule for processing a new loop. Rule 2 is a demon that checks that the arguments of an i~-&~et operation are consistent. Rule 3 is a canonization demon that transforms a case into a test when appropriate. [II A loop consists of an optional initialization, required body, and required pairs of exit tests and exit blocks. Each exir test must be a Boolean expression occurring within the body. [21 Require that the two arguments of an is-J&Jet operation both return collections of the same prototypic element. [31 If the statement is a case, the case has two condition/action pairs, and the first condition is the negation of the second condition, then change the case into a test. Both response rules and simple demons are procedural. Compound demons (i.e., those whose antecedents test more than one object in the model) use declarative antecedent patterns that are expanded automatically into procedural form by a rule “compiler”. 5. Example of PMB in Operation The model building excerpt below displays (1) growth of the program model tree in a fashion that is generally top down, but data driven, and (2) completion and monitoring of parts of the model by demons. Note that this excerpt does justice neither to the concept of arbitrary order 6f fragments nor the types of programming knowledge in PMB. The trace discusses three program fragments generated from an English dialog. Each fragment is followed by a description of how it was processed by PMB, a snapshot of the partial model at that point, and a list of the outstanding demons. A detailed trace for the first fragment shows PMB focusing on individual slots of a fragment, creating model templates, and creating subgoals. The other fragments emphasize the creation and triggering of demons. Names preceding colons are unique template names that allow fragments to refer to different parts of the model. Missing parts of the partial model are denoted by “???“. Newly added or changed lines are denoted by the character ‘I” at the right margin. [The excerpt starts after the first fragment has already caused the partial program model shown below to be created. It only contains the names of the model, CLASSIFY, and the main algorithm, “algorithm-body”. No demons exist.] Current program model: program classify; algorithm-body: ??? Current demons act ive: None [The second fragment describes the top level algoriehm as a control structure having type composite and two “Input-concept” steps called and “classify-loop”. This fragment might have arisen from a sentence from the user such as “The algorithm first inputs the concept and then classifies it.“] Inputting fragment: algorithm-body: begin input-concept classify-loop end CA composite is a compound statement with an optional partial ordering on the execution of its subparts. The response rule that processes the composite creates the following two subgoals, along with response rules to handle them (not shown).] Processing ALGORITHM-BOOY.TYPE = COMPOSITE Creating subgoal: ALGORITHM-BOOY.SUBPARTS - 711 Creating subgoal: ALGORITHM-BOOY.OROERINGS = 711 Done processing ALGORITHM-BOOY,TYPE = COMPOSITE [Within the same fragment the two subparts are defined as operational units with unique names, but of unknown types. An operational unit can be any control structure, primitive operation, or procedure call. Two new templates are created and their types are requested.1 Process i ng ALGORI THM-BODY. SUBPARTS = (INPUT-CONCEPT CLASSIFY-LOOP) Creating template INPUT-CONCEPT with value I NPUT-CONCEPT. CLASS - OPERAT I ONAL-UNI T Creating subgoal: INPUT-CONCEPT.TYPE - 131 Creating template CLASSIFY-LOOP with value CLASSIFY-LOOP.CLASS = OPERATIONAL-UNIT Creating subgoal: CLASSIFY-LOOP. TYPE = ??? Done processing ALGORITHM-BOOY.SUBPARTS = (INPUT-CONCEPT CLASSIFY-LOOP) [At this point, the of the composite.] model IS missing the definitions of the two parts Current program model: program classify; begin Input-concept: ???; classify-loop: ??? end Current demons active: None [The third fragment, which defines “input-concept” to be an input primitive operation, is omitted. Information structures from this fragment are not shown in the models below. The fourth fragment defines the second step of the composite. This fragment might have come from “The classificaeion step is a loop with a single exit condition.“] Inputting fragment: classify-loop: until exit (exit-condition) repeat loop-body finally exit: endloop [This fragment defines a loop that repeats “loop-body” (as yet undefined) until a Boolean expression called “exit-condition” is true. At such time, the loop is exited to the empty exit block, called “exit”, which is associated with “exit-condition”. Since PMB doesn’t know precisely where the test of aexit-condition” will be located, it is shown separately from the main algorithm below. The response rule that processes the loop needs to guarantee that “exit-condition” is contained within the body of the loop. Since 72 this can’t be determined until the location of “exit-condition” is defined in a fragment, the response rule attaches Demon 1 to the template for “exit-condition” to await this event. Similarly, Demon 2 is created to await the location of “exit-condition” and then put it inside a test with an asserl-exit-connlilon as its true branch. This will cause the loop to be exited when the exit condition becomes true.1 Current program model8 program classlf3r; begirl concept c input(concept#otot~pe, user, concept-prompt); until exit repeat I loop-body: ??? finally I exit: endloop I end exit-condition: ??? I Current demons active: Demon 1: awaiting control structure containing “ex i t-cond i t i on” I Demon 2: awaiting control structure containing “ex i t-cond i t i on” I [The fifth fragment defines the body of the loop, thus triggering the two demons set up previously. A possible source of this fragment is ‘The loop first inputs a scene, tests whether the datum that was input is really the signal to exit the loop, classifies the scene, and then outpu&this classification to the user.“] Inputting fragment: loop-body: begin loop-input; exit-condition; classification; output-classification end [“Loop-body” is a composite with four named steps. PMB now knows where “exit-condition” occurs and that it must return a Boolean value. Demon I is awakened to find that “exit-condition” is located inside the composite “loop-body”. Since this isn’t a loop, Demon 1 continues up the tree of nested control constructs. It immediately finds that the parent of “loop-body” is the desired loop, and succeeds. Demon 2 is also awakened. Since it now knows “exit-condition” and its parent, Demon 2 can create a new template between them. The demon creates a test with “exit-condition” as its predicate and an assert-exit-condition that will leave the loop as its true action.] Current program mode I : program classify; begin concept t input(concept.$ototype, user, concept-prompt); until exit repeat begin loop-input: ???; I if exit-condition: ??? then assert-exit-conditio(exit); I classification: ??!; output-classification: ??? end finally exit: endloop end Current demons act i ve: None I [At the end of the excerpt, five of 32 fragments have been processed.] 6. Role of PMB in a Program Synthesis System PMB was designed to operate as part of a more complete program synthesis system with two distinct phases: acquisition and automatic coding. In such a system the program model would serve as the interface between the two phases. Automatic coding is the process of transforming a model into an efficient program without human intervention. The model is acquired during the acquisition phase; the model is coded only when it is complete and consistent. PMB may work within a robust acquisition environment. In such an environment, program fragments may come from many other knowledge sources, such as those expert in traces and examples, natural language, and specific programming domains. However, the operation of PMB is not predicated on the existence of other modules: all fragments to PMB could be produced by a straightforward deterministic parser for a surface language such as the one used to express fragments. 7. Conclusion PMB has been used both as a module of the PSI program synthesis system E31 and independently. Models built as part of PSI have been acquired via natural language dialogs and execution traces and have been automatically coded into LISP by other PSI modules. PMB has successfully built a number of moderately complex programs for symbolic computation. The most important topics for future work in this area include (1) extending and revising the knowledge base, (2) providing an efficient mechanism for testing alternate hypotheses and allowing program modification, and (3) providing a general mechanism for specifying where in the program model a program fragment is to go. The last problem has resulted in a proposed program reference language I1 3. 8. References [I J Brian P. McCune, Building Program Models Incrementally from informal Descriptions, Ph.D. thesis, AIM-333, STAN-CS-79 772, AI Lab., CS Dept., Stanford Univ., Stanford, CA, Oct. 1979. [2J Daniel G. Bobrow and Terry Winograd, “An Overview of KRL, a Knowledge Representation Language”, Cognifive SCienCe, Vol. 1, No. 1, Jan. 1977, pp. 3-46. [33 Cordell Green, Richard P. Gabriel, Elaine Kant, Beverly I. Kedzierski, Brian P. McCune, Jorge V. Phillips, Steve T. Tappel, and Stephen J. Westfold, “Results in Knowledge Based Program Synthesis”, 1 JCAI-79: Proceedings of the Sixth International Joint Conference on Artijciai Intelligence, Vol. 1, CS Dept., Stanford Univ., Stanford, CA, Aug. 1979, pp. 342-344. 73

1980

4

34

SOME REQUIREMENTS FOR A COMPUTER-BASED LEGAL CONSULTANT L. Thorne McCarty Faculty of Law, SUNY at Buffalo Laboratory for Computer Science Research, Rutgers Although the literature on computer-based consultation systems has often suggested the possibility of building an expert system in the field of law (see, e.g., [2]) it is only recently that several AI researchers have begun to explore this possibility seriously. Recent projects include: the development of a computational theory of legal reasoning, using corporate tax law as an experimental problem domain C61 C73 C81; the development of a language for expressing legal rules within a data-base management environment [Y]; the design of an information retrieval system based on a computer model of legal knowledge [ 31; and the design of an artificial intelligence system to analyze simple tort cases [lo]. This paper attempts to identify the principal obstacles to the development of a legal consultation system, given the current state of artificial intelligence research, and argues that there are only certain areas of the law which are amenable to such treatment at the present time. The paper then suggests several criteria for selecting the most promising areas of application, and indicates the kinds of results that might be expected, using our current work on the TAXMAN project [q] as an example. I. Poter&i& &nlications. One can imagine numerous applications of artificial intelligence techniques, in several diverse areas of law, but most of these would fall into one of the following categories: (I.) m-al Svstema. There are a number of systems in operation today which maintain data bases of statutes and decided cases, in full text, and which are capable of searching these texts for combinations of key words, using standard information retrieval techniques. (For a comparative survey of the principal domestic systems, LEXIS and WESTLAW, see cf21.1 These retrieval systems have turned out to be useful for certain kinds of legal research tasks, but only when used in conjunction with the traditional manual digests and indices, all of which are organized according to a rigid conceptual * This research has been funded by the National Science Foundation through Grant SOC-78- 1 1408 and Grant MCS-79-21471 (1979-81). classification of the law (usually: the West "key number systemV1). With the use of artificial intelligence techniques, however, the retrieval systems could be augmented to provide a form of automated conceptual searching as well, and without the rigidities of the manual indices. For a discussion of these possibilities, see [6] and c31. (2.) I&g,& Analvsis and Planni. Svstems. A step more sophisticated than a retrieval system, a legal analysis and planning system would actually analyze a set of facts, or propose a sequence of transactions, in accordance with the applicable body of legal rules. This is the kind of system that most often comes to mind when one speculates about computer-based legal consultation, for it is the system most similar to the successful systems in chemical and medical domains: a lawyer, engaged in a dialogue with a computer, would describe the facts of his case, and the computer would suggest an analysis or a possible course of action. In fact, there are systems of this sort under development today, using techniques much less powerful than those available to the artificial intelligence community, and they seem close to commercial application: see, e.g., [I31 . The advantages of artificial intelligence techniques for these applications have been discussed by [ 61. (3.) Integrated u Information Svstems. Instead of looking only at the tasks of the private attorney, we could focus our attention more broadly on the legal system as a whole. One of the tasks of the legal system is to manage information and to make decisions about the rights and obligations of various individual actors, and there seems to be no reason, in principle, why some of this information and some of these decisions could not be represented entirely within a computer system. For a current example, using conventional programming technology, consider the computerized title registration systems which are now being used to manage real estate transactions (see, e.g., c511. With the availability of artificial intelligence techniques, a large number of additional applications come to mind: financial transactions, securities registration, corporate taxation, etc. At present it appears that these possibilities are being taken more seriously by European lawyers than by American lawyers (see, e.g., [II and [Ill). From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. If we consider the potential role of artificial intelligence techniques in all of these applications, a basic paradigm emerges. A computer-based legal consul tat ion system must represent the 11facts11 of a case at a comfortable level of abstraction, and it must represent the I1 1 awl1 in the chosen area of application. The “lawl’ would consist of a system of “concepts” and ltrulesl’ with the following characteristics : (a. 1 they would be relatively abstract, that is, they would subsume large classes of lower-level factual descriptions; and (b. ) they would have normative implications, that is, they would specify which actions were permitted and which actions were obligatory in a given situation. Legal anal ysi 3, in its simplest form, would then be a process of applying the lllawll to the “facts.” . Put this way, the paradigm seems to be an ideal candidate for an artificial intelligence approach: the “facts11 would be represented in a lower-level semantic network, perhaps ; the ltlawll would be represented in a higher-level semantic description ; and the process of legal analysis would be represented by a pattern-matching routine. The difficult problems with this paradigm, however, are the representation problems. In the existing knowledge-based systems, in other domains, the representation of the llfactsll and the I1 1 awl1 has been relatively straight forward. In DENDRAL, for example, the ground-level description of all possible chemical strut tures could be represented in a simple graphical notation, and the rules for splitting these structures in a mass spectrograph coul d be -represented as simple operations on the links of the graphs. In MYCIN, the basic facts of a case could be represented as a set of features listing the presence or absence of certain symptoms and the results of certain laboratory tests, and the diagnostic rules could then be represented as a probabilistic judgment that a given symptom or test result implied a certain disease. By contrast, the facts of a legal case typically involve al 1 the complexities of daily life : human actions, beliefs, intentions, motivations, etc. , in a world of ordinary objects like houses and automobil es, and complex institutions like businesses and courts. Even if the facts of a particular case could be represented in a computer system, the rules themselves would often be problematical. Some rules, usually those embodied in statutes, have a precise logical strut ture , and this makes them amenable to the existing artificial intelligence techniques. But it is a commonplace among lawyers that the most important legal rules do not have this form at all: instead they are said to have an ltopen texture”; their boundaries are not fixed, but are llconstruc ted” and ltmodifiedll as they are applied to particular fat tual situations. A sophisticated legal consul tation system would not be able to ignore these complexities, but would have to address them directly.- II. Possible mroaches. Since the representation problems for a legal consul tation system are so difficult, it is tempting to start with the “simplestl’ possible 1 egal issue 3, such as the subject matter of the first-year law school courses. We might therefore be tempted to investigate assault and battery cases from the first-year torts course [ lOI, or offer and acceptance cases from the first-year contracts tour se . But these cases are ltsimplelV for law students primarily because they draw upon ordinary human experience, and this is precisely what makes them so difficult for an artificial intelligence system. To understand tort cases, we must understand all the ways in which human beings can be injured, intentionally and unintentionally, mental 1 y and physically , with and without justification. To understand contract cases, we must understand the expectations of real people in concrete business situations, and the ambiguities of human language in expressing particular contractual intentions. If we abstract away these details, we will miss entirely the central features of legal reasoning, and our consultation systems will tend to produce only the more trivial results. Paradoxically, the cases that are most. tractable for an artificial intelligence system are those cases, usually involving commercial and corporate matters, which a lawyer finds most complex . There is a simple reason why this is so. A mature 1 egal system in an industrialized democracy i a composed of many levels of legal abstractions: the law initially defines llrightsll in terms 0 f concrete objects and ordinary human actions , but these rights are then treated as llobjects I1 themselves, and made subject to further human llactionsll ; by repeating this process of reification many times, a complex body of commercial law can be developed. Because of their technical complexity, the legal rules at the top levels of this conceptual hierarchy are difficult for most lawyers to comprehend, but this would be no obstacle for an artificial intelligence system. The commercial abstractions, in fact, are artificial and formal systems themselves, drained of much of the content of the ordinary world; and because of the commercial pressures for precision and uniformity, they are, by legal standards, well structured. A reasonable strategy for developing a computer-based legal consul tation system, then, would be to start here. This is the strategy we have followed in the TAXMAN project C61 [7] . The TAXMAN system operates in the field of corporate tax law, which is very near the apex of the hierarchy of commercial abstractions. The basic 11facts11 of a corporate tax case can be captured in a relatively straightforward representation: corporations issue securities, transfer property, distribute dividends, etc. Below this level there is an expanded representation of the meaning of a security interest in terms of its component rights and obligations: the owners of the shares of a common stock, for example, have certain rights to 299 the "earnings", the "assets", and the llcontrol" of the corporation. Above this level there is the lllawll : the statutory rules which classify transactions as taxable or nontaxable, ordinary income or capital gains, dividend distributions or stock redemptions, etc. Although these rules are certainly complex, the underlying representations are manageable, and we have concluded from our earlier work that the construction of an expert consultation system in this area of the law is a feasible proposition C61 . In our current work [7] we are taking this model one step further, in an attempt to account for the "open texture" of legal concepts and the process by which a legal concept is "constructed" and "modified" during the course of an argument over a contested case. In many areas of the law this would be an impossible task: the complexity of the representation would be overwhelming, and the structural and dynamic properties of the concepts would be obscured. But in the world of corporate abstractions the task appears to be feasible. Looking at a series of corporate tax cases before the United States Supreme Court in the 1920's and 19301s, we have been able to construct a model of the concept of "taxable income" as it appeared to the lawyers and judges at the time (see [9]) . Although the concept is sometimes represented as a "logicalfl pattern which can be "matched" to a lower-level semantic network (we call this a -ical temblate structure), the more important representation for the process of legal analysis consists of a prototvpe structure and a sequence of deformations of the prototype. We are currently involved in an implementation of these ideas, and we will describe them in detail in a future paper (for an initial description of the implementation, see [81). We believe that the "prototype-plus-deformation" structure is an essential component of a system of legal rules, and that it ought to play an important role in a sophisticated legal consultation system. III. Prosbects. This paper has emphasized the difficulties in constructing a computer-based legal consultation system, but it has also suggested some feasible approaches. The main difficulty is the representation problem: the factual situations in a legal problem domain involve complex human actions, and the most important legal rules tend to contain the most amorphous and malleable legal concepts. By selecting legal problems from the commercial and corporate areas, however, we can construct a representation of the legally relevant facts without having to model the entire human world, and we can begin to develop the necessary structures for the representation of the higher- level legal rules. We have had some success with this strategy in the TAXMAN project, and we believe it can be applied elsewhere as well. REFERENCES T: 11 Bing, J., and Harvold, T., &Q&- Decisions m Information Svstems (Universitetsforlaget, Oslo, 1977). [2] Buchanan, B.G., and Headrick, T.E., "Some Speculation About Artificial Intelligence and Legal Reasoning," 23 Stanford ti Review 40- ( 1970). [3] Hafner, C., "An Information Retrieval System Based on a Computer Model of Legal Knowledge," Ph.D. Dissertation, University of Michigan (1978). [4] Jones, S., Mason, P., and Stamper, R., "LEGOL 2.0: A Relational Specification Language for Complex Rules, )1 4 Information Svstems 293-305 ( 1979). [5] Maggs, P.B., "Automating the Land Title System," 22 American Universitv Law Review 369-91 (1973). C63 McCarty, L.T., flReflections on TAXMAN: An Experiment in Artificial Intelligence and Legal Reasoning,11 90 Harvard & Review 837-93 (1977). [7] McCarty, L.T., "The TAXMAN Project: Towards a Cognitive Theory of Legal Argument," in B. Niblett, ed., Computer &g& & Legal &gl&@Jg (Cambridge University Press, forthcoming 1980). 183 McCarty, L.T., and Sridharan, N.S., "The Representation of an Evolving System of Legal Concepts: I. Logical Templates," in Proceedings. Third National Conference of the Canadian Societv for Computational Studies of A- Intelligence, Victoria, British Columbia, May 14-16, 1980. [9] McCarty, L-T., Sridharan, N-S., and Sangster, B.C., "The Implementation of TAXMAN II: An Experiment in Artificial Intelligence and Legal Reasoning," Report LRP-TR-2, Laboratory for Computer Science Research, Rutgers University (1979). [ IO1 Meldman, J.A., "A Preliminary Study in Computer-Aided Legal Analysis," Ph.D. Dissertation, Massachusetts Institute of Technology, Technical Report No. MAC-TR-157 (November, 1975). [ 111 Seipel, P., Computing & (LiberForlag, Stockholm, 1977). [I21 Sprowl, J.A., A Manual for Commuter-Assisted Legal Research (American Bar Foundation, 1976). El31 Sprowl, J.A. "Automating the Legal Reasoning Process: A Computer that Uses Regulations and Statutes to Draft Legal Documents," 1979 American & Foundation Research Journal l-81. 300

1980

40

35

A FRAME-BASEDPRODUCTION SYSTEM ARCHITECTURE David E. Smith and Jan E. Clayton Heuristic Programming Project* Department of Computer Science Stanford University ABSTRACT We propose a flexible frame-structured representation and agenda-based control mechanism for the construction of production-type systems. Advantages of this architecture include uniformity, control freedom, and extensibility. We also describe an experimental system, named WHJXZE, that uses this formalism. The success of MYCIN-like production systems 141 [7] [9] has demonstrated that a variety of types of expertise can be successfully captured in rules. In some cases, however, rules alone are inadequate necessitating the USC of auxiliary representations (e.g. property lists for paramzters in MYCIN). Other limitations result from the use of goal-directed control. In this paper we outline a flexible schemata for constructing high performance production-like systems. The architecture consists of two components: 1. An extensible representation (utilizing a frame-structured language) which captures production rule knowledge. 2. An agenda-based control mechanism allowing considerable freedom in tailoring control flow. We have used this architecture in the development of a system named WHEEZE, which performs medical pulmonary function diagnosis based on clinical test results. This syst%m is based on two earlier efforts, PUFF [7], an EMYCIN-based production rule system Ill], and CENTAUR [l] 121, a system constructed of both rules and prototypes. AN ~LTERNATXVE RI~PRESENTATION FOH PRODUCTIONS Figure 1 shows how a typical PUFF rule would be transformed into our representation. Each assertion is represented as a frame in the knowledge. base, with anteccdcnt sub-assertions appearing in its h4auzjdarion slot. The number associated with each manifestation is its corresponding importance. Similarly, the certainty factor and findings from the rule are given separate slots in the assertion. Assertions appearing in the SuggesliveOf and ComplemenlaryTo . . ..-........... *This research was supported in part by ONR contract N00014- 79C-0302 and by .NIH Biotechnology Resource Grant RR-00785. 54 slots are those worth investigating cornfirmed or denied respectively assertions are suggesfivilies). Implicit in the production rule if the original assertion is (numbers following these representation is a fimction which indicates how to compute the “belief’ of the consequent assertions given belief in the antecedent assertion. Unfortunately, evaluation of the antecedent assertion involves modal logic (since greater shading is required than simple binary values for belief and disbelief). Therefore, a “HowToDetermineBelief’ slot is associated with each assertion indicating how its belief is to be computed. If: 1) The severity of Obstructive Airways Disease of the patient is less than or greater to mild, al’;1 2) The number of pack-years smoked is greater than 0, and 3) The number of years ago that the patient quit smoking is 0 Then: It is definite (1000) that the following is one of the conclusion statements about this interpretation: Discontinuation of smoking should help re1iev.e the symptoms. OADwithSmoking: Manifestation SuggestiveOf ComplementaryTo Certainty Findings ((OAD-Present 10) (PatientHasSmoked 10) (PatientSillSmoking 10)) ((SmokingExaccrbatedOAD 5) (SmokingInduccdOAD 5)) ((OADwithSmoking-None 5)) 1000 “Discontinuation of smoking should help relieve the symptoms. ” HowToDetermincBelicf finelion for compuring the minimum of the beliefs of the manifestations Figure 1. English translation of PUFF rule (top) and Corresponding WHEEZE Frame for OADwithSmoking (bottom). Numbers appearing in the Manifestation, SuggcstiveOf and Complcmcntaryl’o slots arc importance and suggcstivity wcightings. From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. The declarative nature of this representation facilitates modification and extension. For example, the addition of related knowledge, such as justifications, explanations, and instructional material, can be accomplished by the addition of slots to already existing assertions. The single uniform structure alleviates the need for any auxilliary means of representation. Considerable efficiency has been gained by the use of rule compilation on production systems [lo] [ll]. We feel that a technique similar to this could also be used effectively on our representation but have not yet fully investigated this possibility. AN AGENDA BASED CONTROL MECHANISM Depth-first, goal-directed search is often used in production systems because questions asked by the system are focused on specific topics. Thus, the system appears to follow a coherent line of reasoning, more closely mimicking that of human diagnosticians. There are, however, many widely recognized limitations. No mechanism is provided for dynamically selecting or ordering the initial set of goals. Consequently, the system may explore many “red herrings” and ask irrelevant questions before encountering a good hypothesis. In addition, a startling piece of evidence (strongly suggesting a different hypothesis) cannot cause suspension of the current investigation and pursuit of the alternative. Expert diagnosticians use more than simple goal-directed reasoning. They seem to work by alternately constructing and verifying hypotheses, corresponding to a mix of data- and goal- directed search. Furthermore, they expect these systems to reason in an analogous manner. It is desirable, therefore, that the system builder have control over the dynamic reasoning behavior of the system. To provide this control, we employ a simple relaxation of goal- and data-directed mechanisms. This is facilitated by the use of an agenda to keep track of the set of goals to be examined, and their relative priorities. The control strategy is: 1. Examine the top assertion on the agenda. 2. If its sub-assertions (manifestations) arc known, the relative belief of the assertion is determined. If confirmed, any assertions that it is suggestive of are placed on the agenda according to a specified measure of suggestivity. If denied, complementary assertions are placed on the agenda according to a measure of suggestivity. 3. If it cannot be immediately verified or rejected then its unknown sub-assertions are placed on the agenda according to a measure of importance, and according to the agenda level of the original assertion. By varying the importance factors, SuggesfiveOf values, and the initial items placed on the agenda, numerous strategies are possible. For example, if high-level goals are initially placed on the agenda and subgoals are always placed at the top of the agenda, depth-first goal-directed behavior will result. Alternatively, if low-level data are placed on the agenda initially, and assertions suggested by these data assertions arc always placed below them on the agenda, breadth-first data driven behavior will result. More commonly, what is desired is a mixture of the two, in which assertions suggest others as being likely, and goal directed verification is employed to investigate the likely assertions. The example below illustrates how this can be done. FEVl/FVC<80 QAD FEVl/FVC>SO Figure 2. A simplified portion of the WHEEZE knowledge base. The solid lines indicate Manifesatation links (e.g. OAD is a manifestation of Asthma), the dashed lines represent SuggestiveOf links. The numbers associated with the links are the corresponding “importances” and “suggestivities” of the connections. In -the knowledge base of figure 2, suppose that RDX-ALS is confirmed, suggesting RLD to the agenda at level 6 and ALS at level 4. ‘RLD is then examined, and since its manifestations are unknown, they are placed at the specified level on the agenda. The agenda now contains FEVl/FVC>80 at level 8, RV<80 and RLD at level 6, and ALS at level 4. FEVl/FVC>80 is therefore selected, and suppose. that it is found to be false. Its complementary assertion (FEVl/FVC<80) is placed at level 8 on the agenda. and is immediately investigated. It is, of course, true, causing OAD to be placed at level 8 on the agenda. The diagnosis proceeds by investigating the manifestations of OAD; and, if OAD is confirmed, Asthma and Bronchitis are investigated. While many subtleties have been glossed over in this example, it is important to note that: 1. The manipulation of SuggestiveOf and importance values can change the order in which assertions are examined, therefore changing the order in which questions are asked and results printed out. (In the example, FEVUFVC was asked for before RV.) 155 2. Surprise data (data contrary to the hypothesis currently being investigated) may suggest goals to the agenda high enough to cause suspension of the current investigation. (The surprise FEVl/FVC value caused suspension of the RLD investigation in favor of the OAD investigation. If the suggestivity of the link from FEVl/FVC<80 to OAD were not as high, this would not have occurred.) 3. Low-level data assertions cause the suggestion of high- level goals, thus selecting and ordering goals to avoid irrelevant questions. (In the ex‘ample, RLD and ALS were suggested and ordered by the low-level assertion RDX-ALS.) Thus, extreme control flexibility is provided by this mechanism. Besides the mechanism proposed above, there have been several other attempts to augment simple goal directed search. Meta-rules [S] can be used to encode strategic information, such as how to order or prune the hypothesis space. They could also be used, in principle, to suspend a current investigation when strong alternatives were discovered. In ‘practice, however, meta-rules for accomplishing this task could be quite clumsy. In the CENTAUR system [l] [2], procedural attachment mechanisms (in disease prototypes) are used to capture the control information explicitly, and “triggering” rules serve to order the initial hypothesis space. Our solution differs from these earlier attempts by proposing a single uniform control mechanism. lt is sufficiently straightforward that tailoring of the control flow could potentially be turned over to the domain expert. RIZSULTS Not suprisingly, WHEEZE exhibits the same diagnostic behavior as its predecessors, PUFF and CENTAUR, on a standard set of 10 patient test cases. In refining the knowledge base, suggestivities and importance factors were used to great advantage to change the order in which questions were asked and conclusions printed out. This ‘eliminated the need to carefully order sets of antecedent assertions. The reprcscntation described has proven adcquatc for capturing the domain knowledge. In some cases, several rules were collapsed into a single assertion. In addition, the combination of representation and control structure eliminated the need for many awkward interdependent rules (e.g. rules with screening clauses). Representation of both the rule and non-rule knowledge of the PUFF and CENTAUR systems has been facilitated by the flexibility of the architecture described. This flexibility is the direct result of the uniform representation and control mechanism. Further exploitations of this architecture appear possible, providing directions for future research. ACKNOWLEDGEMENTS We would like to thank Jan Aikins, Avron Barr, James Bennett, Bruce Buchanan, Mike Genesereth, Russ Greiner, and Doug Lenat for their help and comments. 111 PI t31 141 151 161 171 PI [91 [lOI REFERENCES Jan. S. Aikins. Prototypes and Production Rules: an Approach to Knowledge Representation for Hypothesis Formation. Proc 6th IJCAI, 1979. Pp. l-3. Jan S. Aikins. Profolypes and Producrion Rules: A Knowledge Represenrarion for Computer Consullations Doctoral dissertation, Dept. of Computer Science, Stanford University. Jan S. Aikins. Representation of Control Systems. Proc Zsr AAAI, 1980. Knowledge in Expert James Bennett, Lewis Creary, Robert Englemorc and Robert Melosh. SACON: A Knowledge Based Consultant for Sfructural Analysis. Computer Science Report CS-78-699, Dept. of Computer Science, Stanford University, September 1978. Randall Davis and Bruce G. Buchanan. Meta-Level Knowledge: Overview and Applications. Proc Slh IJCAI, 1977. Pp. 920-927. Russell Greiner, Douglas Lenat. A Language. Proc Zsf AAAI, 1980. Representation Language J. C. Kunz, R. J. Fallet, et. al.. A Physiological Rule Based System for Inlerpreting Pulmonary Function Test Results. HPP- 78-19 (Working Paper), Heuristic Programming Project, Dept. of Computer Science, Stanford University, December 1978. Stephen G. Pauker and Peter Szolovits. Analyzing and Simulating Taking the History of the Present Illness: Context Formation. In Schneider and Sagvall Hein (Eds.), Compulational Linguislics in Medicine. North-Holland, 1977. Pp. 109-118. E. hi. Shortliffe. Computer-based Medical Consultations: MYCIN. New York: American Elsevier, 1976. William van Melle. A System for Consultation 923-92s. Domain-independent Production-rule Programs. Proc 6th IJCAI, 1979. Pp. [ll] William van Melle. A Domain-independenr Syslem thai Aids in Construcling Knowledge-based Consultation Programs. Doctoral dissertation, Dept. of Computer Science, Stanford University. 156

1980

41

36

ABSTRACT Knowledge Embedding in the Description System Omega Carl Mewitt, 6iuseppe Attardi, and Maria Simi M.I.T. 545 Technology Square Cambridge, Mass 02139 Omega is a description system for knowledge embedding which combines mechanisms of the predicate calculus, type systems, and pattern matching systems. It can express arbitrary predicates (achieving the power of the o-order quantificational calculus), type declarations in programming systems (Algal, Simula, etc.), pattern matching languages (Planner, Merlin, KRL, etc.). Omega gains much of its power by unifying these mechanisms in a single formalism In this paper we present an axiomatization of basic constructs in Omega which serves as an important component of the interface between implementors and users+ Omega is based on a small number of primitive concepts. It is sufficiently powerful to be able to express its own rules of inference In this way Omega represents a self-describing system in which a great deal of knowledge about itself can be embedded. The techniques in Omega represent an important advance in the creation of self-describing systems without engendering the problems discovered by Russell Meta-descriptions (in the sense used in mathematical logic) are ordinary descriptions in Omega. Together with Jerry Barber we have constructed a preliminary implementation of Omega on the MLT. CADR System and used it in the development of an office workstation prototype. I -- Introduction First Order Logic is a powerful formalism for representing mathematical theories and formalizing hypotheses about the world. Logicians have developed a mathematical semantics in which a number of important results have been established such as completeness. These circumstances have motivated the development of deductive systems based on first order predicate calculus [FOL, PROLOG, Bledsoe’s Verifier, etc. ] However, First Order Logic is unsatisfactory as a language for embedding knowledge in computer systems. Therefore tnany recent reasoning system have tried to develop their own formalisms [PLANNER, FRL, KL-ONE, KRL, LMS, NETL, AMORD, XPRT, ETHERJ The semantics and deductive theory of these new systems however has not been satisfactorily developed. The only rigorous description of most of them has been their implementations which are rather large and convoluted programs. 2 -- Overview The syntax of Omega is a version of template English, For example we use the indefinite article in instance descriptions such as the one below: (a Son) Instance descriptions like the previous one in general describe a whole category of objects, like the category of sons, in this example Such description however can be tnade more specific, by ‘prescribing particular attributes for the instance ‘description So for example, (a Son (With father Paul) (With mother Mary)) describes a son with father Paul and with mother Mary. Otnega differs from systems based on records with attached procedures (SIMULA and its descendants), generalized property lists (FRL, XRL, etc.), frames ‘(Minsky), and units (KRL) in several important respects. One of the most important differences is that instance descriptions in Omega cannot be updated. This is a consequence of the monotonicity of knowledge accutnulation in Omega Change in Omega is modeled through the use of viewpoints [Barber: 19801 Another difference is that in Omega an instance description can have more than one attribution with the same relation, For example (a Human (with child Jack) (with child Jill)) is a description of a human with a child Jack and a child Jill Statements inheritance can be relation deduced is. For of because example transitivity of the 157 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. (John is (a Man)) can be deduced from the following statements (John is (a Son)) ((a Son) is (a Man)) In order to aid readability we will freely mix infix and prefix notations. For example the statement (John is (a Man)) is completely equivalent to (is John (a Man)) 3 -- Inheritance The inheritance relation in Omega differs somewhat from the usual ISA relation typically found in semantic networks. For example from (John iS (a Human)) ((a Human) is (a Mammal)) (Human is (a Species)) we can deduce (John is (a Mammal)) but cannot conclude that (John is (a Species)). However we can deduce that (John is (a (a Species))) which says that John is something which is a species We can often avoid the use of explicit universal quantifiers. For instance the following sentence in quantificational calculus Vx Man(x) * Mortal(x) can be expressed as ((a Man) is (a Mortal)) In this case we avoid the need to restrict the range of a universal quantifier by means of a predicate, as it is usually necessary in the quantificational calculus. When unrestricted quantification is required, or when we need to give a name to a description which occurs several times in the same expression, we denote a universally quantified variable by prefixing the symbol = to the name of the variable, wherever it appears in the statement, as in: (=x is (a Man)) * (=x is (a Mortal)) The scope of such a variable is the whole statement. Thus the above statement is an abbreviation for (for-a// =x ((=x is (a Man)) 3 (=x is (a Mortal)))) Occasionally it is necessary to use a for-all in the interior of a statement. For example in the following statement expresses the Axiom of Extensionality which is one of the most fundamental principles in Omega: ((for-a// =d ((=d is =dr) * (=d is =d2))) 3 (=dl is =d2)) In the above statement, the scope of =dl and =d2 is the whole statetnent, while the scope of =d is only the statetnent ((=d is =dl) * (=d is =d2)). A form of existential quantification is implicit in the use of attributions+ For instance (Pat is (a Man (with father (an Irishman)))) says that there is an Irishman who is Pat’s father. Omega makes good use of the ability to place nontrivial descriptions on the left hand side of an inheritance relation. For example from the following statements (a Teacher (with subject =s)) iS (an Expert (W/f/l field ES)) (John is (a Teacher (With subject Music))) we get the following by transitivity of inheritance: (John iS (an Expert (Wifh field Music))) Note that statements like the following (is (and (a WarmBloodedAnimal) (a BearerOfLiveYoung)) (a Mammal)) are much more difficult to express in systems such as KRL and FRL which are based on generalized records and property lists. If it happens that two descriptions inherit from each other, we will say they are the same. For example if ((a Woman) is (a Human (Wifh sex female))) ((a Human (wifh sex female)) is (a Woman)) then we can conclude (a Woman) Same (a Human (wifh sex female)) We can express the general principle being used here in Omega as follows 158 ((=dl same =d2) <=> (A f=dl is =d2) (=d2 is =dl))) 4 -- Lattice Operators and Logical Operators The domain of descriptions constitutes a complemented lattice, with respect to the inheritance ordering is, meet and join operations and and or, complementation operation not, and Nothing and Something as the bottom and top respectively of the lattice Some axioms are required to express these relations For example all descriptions inherit from Something. (=d is Something) Furthermore Nothing is the complement of Something (Nothing same (not Something)) The usual logical operators on statements are A, V, -, => for conjunction, disjunction, negation, and implication respectively. The description operators not, and, or, etc. apply to all descriptions including statements. It is very important not to confuse the logical operators with the lattice operators in Omega. Note for example: ((A true false) is false) ((and true false) is /Vothingj ((and true true) is true) Unfortunately most “knowledge representation languages” have not carefully distinguished between lattice operators and logical operators leading to a great deal of confusion. Note that a statement of the form (p is true) does not in general imply that (p same true). For example ((Nixon iS (a UnindictedCoConspirator)) is true) (((a Price (with merchandise Tea) (with place China)) is 81) is true) does not imply (same (Nixon is (a UnindictedCoConspirator)) (a Price (with merchandise Tea) (with place China))) 5 -- Basic Axioms We will state some of the axioms for the description system. The axioms for a theory are usually stated in a metalanguage of the theory. However, since our language contains its metalanguage, we can here give the axioms as ordinary statements in the description system itself. 5.1 Extensionality Inheritance obeys an Axiom of Extensionality which is one of the most fundamental axioms of Omega. Many important properties can be derived from extensionality which can be expressed in Omega as follows: (W (=descriptionl iS =description2) (for-a// =d (=j (=d iS =descriptionl) (=d iS zdescription2)))) Note that the meaning of the above statement would be drastically changed if we simply omitted the universal quantifier as follows (W (=descriptionl iS =description2) (3 (=d iS =descriptionl) (=d is =description2))) The axiom extensionality illustrates the utility of explicitly incorporating quantification in the language in contrast to some programming languages which claim to be based on logic. From this axiom alone we are able to derive most of the lattice-theoretic properties of descriptions. In particular we can deduce that is is a reflexive and transitive relation. The following (=description iS =description) expresses the reflexivity of inheritance whereas the following (4 (A (=descriptionl /S =description2) (=description2 is =descriptions)) kdescriptionl is zdescriptiong)) expresses the transitivity of inheritance. 5.2 Commutativity Commutativity says that the order in which attributions Of a concept are written is irrelevant We use the notation that an expression of the form XC..>> is a sequence of 0 or more elements 159 (same (a =descriptionl <<=attributionsl>> =attribution2 <<=attributions3>> =attributionq <<=attributionsg>>) (a =descriptionl <<=attributionsl>> =attributionq <<=attributionsg>> =attribution2 <<=attributionsg>>)) (Susan is (a Mother (with child Jim) (with father Bill) (with child (a Female)))) 5.5 Monotonicity of Atributes Monotonicity of attributes is a fundamental property of instance descriptions which is close)y related to transitivity of inheritance. (=descriptionl is =description2) (is (a =concept (With =attribute =descriptionl)) (a =concept (With =attribute =descriptionp)))) For example ((a Father (with child Henry) (With mother Martha)) same (a Father (with mother Martha) (with child Henry))) For example if 5.3 Deletion (Fred is (an American)) (Bill is (a Person (with The axiom of Deletion is that attributions of an instance description can be deleted to produce a more general instance description. (is (a =descriptionl <<=attributions-l>> zattribution-2 <<=ettributiona-3 (a =descriptionl <<=attributions-I>> <<=attributions-3)))) father Fred))) then (Bill is (a Person (With father (an kkTWkaI7~~~~ Note that the complementation in monotonic. For example Omega is not For example ((a Bostonian) is (a NewEnglander)) (is (a Father (With child Henry) (With mother Martha)) (a Father (with mother Martha))) does not imply that ((not (a Bostonian)) is (not (a NewEnglander))) 5.6 Constraints 5.4 Merging Constraints can be used to restrict the objects will satisfy certain attributions. For example which One of the most fundamental axioms in Omega is Merging which says that attributions of the same concept can be merged. (a Human (withh?straint child (a Male))) describes humans who have Axiom for Constraints is only male children. The (=descriptionl is (a =description2 <<=attributions-l>>)) kdescriptionl is (a =description2 <<=attributions-2))))) (is =descriptionl (a =description2 <<=attributions-1>> <<=attributions-2))))) ((a =C (withconstraint =R =dl) (with =R =d2)) is (a =C (with =R (and=dl =d2)))) If For example if (Joan is (a Human (WithConstraint child (a Male)) (With child Jean))) (Susan is (a Mother (with child Jim))) (Susan is (a Mother (with father Bill))) (Susan is (a Mother (With child (a Female)))) 160 then 6.2 Projective Relations (Joan iS (a Human (with child (SfK! (a Male) Jean)))) Note that solely from the statement (Ann iS (a Human (with child (a Male)) (With child Jean))) no important conclusions can be drawn in Omega. It entirely possible that Jean is a female with a brother. is We have found the constrained attributions in Omega to be useful generalizations of the increasingly popular “constraint languages” which propagate values through a network of property lists. If (2 is (a Complex (with real-part (> 0)))) and (2 iS (a Complex (With real-part (an Integer)))) then by merging it follows that (z iS (a Complex (With real-part (> 0)) (with real-part (an Integer)))). However in order to be able to conclude that (z iS (a Complex (With real-part (and (> 0) (an Integer))))) some additional information is needed. One very general way to provide this information is by (rsalgart iS (a Projective-relation (with concept Complex))) 6 -- Higher Order Capabilities and by the statement In this section we present examples which illustrate power of the higher-order capabilities of Omega. the 6.1 Transitive Relations If (3 is ($7 Integer (with larger 4))) and (4 is (an Integer (with larger 5))), we can conclude by monotonicity that (3 is (an Integer (With larger (an Integer (with larger 5))))) From the above statement, we would like to be able to conclude that (3 is (an Integer (with larger 5))). This goal can be accomplished by the statement (larger is (a Transitive-relation (with concept Integer))) which says that larger is a transitive relation for the concept Integer. The Axiom for Transitive Relations states that if R is a transitive relation for a concept c and x is an instance of c which is R-related to an instance of c which is R-related to m, then x is R-related to m. (=> (=R iS (a Transitive-relation (With concept =C))) (is (a =C (with =R (a =C (with =R =m)))) (a =C (with =R m)))) The desired conclusion can be reached by using the above description with c bound to Integer, R bound to larger, and m bound to 5. (=R k (a Projective-relation (with concept =C))) (is (a =C (with =R =d)) (a =C (wifhConsfrainf =R =d)))) The desired conclusion is reached by using the above description with =C bound to Complex, =R bound to real-part, =descriptionl bound to (> 01, and =description2 bound to (an Integer). 6.3 Inversion Inverting relations for efficiency of retrieval is a standard technique in data base organization. Inversion makes use of the converse of a relation with respect to a concept which satisfies the following Axiom for Converse: (=R same (a Converse (With relation (a Converse (with relation =R) (With concept =C))) (With concept =C))) The Axiom of Inversion expresses how to invert inheritance relations for constrained instance descriptions (<=> (=dl is (a =C (withConstraint =R (an =d2)))) ((a =R (wifh (a Converse (with relation =R) (With relation =C) =dl)) is =d2))) 161 For example suppose ((a Converse (with relation son) (with concept Person)) same Parent) we can conclude (Sally iS (a Person (wifhConsfrainf son (an American)))) if and only if ((a Son (with parent Sally)) is (an American)) We have inversion to be a useful generalization of the generalized selection mechanisms in Simula, SmallTalk, and KRL as well as the generalized getprop mechanism in FRL. The interested reader might try to define the transitivity, projectivity, and converse relations in other “knowledge representation languages” 7 -- Conclusions Omega encompasses ‘?he capabilities of both the w-order quantification calculus, type theory, and pattern matching languages in a unified way. We have illustrated how Omega is more powerful than First Order Logic by showing how it can directly express important properties of relations such as transitivity, projectivity, and converse that . are not first order definable Omega is based on a small number of primitive concepts including inheritance, instantiation, attribution, viewpoint, logical operations (conjunction, disjunction, negation, quantification, etc.) and lattice operations ( meet, join, complement, etc.) It makes use of inheritance and attribution between descriptions to build a network of descriptions in which knowledge can be embedded. Omega is sufficiently powerful to be able to express its own rules of inference. In this way Omega represents a self-describing system in which a great deal of knowledge about itself can be embedded. Because of its expressive power, we have to be very careful in the axiom system for Omega in order to avoid Russell’s paradox. Omega uses mechanisms which combines ideas from the Lambda Calculus and Intutionistic Logic to avoid contradictions in the use of self reference We have found axiomatization to be a powerful technique in the development, design, and use of Omega Axiomatization has enabled us to evolve the design of Omega by removing many bugs which have shown up as undesirable consequences of the axioms. The axiomatization has acted as a contract between the implementors and users of the system. The axioms provide a succinct specification of the rules of inference that can be invoked. The development of Omega has focused on the goals of conceptual simplicity and power. The axiomatization of Omega in itself is a measure of our progress in achieving these goals. 8 -- Related Work The intellectual roots of our description system go back to von Neumann-Bernays-Goedel set theory [Goedel: 19401, the o-order quantificational calculus, and the lambda calculus. Its development has been influenced by the property lists of LISP, the pattern matching constructs in PLANNER-71 and its descendants QA-4, POPLER, CONNIVER, etc., the multiple descriptions and beta structures of .MERLIN, the class mechanism of SIMULA, the frame theory of Minsky, the packagers of PLAShlA, the stereotypes in [Hewitt: 19751, the tangled hierarchies of NETL, the attribute grammars of Knuth, the type system of CLU, the descriptive mechanisms of KRL-0, the partitioned semantic networks of [Fikes and Hendrix: 19771, the conceptual representations of [Yonezawa: 19771, the class mechanism of SMALL-TALK [Ingalls: 19781, the goblets of Knowledge Representation Semantics [Smith: 19781, the selector notation of BETA, the inheritance mechanism of OWL, the mathematical semantics of actors (Hewitt and Attardi: 19781, the type system in Edinburgh LCF, the XPRT system of Luc Steels, the constraints in [Borning: 1977, 1979 and Steele and Sussman: 1978] 9 -- Further Work We have also developed an Omega Machine (which is not described in this paper) that formalizes the operational semantics of Omega. Mike Brady has suggested that it might be possible to develop a denotational semantics for Omega along the lines of Scott’s model of the lambda calculus. This development is one possible approach to establishing the consistency of Omega. 162 10 -- Acknowledgments We are grateful to Dana Scott for pointing out a few axioms that were incorrectly stated in a preliminary version of this paper. Jerry Barber has been extremely helpful in aiding us in developing and debugging Omega. Brian Smith and Gene Ciccarelli helped us to clear up some important ambiguities. Conversations with Alan Borning, Scott Fahlman, William Martin, Allen Newell, Alan Perlis, Dana Scott, Brian Smith, and the participants in the “Message Passing Systems” seminar were extremely helpful in getting the description system ‘nailed down. Richard Weyhrauch has raised our interests in meta-theories. His system FOL is one of lthe first to exploit the classical logical notion of metatheory in AL systems Several discussions with Luc Steels have been the source of cross-fertilization between the ideas in our system and his XPRT system ‘Roger Duffey and Ken Forbus have served as extremely able teaching assistants in helping to develop this material for the Paradigms for Problem Solving Course at MIT. Comments by Peter Deutsch and Peter Szolovits have materially helped to improve the presentation. Our logical rules of inference are a further development of a natural deduction system by Kalish and Montague. Some of the axioms for inheritance were inspired by Set Theory. 11 -- Bibliography Barber, G. “Reasoning About Change in Knowledgeable Office Systems” 1980. Birtwistle, G. M; Dahl, 0.; Myhrhaug, B.; and Nygaard, K. “SIM..JLA Begin” Auerbach 1973. Bobrow, D. G. and Winograd, T. “An Overview of KRL-0 Knowledge Representation Language” ’ Co:nitive Science VoL 1 No. I. 1977. Borning, A. “ThingLab -- An Object-Oriented System for Building Simulations Using Constraints” Proceedings of IJCAI-77. August, 1977. Bourbaki, N. “Theory of Sets” Book I of Elements of Mathematics. Addison-Wesley. 1968. Church, A.. “A Formulation of the Simple Theory of Types”, 194 1. Dahl, 0. J. and Nygaard, K. “Class and Subclass Declarations” In Simulation Programming Languages J. N. Buxton (Ed.) North Holland. 1968. pp. 158-174. Fahlman, Scott “Thesis Progress Report” MIT AI Memo 331. May, 1975. Fikes, R. and Hendrix, G. “A Network-Based Knowledge Representation and its Natural Deduction System” IJCAI-77. Cambridge, Mass. August 1977. pp 235-246. Goedel, K “The Consistency of the Axiom of Choice and of the Generalized Continuum Hypothesis with the Axioms of Set Theory” Annals of Mathematics Studies No. 3, Princeton, 1940. Hammer, M. and McLeod, D. “The Semantic Data Model: A Modeling &Mechanism for Data Base Applications SIGMOD Conference on the Management of Data. Austin Texas. May 31-June 2, 1978. Hawkinson, Lowell “The Representation of Concepts in OWL” Proceedings of IJCAI-75. September, 1975. Tiblisi, Georgia, USSR pp. 107-114. Hewitt, C. “Stereotypes as an ACTOR Approach Towards Solving the Problem of Procedural Attachment in FRAME Theories” “Proceedings of Interdisciplinary Workshop on Theoretical Issues in Natural Language Processing” Cambridge, June 1975. Kalish and Montague Kristensen, B. B.; Madsen, 0. L.; Moller-Pedersen, B.; and Nygaard, K. “A Definition of the BETA Language” TECHNICAL REPORT TR-8. Aarhus University. February 1979. Moore, J. and Newell, A. “How Can MERLIN Understand?” CMU AM. November, 1973. Quine, W. K. “New Foundations of Mathematical Logic” 1952 Burstall, R and Goguen , J. “Putting Theories Together to Make Specifications”, Proceedings of IJCAI-77. August, 1977. 163 Rulifson, J. F.; Derksen, J. A.; and Waldinger, R. J. “QA4: A Procedural Calculus for Intuitive Reasoning” SRI Technical Note 73. November 1972. Schubert, L. K. “Extending the Expressive Power of Semantic Networks” Artificial Intelligence 7. Steele, G. L. and Sussman, G. J. “Constraints” MIT Artificial Intelligence Memo 502. November 1978. Steels, L. Master Thesis, MIT 1979. Weyhrauch, R. “Prolegomena to a Theory of Formal Reasoning”, Stanford AI Memo AIM-315, December 1978. Forthcoming in AL JournaL 164

1980

42

37

A Representation Language Language Russell Greiner and Douglas B. Lenat Computer Science Deptartment Stanford University ABSTRACT The field of AI is strewn with knowledge representation languages. The language designer typically has one particular application domain in mind: as subsequent types of applications are tried, what had originally been US&II j2urure.y become undesirable limitations, and the language is overhauled or scrapped. One remedy to this bleak cycle might be to construct a representational scheme whose domain is the field of representational languages itself. Toward this end, we designed and implemented RLL, a frame-based Representation Languange Language. The components of representation languages in general (such ti slots and inheritance mechanisms) and of RLL itself are encoded declaratively as frames. Modifvinn these frames can change the of the RLL environment. semantics of RLL, by altering the 1. MOTIVATION “One ring to rule them all... and in the darkness bind them. ” Often a large Artificial Intelligence project begins by designing and implementing a high-level language in which to easily and precisely specify the nuances of the task. The language designer typically builds his Representation Language around the one particular highlighted application (such as molecular biology for Units [Sfe$k], or natural language understanding for KRL [Bobrow & Winogrudj and OWL [Szolovifs, et al.]). For this reason, his language is often inadequate for any subsequent applications (except those which can be cast in a form similar in structure to the initial task): what had originally been useful features become undesirable limitations (such as Units’ explicit copying of inherited facts, or KRL’s sophisticated but slow matcher). Building a new language seems cleaner than modifying the flawed one, so the designer scraps his “extensible, general” language after its one use. The size of the February 1980 SIGART shows how many similar yet incompatible representation schemes have followed this evolutionary path. One remedy to this bleak cycle might be to construct a representation scheme whose domain is the field of representational languages itself. a program which could thc’n bc tailored to suit many specific applications. Toward this end, WC arc designing and implementing RLL, an object-centered’ Representation Languange I,anguage.2 This paper reports on the current state of our ideas and our implementation. 1 This “object-centering” does not represent a loss in geneiality. We will soon see that each part of the full system, including procedural information, is reified as a unit. * As RLL is itself a completely self-descriptive there is no need for an RLLL. representation language, 2. INTRODUCTION RLL explicitly represents (i.e. contains units-? for) the components of reprcscntation languages in general and of itself in particular. The programming language LISP derives its flexibility in a similar manner: it, too, encodes many of its constructs within its own formalisms. Representation languages aim at easy, natural interfacing to users; therefore their primitive building blocks are larger, more abstract, and more complex than the primitives of programming languages. Building blocks of a representation language include such things as control regimes (ExhaustiveBackwardChaining, Agendas), methods 3f associating procedures with relevant knowledge (Footnotes, Demons), tindamental access functions (Put/Get, Assert/Match), automatic inference mechanisms (InheritFromEvery2ndGeneration, InheritBut-PermitExceptions), and even specifications of the intended semantics and epistemology (ConsistencyConstraint, EmpiricalHeuristic). of the components RLL is designed to help manage these complexities, by providing (1) an organized library of such representation language components, and (2) tools for manipulating, modifying, and combining them. Rather than produce a new representation language as the “output” of a session with RLL, it is rather the RLL language itself, the environment the user sees, which changes gradually in accord with his commands. 3. HOWJS A REPRESENTATION LANGUAGE LIKE AN ORGAN? When the user starts RLL, he finds himself in an environment very much like the Units package [Slefik], with one major difference. If he desires a new type of inheritance mechanism, he need only create a new Inheritance-type of unit, initialize it with that desired property; and that new mode of inheritance will automatically be enabled. This he can do using the same editor and accessing functions hc uses for cntcring and codifying his domain knowledge (say, poultry inspection); only hcrc the information pertains to the actual Knowledge Base system itself, not turkeys. The Units package has Get and Put as its tindamcntal storage and retrieval functions; therefore RLL also begins in that state. But there is nothing sacred about cvcn these two “primitives”. Get and Put arc encoded as modifiable units: if they are altered, the nature of accessing iI slot’s value will change correspondingly. In short, by issuing a small number of commands hc can radically alter the character of the RLI. cnvironmcnt. molding it to his personal -----______________ 3 RLL is a frame-based system [Minsky], whose building blocks are called Units [Stefik], [Bobrow & Winograd]. Each unit consists of a set of Slots with their respective values. 165 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. preferences and to the specific needs of his application. RLL is responsible for performing the necessary “truth maintainence” operations, (e.g. retroactive updates) to preserve the overall correctness of the system as a whole. For example, Get and Put can be transformed into units which, when run as functions, resemble Assert (store proposition) and Match (retrieve propositions), and the user need never again mention “slots” at all. RLL is more like a stop organ than a piano. Each stop corresponds to a “pre-fabricated” representational part (e.g. a slot, inheritance, format, control regime, etc.), which resides in the overall RLL system. The initial RLL is simply one configuration of this organ, with certain stops “pulled out” to mimic the Units package. These particular stops reflect our intuitions of what constitutes a general, powerful system. Some of the units initially “pulled out” (activated) define more or less standard inheritance regimes, such as Inherit-Along-IS-A-Links, which enables Fido to gather default information from AnyDog. We chose to include a large assortment, of common slots. One hundred and six types of slots, including IS-A, SuperClass. BroaderHeuristics, and TypicalExamples, are used to hierarchically organize the units. That number grows daily, as we refine the organizing relationships which were originally “smeared” together into just. one or two kinds of slots (e.g. A-Kind-Ofl. An additional fifteen types of slots, including ToGerValue, ToPut Value, ToKillUnit. and To Add Value, collectively define the accessing/updating functions. This bootstrapping system (the initial configuration of “organ stops”,) dots not span the scope of RLL’s capabilities: many of its stops are initially in the dormant position. Just as a competent musician can produce a radically different sound by manipulating an organ’s stops, so a sophisticated RLL user can define his own representation by turning off some features and activating others. For instance, an FRL devotee may notice -- and choose to use exclusively -- the kind of slot called A-Kind-Ox which is the smearing together of Is- A, SuperSet, Abstraction, TypicalExampleOJ PartOf etc. He may then deactivate those more specialized units from his system permanently. A user who does not want to see his system as a hierarchy at all can simply deactivate the A-Kind-Of unit and its progeny. The user need not worry about the various immediate, and indirect, consequences of this alteration (e.g., deleting the Inherit-Along-IS-A-Links unit); RLL will take care of them. By selectively pushing and pulling, he should be able to construct a system resembling almost any currently used representational language, such as KRL, OWL and KLONE;4 after all, an organ can be made to sound like a piano. Unlike musical organs, RIL also provides its user with mechanisms for building his own stops (or even type of stops, or even mechanisms for building stops). With experience, one can use RLL to build his own new components. Rather than building them from scratch, (e.g., from CAR, CDR, and CONS,) he can modi& some existing units of RLL (employing other units which are themselves tools designed for just such manipulations.) 4 This particular task, of actually simulating various existing Representation languages, has not yet been done. It is high on our agenda of things to do. We anticipate it will require the addition of many new components (and types of components) to RLL, many representing orthogonal decompositions of the space of knowledge representation. The following examples convey the flavor of what can currently done with the default settings of the RLL “organ stops”. 4. EXAMPLES be 4.1. EXAMPLE: Creating a New Slot In the following example, the user wishes to define a Fafher slot, in a sexist genealogical knowledge base which containes only the primitive slots Morher and Spouse. As RLL devotes a unit to store the necessary knowledge associated with each kind of slot, (see Figure 1,) defining a new kind of slot means creating and initializing one new unit. In our experience, the new desired slot is frequently quite similar to some other slot(s), with but a few distinguishing differences. We exploited this iegularity in developing a high level “slot-defining” language, by which a new slot can be defined precisely and succinctly in a single declarative statement. 1 Name: IS-A 1 Description: Lists the classes I AM-A member of. I Format: List-of-Entries 1 Datatype: Each entry represents a class of objects. I Inverse: Examples 1 IS-A: (AnySlot) 1 UsedSylnhefifance: Inherit-Along-/S-A-Links 1 MyTimeOfCfeation: 1 April 1979, 12:Ol AM 1 MyCreator: D.B.Lenat Figure # 1 - Unit devoted to the “IS-A” slot. There are many other slots which are appropriate for this unit; whose value will be deduced automatically (e.g. inherited from AnySlot) if requested. Creating a new slot for Father is easy: we create a new unit called Father, and fill its HighLevelDefn slot with the value (Composition Spouse Mofher). Composition is the name of a unit in our initial system, a so-called “slot-combiner” which knows how to compose two slots (regarding each slot as a function from one unit to another). We also fill the new unit’s Isa slot, deriving the unit shown in Figure 2. Name: 1 IS-A: I HighLevelDefn: Father (AnySlot) (Composition Spouse Mother) Figure # 2 - Slots filled in by hand when creating the unit devoted to the”Father” slot. Several other slots (e.g., the syntactic slots MyCreator, MyTimeOfCreation) are filled in automatically at this time. The user now asks for KarlPhilippEmanuel’s father, by typing (GetVal ue 'KPE 'Father). GetValue first tries a simple associative lookup (GET), but finds there is no Fafher property stored on KPE, the unit representing KarlPhilippEmanuel. GctValue then tries a more sophisticated approach: ask the Father unit how to compute the Father of any person. Thus the call becomes [Apply* (GetValue 'Father 'ToCompute) 'KPE]. Notice this calls on GctValue recursivelv. and once again there is no value stored on the ToCompure slot-of the unit tilled Father. The call now has expanded into [Apply* (Apply* (GetValue 'ToCompute 'ToCompute) 'Father) 'KPE]. Luckily, there is a value on the ToCompute slot of the unit ToCompute: it says to find the HighLevelDefn, find the slot- combiner it employs, find its ToComput$, and ask it. Our call is now cxpandcd out into the following: [Apply* (Apply* (GetValue 'Composition 'ToCompute) 'Spouse 'Mother) 'KPE]. ’ Each unit which represents a function has a ToCompure slot, which holds the actual LISP function it encodes. Associating such a ToCompute slot with each slot. reflects our view that each slot is a function, whose argument happens to be a unit, and whose computed value may be cached away. 166 The unit called Composition does indeed have a ToCompute Slot; after applying it, we have: CApply* ‘(h (x) (GetValue (GetValue x ‘Mother) ‘Spouse)) ‘KPE]. This asks for the Mother slot of KPE, which is always physically stored in our knowledge base, and then asks for the value stored in her Spouse slot. The final result, JohannSebastian, is then returned. It is also cached (stored redundantly for future use) on the Fafher slot of the unit KPE. See [Lenat et al., 19791 for details. Several other slots (besides ToCompufe) are deduced automatically by RLL from the HighLevelDefn of Father (see Figure 3) as they are called for. The Format of each Fafher slot must be a single entry, which is the name of a unit which represents a person. The only units which may have a Fafher slot are those which may legally have a Mofher slot, viz., any person. Also, since Fafher is defined in terms of both Mother and Spouse, using the slot combiner Composition, a value stored on KPE:Fafher must be erased if ever we change the value for KPE’s M&her or AnnaMagdelena’s Spouse, or the definition (that is, 7’oCompufe) of Composition. Name: 1 IS-A: Father (AnySlot) I 1 HighLevelDefn: (Composition Spouse Mother) 1 Description: Value of unit’s Mother’s Spouse. 1 1 Format: SingleEntry 1 Datatype: Each entry is unit representing a person. I 1 MahesSenseFor: AnyPerson 1 DefinedlnTermsOf: (Spouse Mother) I 1 DefinedUsing: Composition 1 ToCompute: (A (x) (GetValue (GetValue x ‘Mother) ‘Spouse) i Figure # 3 - Later form of the Father unit, showing automatically. slots filled in Notice the the ease with which a use; can currently :‘extend his representation”, enlarging his vocabulary of new slots. A similar, though more extravagant example would be to define FavorifeAunf *as (SingieMost (Unioning (Composition Sister Parenfs) (Composition Spouse Brofher Parenfs)) $$). Note that “Unioning” and “SingleMost” are two of the slot combiners which come with RLL, whose definition and range can be inferred from this example. It is usually no harder to create a new type of slot format (OrderedNonemptySet), slot combiner (TwoMost, Starring), or datatype (MustBePersonOverl6), than it was to create a new slot type or inheritance mechanism. Explicitly encoding such information helps the user (and us) understand the precise function of each of the various components. We do not yet feel that we have a complete set of any of these components, but are encouraged by empirical results like the following: The first two hundred slots we defined required us to define thirteen slot combiners, yet the lasf two hundred slots required only five new slot combiners. 4.2. EXAMPLE: Creating a New Inheritance Mode Suppose a geneticist wishes to have a type of inheritance which skips every second ancestor. He browses through the hierarchy of units descending from the general one called Inheritance, finds the closest existing unit, InheritSelectively, which he copies into a new unit, InheritFromEvery2ndGeneration. Editing this copy, he finds a high level description of the path taken during the inheritance, wherein he replaces the single occurrence of “Parenf” by “GrandParent” (or by two occurrences of Parenf, or by the phrase (Composition Parenf Parent)). After exiting from the edit, the new type of inheritance will be active; RLL will have translated the slight change in the unit’s high-level description into a multitude of low-level changes. If the geneticist now specifies that Organism # 34 is an “InheritFromEvery2ndGeneration offspring” of Organism#20, this will mean the right thing. Note that the tools used (browser, editor, translator, etc.) are themselves encoded as units in RLL. 4.3. EXAMPLE: Epistemological Status Epistemological Status: To represent the fact that John believes that Mary is 37 years old, RLL adds the ordered pair (SeeUnit AgeOfMaryOOOl) to the the Age slot of the Mary unit. RLL creates a unit called AgeOfMaryOOOl, fills its *value* slot with 37 and its EpiSfa/us slot with “John believes”. See Figure 4. Note this mechanism suffices to rcprcscnt belief about belief (just a second chained SeeUnit pointer), quoted belief (“John thinks he knows Mary’s age”, by omitting the *value* 37 slot in AgeOfMaryOOOl), situational fluents, etc. This mechanism can also be used to represent arbitrary n-ary relations, escaping the associative triple (i.e. Unit/SZof/value) limitation. Name: ] IS-A: Mary I Description: (Person Female ContraryActor) 1 The grower of silver bells etc. I i Age: 1 Name AgeOfMaryOOOl il Name- 1 Isa (UnitForASlotFiller) I I lsa AgeOfMaryOO02 I I LiveslnUnit I I LiveslnUnit (UnitForASlotFiller) I Mary Mary I 1 LiveslnSlot Age- I ~Zs 37 John believes I Teleology: Epistemic . . I 1 LiveslnSlot Age- 1 1 *value* 1 I I Epistatus gring Wedding805 I 1 I Teleology: Historic I Figure 4 - Representing “John believes that Mary is 37, but she’s When she was married. she was 21”. really 39. 4.4. EXAMPLE: Enforcing Semantics Suppose that Lee, a user of RLL, is constructing HearSayXXIV, a representation language which contains cooperating knowledge sources (KSs). He specifies that each unit representing a knowledge source should have some very precise applicability criteria (he defines a FuZZRelevancy slot) and also a much quicker, rougher check of its potential relevance (he defines a PrePreCondifions slot). If HearSayXXIV users employ these two slots in just the way he intended, they will be rewarded with a very efficiently-running program. But how can Lee be sure that users of HearSayXXIV will use these two new slots the way he intended? He also defines a new kind of slot called Semantics. The unit for each type of slot can have a Semanfics slot, in which case it should contain criteria that the values stored in such slots are expected to satisfy. Lee fills the Semanfics slot of the unit called PrePreConditions with a piece of code that checks that the PrePreConditions slot of every KS unit is filled by a Lisp predicate, which is very quick to execute, which (empirically) correlates highly to the FullRelevancy predicate, and which rarely returns NIL when the latter would return T. This bundle of constraints captures what he “really means” by PrePreCondifions A user of HearSayXXIV, say Bob, now builds and runs a speech understanding program containing a large collection of cooperating knowledge sources. As he does so, statistics are gathered empirically. suppose Bob frequently circumvents the PrePreCondifions slot cntircly, by placing a’ pointer there to the same long, slow, complete criteria he has written for the FuNReZevancy slot of that KS. This is empirically caught as a violation of one of the constraints which Lee recorded in the Semanfics slot of the unit PrePreConditions. As a result, the Semanfics slot of the Semantics unit will be consulted to find an appropriate reaction; the code therein might direct it to print a warning message to Bob: “The PrePreCondifions slot of a KS is meant to run very quickly, compared with the F&Relevancy slot, but 70% of yours don’t; please change your PrePreCondifions slots, 167 or your FullRelevancy slots, or (if you insist) the Semantics slot of the PrePreConditions unit”.6 5. SPECIFICATIONS FOR ANY REPRESENTATION LANGUAGE LANGUAGE The following are some of the core constraints around which RLL was designed. One can issue commands to RLL which effectively “turn off’ some of these features, but in that case the user is left with an inflexible system we would no longer call a representation language language. Further details may be found in [Lena& Haye* Rorh, & Klahr] and in [Geneserelh & Lenaf]. Self-description: No part of the RLL system is opaque; even the primitive Get and Put and Evaluate finctions are represented by individual units describing their operation.2 Current status: complete (to a base language level). Self-modification: Changes in the high-level description of an RLL process automatically result in changes in the Lisp code for -- and hence behavior of -- RLL. Current status: this works for changes in definition, format, etc. of units representing slots and control processes. Much additional effort is required. Codification of Representation Knowledge: Taxonomies of inheritance, function invocation, etc. Tools for manipulating and creating same. These correspond to the stops of the organ, illustrated above. Current status: this is some of the most exciting research we foresee; only a smattering of representation knowledge has yet been captured. 6. INITIAL “ORGAN STOPS” The following characteristics pertain especially to the initirll state of the current RLL system, wherein all “organ stops” are set at their default positions. Each RLL user will doubtless settle upon some different settings, more suited to the representation environment he wishes to be in while constructing his application program. For details, see [Greiner]. Cognitive economy: Decision-making about what intermediate values to cache away, when to recompute values, expectation- f-iltcring. Current status: simple reasoning is done to determine each of these decisions; the hooks for more complex procedures exist, but they have not been used yet. Syntactic vs Semantic slots: Clyde should inherit values for many slots from TypicalElephant, such as Color, Diet, Size; but not from slots which refer to TypicalElephant qua data structure, slots such as NumerOjFilledInSlots and DaleCreated. Current status: RLL correctly treats these two classes of slots differently, e.g. when initializing a new unit. Onion field of languages: RLL contains a collection of features (e.g., automatically adding inverse links) which can be individually enabled or disabled, rather than a strict linear sequence of higher and higher level languages. Thus it is more like an onion field than the standard “skins of an onion” layering. Current status: Done. Three of the most commonly used settings are bundled together as CORLL, ERLL, and BRLL. ’ This work has led us to realize the impossibility of unambiguously stating semantics. Consider the case of the semantics of the Lisp function “OR”. Suppose one person believes it evaluates its arguments left to right until a non- null value is found: a second person believes it evaluates right to left: a third person believes it evaluates all simultaneously. They go to the Semanfics slot of the unit called OR, to settle the question. Therein they find this expression: (OR (Evaluate the args Iefi to right) [Evaluate the args right to leJ)J. Person #3 is convinced now that he was wrong, but persons 1 and 2 point to each other and exclaim in unison “See, I toId you!” The point of all this is that even storing a Lisp predicate in the Semanfics slots only specifies the meaning of a slot up to a set of fixed points. One approaches the description of the semantics with some preconceived ideas, and there may be more than one set of sucpi hypotheses which are consistent with everything stored therein. See [Genesereth & Lenar]. Economy via Appropriate Placement: Each fact, heuristic. comment, etc. is placed on the unit (or set of units) which are as general and abstract as possible. Frequently, new units are created just to facilitiate such appropriate placement. In the long run, this reduces the need for duplication of information. One example of this is the use of of appropriate conceptual units: Clarity of Conceptual Units: RLL can distinguish (i.e. devote a separate unit to each of) the following concepts: TheSetOfAllElephants, (whose associated properties describe this as a set -- such as #OjMembers or SubCaregories), TypicalElephant, (on which we might store Expected-TuskLenglh or DefauKolor slots), ElephantSpecies, (which EvoZvedAsASpecies some 60 million years ago and is CloselyRelaredTo the HippopatamusSpecies,) ElephantConcept, (which QuaZifiesAsA BeastOfBurden and a TuskedPackyderm,) ArchetypicalElephant (which represents an elephant, in the real world which best exemplifies the notion of “Elephant-ness”). It is important for RLL to be able to represent them distinctly, yet still record the relations among them. On the other hand, to facilitate interactions with a human user, RLL can accept a vague term (Elephant) from the user or from another unit, and automatically refine it into a precise term. This is vital, since a term which is regarded as precise today may be regarded as a vague catchall tomorrow. Current status: distinct representations pose no problem; but only an exhaustive solution to the problem of automatic disambiguation has been implemented. 7. CONCLUSION “‘...in Mordor, where the Shadow lies. ” The’system is currently usable, and only through use will direction.* for future effort be revealed. Requests for documentation and access to RLL are encouraged. There are still many areas for tirther development of RLL. Some require merely a large amount of work (e.g., incorporating other researchers’ representational schemes and conventions); others require new ideas (es., handling intensional objects). To provide evidence for our arguments, we should exhibit a large collection of distinct representation languages which were built out of RLL; this we cannot yet do. Several specific applications systems live in (or are proposed to live in) RLL; these include ELI-RISK0 (discoiery of heuristic rules), E&E (combat gamina). FUNNEL (taxonomv of Lisp objects, with an aim toward-autor&tic programming), ROGET (Jim Bennett: guiding a medical expert to directly construct a knowledge based system), VLSI (Mark Stefik and Harold Brown: a foray of AI into the VLSI layout area), and WHEEZE (Jan Clayton and Dave Smith: diagnosis of pulminary function disorders, reported in [Smifh & Clayton]). Experience in AI research has repeatedly shown the need for a flexible and extensible language -- one in which the very vocabulary can be easily and usefully augmented. Our representation language language addresses this challenge. We leave the pieces of a representation in an explicit and modifiable state. By performing simple modifications to these representational parts (using specially-designed manipulation tools), new representation languages can be quickly created, debugged, modified, and combined. This should ultimately obviate the need for dozens of similar yet incompatible representation languages, each usable for but a narrow spectrum of task. 68 ACKNOWLEDGEMENTS The work reported here represents a snapshot of the current state of an on-going research effort conducted at Stanford University. Researchers from SAIL and HPP are examining a variety of issues concerning representational schemes in general, and their construction in particular (viz., [Nii & Aiello]). Mark Stefik and Michael Geneseretb provided frequent insights into many of the underlying issues. We thank Terry Winograd for critiquing a draft of this paper. He, Danny Bobrow, and Rich Fikes conveyed enough of the good and bad aspects of KRL to guide us along the path to RLL. Greg Ilarris implemented an early system which perfomred the task described in Section 4.1. Others who have directly or indirectly influenced this work include Bob Balzer, John Brown, Cordell Green, Johan deKleer, and Rick Hayes-Roth. To sidestep InterLisp’s space limitation, Dave Smith implemented a demand unit swapping package (see [Smith]). The research is supported by NSF Grant #MCS-79-01954 and ONR Contract #NOOO14-80-C-0609. BIBLIOGRAPHY Aikins, Jan, “Prototypes and Production Rules: An Approach to Knowledge Representation from Hypothesis Formation”, HPP Working Paper HPP-79- 10, Computer Science Dept., Stanford University, July 1979. Bobrow, D.G. and Winograd, T., “An Overview of KRL; a Knowledge Representation Language”, SIJCAI, August 1977. a Brachman, Ron, “What’s in a Concept, Structural Foundations for Semantic Networks”. BBN report 3433, October 1976. Findler, Nicholas V. (ed.). Associafive Networks. NY. hcadcmic Press, 1979. Gencsereth, Michael, and Lenat, Douglas B.. SeIfDescripfion and -Modijicnrion in a Knowledge Representation Language, HPP Working Paper HPP-80-10, June, 1980. Greiner. Russell. “A Representation Language Language”, HPP Working Paper HPP-80-9, Computer Science Dept., Stanford University, June 1980. Lenat, Douglas B.. “AM: Automated Discovery in Mathematics”, SIJCAI, August 1977 Lenat, D. B., Hayes-Roth, F. and Klahr. P., “Cognitive Economy”, Stanford HPP Report HPP-79-15. Computer Science Dept., Stanford University, June 1979. Minsky, Marvin, “A Framework for Representing Knowledge”, in 77re I’s#rofogv OJ Computer Vision. P. Winston (ed.), McGraw-Ilill, New York, 1975. Nii, II. Penny, and Aiello, N., “AGE (Attempt to Generalize): A Knowledge- Based Program for Building Knowledge-Based Program”, 6IJCAI. August 1979. SIGART Newsletter, February 1980 (Special Representation Issue; Brachman & Smith, eds.). Smith. David and Clavton. Jan. “A Frame-based Production System Architecture”, AAAI Conference, .1980. Smith, David, “CORLL: A Demand Paging System for Units”, HPP Working Paper HPP-80-8, June 1980. Stefik, Mark J.. “An Examination of a Frame-Structured Representation System”, SIJCAI, August 1977. Szolovits, Peter, Hawkinson, Lowell B., and Martin, William A., “An Overview of OWL, A Language for Knowledge Representation”, MITLLCS/TM-86, Massachusetts Institute of Technology, June 1977. Woods, W. A., “What’s in a Link, Foundations for Semantic Networks”, in D. G. Bobrow & A. M. Collins (eds.), Representation and Understanding, Academic Press, 1975. , 169

1980

43

38

SPATIAL AND QUALITATIVE ASPECTS OF REASONING ABOUT MOTION Kenneth D. Forbus MIT Artificial Intclligcncc Laboratory 545 Technology Square Cambridge, Mass. 02139 ABSTRACT Reasoning about motion is an important part of common sense knowledge. The spatial and qualitative aspects of reasoning about motion through fret space arc studied through the construction of a program to perform such reasoning. An analog gcomctry rcprcscntation serves as a diagram, and descriptions of both the actual motion of a ball and envisioning are used in answering simple questions. 1 Introduction People reason fluently about motion through space. For example, WC know that if two balls arc thrown into a well they might collide, but if one ball is always outside and the other always inside they cannot. The knowl”cdgc involved in this qualitative kind of reasoning seems to be simpler than formal mechanics and appears to be based on our expcricnce in the physical world. Since this knowledge is an important part of our common SCIISC, capturing it will help us to understand how pcoplc think and enable us to make machines smarter. The issues involved in reasoning about motion through space wcrc studied by constructing a program, called FROB, that reasons about motion in a simple domain. 1 believe three important ideas have been illustrated by this work: 1. A quantitative geometric rcprescntation simplifies reasoning about space. It can provide a simple method for answering a class of gcomctric questions. The descriptions of space required for qualitative reasoning can bc defined using the quantitative representation, making it a communication device between several rcprcsentations of a situation 2. Describing the actual motion of an object can be thought of as creating a network of descriptions of qualitatively distinct types of motion, linked by descriptions of the state of the object before and after each of these motions. This network can bc used to analyze the motion and in some cases can be constructed by a process of simulation. 3. The description of the kinds of motion possible from some state (called the cnvisionment) is useful for answering certain questions and for checking an actual motion against assumptions made about it. The assimilation of assumptions about global properties of the motion into this description makes heavy use of spatial descriptions. FRO11 reasons about motion in a simplified domain called the “Bouncing Ball” world. A situation in the Bouncing IjJl world consists of a two dimensional scent with surfaces represented by line scgmcnts, and one or more balls which are mod&d as point masses. WC ignore the exact shape of balls, motion after two balls collide, spin, motion in a third spatial dimension, air resistance, sliding motion, and all forces other than gravity. The initial description of a situation is a diagram containing a description of the surfaces and one or more bails, as in ligure 1. Fig. 1. A typical sccnc from the Rouncing Ball world A silualion in the Bouncing Ml World consisls or a diagram that q.x~fies surfaces and one or more balls This drawing only shows lhe geometric aspects of the descriptions involved. When given a description of a situation in the Ijouncing hall world, I-ROB analyzes the surface geometry and computes qualitative descriptions of the free space in the diagram. ‘l’hc person using the program can describe bails. properties of their states of motion, rcqucst simulations, and rnakc global assumptions about the motion. FROl? incrcmcntally crcatcs and updates its descriptions to accommodate this information, complaining if inconsistencies are dctcctcd. Questions may bc asked by calling procedures that intcrrogatc thcsc descriptions. The four basic questions FROR can answer are: (1) What can it (a ball) do next?, (2) Where can it go next?, (3) Where can it end up?, and (4) Can thcsc two balls collide? ,Ll Spatial descriptions WC do not yet know why people are so good at reasoning about space. Thcorcm proving and symbolic manipulation of algebraic expressions do not seem to account for this ability. Arguments against the former mdy bc found in [l], while the sheer complexity of algebraic manipulations argues against the latter. I conjccturc that the fluency pcoplc exhibit in dealing with space comes mainly from using their visual apparatus. One example is the 170 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. USC of diagrams. Thcmarks in a diagram reflect the spatial relations bctwcen the things they rcprcscnt. which allows us to use our visual apparatus to interpret these relationships as WC would with real objects. In this case, perception provides a simple (at least for the processes that use it) decision procedure for a class of spatial questions. We do not yet understand the complcxitics of human vision, but the techniques of analytic gcomctry can be used to provide decision proccdurcs for geometrically simple cases. FROR uses a Metric Diagram. which is a representation of geometry that combines symbolic /and numerical information. The geometric aspects of a problem are rcprcscntcd by symbolic elements whose parameters are numbers in ;I bounded global coordinate system. The rcprescntation is used to answer questions about basic spatial relationships between elements, such as which side of a lint a particular point lies or whcthcr or not two lines touch. Calculation based on properties of the clcmen ts su ffkes to answer these questions. My conjecture about qualitative spatial reasoning is that it involves a vocabulary of PLACES whose relationships are described in symbolic terms. By PLACE, I mcan a piece of space (point, line, region, volume, etc.) such that ail parts of it share some property. The nature of a domain determines the notion of place appropriate to it. In FROII the Space Graoh provides the vocabulary of places. Since all balls arc point masses and arc subject to the same forces, the Space Graph is independent of them and dcpcnds only on the surface geometry in the Metric Diagram. Free space is divided into regions in a way that insures the description of qualitative state (dcscribcd below) will bc unique and simple, and these regions and the edges that bound them are the n_odcs of the Space Graph. These nodes are connected by arcs that are labclled with the name of the relationship between them (such as IXFT or UP). Any other place rcquircd for qualitative reasoning can be dcscribcd by composing these nodes, and the graph structure provides a framework for cfficicnt processing (see section 4). An example of the places in a scene and the graph structure they produce is contained in figure 2. Pointu- Fig. 2. Space Graph for a scene The free space in the diagram is broken up into regions in a way that simplifies the description of the kinds of motion possible. The labels on the pointers which indicate the spatial relationships between the nodes are not shown due to lack of space. III Describing a Particular Motion When we watch an object move, we generally couch OUT description in terms of a sequence of qualitatively distinct motion types. WC will call a network built from descriptions of motions linked by descriptions of the state of the object before and after each motion an Action SCQUCI~. ‘I’he knowledge associated with each type of motion allows it to IX further analy?cd, the consistency of the proposed description to be checked, and permits making predictions about what will happen next. A drawn trajectory of motion in the Bouncing Ml domain and the schema of its associated Action Sequence is illustrated in figure 3. Fig. 3. Action Sequence Schema for Bouncing Balls This schema describes the motion depicted in Figure 1 The PilYSOR constrarnt dcscribcs Lhc state of Ihc ball al some instant in time. and the ACT constraints dcscrlbe a piece of the ball’s history. l’hc two basic types of motion in the Bouncing Hall world are FLY and COLI.II)F.. The difference in computing boundary conditions between flying up and flying down rcquircs their consideration as separate acts in the sequence, and additional motion types are defined for transformations to motions outside the Bouncing I\all world (such as CONTINUE for leaving the diagram and SI,IDE/S’I‘OP when a ball is travclling along a surface). The description of a ball’s state includes such information as its velocity (quantitative if known, or just in terms of a rough heading like (LEFT UP)) and what it is touching. In FROB the Action Sequence descriptions are embedded in a constraint language (SW [2] for an overview), and include equations describing projectile motion to compute numerical values if numerical descriptions of the state parameters are obtained. The use of quantitative parameters in the qualitative description of motion makes possible a diffcrcnt kind of simulation from the usual incremental time simulations used in physics. When numbers are provided, an Action Sequence can bc produced by gcncrating a description of the next motion from the last known state of motion. The time to gcncrate the description, as well as the complexity of the result, depends on the qualitative complexity of the motion rather than some fixed increment of time used to evolve a set of state parameters. Simulation is not the only way an Action Sequence can be created. A network of constraints can bc built to describe some proposed motion and the knowlcdgc of the equations of motion can be used to analyze it to see if it is consistent. ‘The dependence on quantitative paramctcrs in FliOl~s analysis is a drawback. For example, FROB can detect that the situation -in figure 4 is 171 inconsistent only after being given some final height for the ball and a value for the elasticity. People can argue that this proposed motion is impossible with simpler arguments that require less information. Fig. 5. A Sequence Graph The arrows represent the Arcction of a quahtative state at the place the arrow is drawn Circles represent states without well defined direclions. The pomters expressing the possible temporal orderings of the states are not shown. Fig. 4. An inconsistent description of motion This motion is impossible hecause the ball could not get as high as it does after the second collision unless it had gone higher on the tirst. If it had gone higher after the first, the second collision would not even have happened ‘To discover that this description is inconsistent l:KOl3 requires a specific re!ocity a( the highest point and a specific value for the elasticity of the ball as well as the coordinates of the collision points. The basic idea of an Action Sequence stems highly suited as a target representation for parsing quantitative data about motion, perhaps gleaned by perception. For this purpose a more qualitative set of methods for analysis would havP to bc encoded. An example of such a rule for the Bouncing Ball domain would be “A ball cannot increase its energy from one act to the next”. E Describing Possible Motions The quantitative state of a ball consists of its position and velocity. A notion of qualitative state can be defined which generalizes position to be a PLACE, the velocity to be a symbolic heading (such as (RIGHT DOWN)), and makes explicit the type of motion that occurs. A set of simulation rules can be written to operate on qualitative state descriptions, but because of the ambiguity in the description the rules may yield several motions possible from any given state. Since there arc only a small number of places and a small number of possible qualitative states at each place, all the possible kinds of motion from some given initial qualitative state can easily bc computed. This description is called the cnvisionmcnf (after [3]) for that state. dcKleer used this description to answer simple questions about motion directly and plan algebraic solutions to physics problems. In FROB envisioning results in the Scclucncc Graph, which uses the Space Graph for its spatial framework (see Figure 5). It is ----lnll&lstate- (FLY (SREGIONl) (LEFT Knowing more about a ball than its state of motion at some time can restrict the kinds of motion possible to it. Energy limits the height a ball can reach, and knowing that a ball is pcrfcctly elastic or completely inelastic excludes certain results of a collision. Assumptions about whether a ball must or may not reach a particular place or qualitative state can restrict the possibilties as well. The Scqucncc Graph can bc modified by pruning states to reflect this information about the ball and its motion. of Fach of the constraints above directly rules out some states mot&n. The full consequences of eliminating such states are determined by methods that rely on specific properties of space and motion. Among these properties arc the fact that a mo:ion of an object must be “continuous” in its state path (which means that the active part of a Sequence Graph must be a single connected component) and that the space it moves in must bc connected (which is uscfi11 because there are many fcwcr places than qualitative states in any problem). Dependency information is stored so that the cffccts of specific assumptions may be traced. Conflicting assumptions, overconstraint. and conflicts bctwcen a description of the actual motion (as spccificd by an Action Sequence) and its constrained possibilties arc dctcctcd by FROB and the underlying assumptions arc offcrcd up for inspection and possible correction. v Answering Oucstions Many of the questions that could be asked of the Bouncing Ball domain can bc answered by direct examination of the descriptions built by FROB. These include questions (1) and (2) above. The three levels of description of motion in FROB (the Action Sequence, the Sequcncc Graph, and the path of qualitative states corresponding to the Action Sequence) allow some kind of answer to be given even with partial information. used for summarizing propcrtics of the long term motion of an object, evaluating collision possibilties, and,assimilating assumptions More complicated questions (such as (3) and (4) above) can about the global propcrtics of motion. Only the assimilation of be answer-cd with additional computation using these descriptions. Determining whcthcr or not a ball is trapped in a well (see figure 6) assumptions will be discussed here. can be done by examining a Scqucnce Graph for the last state in an 172 Fig. 6. Summarizing Motion ->>(motion-summary-for bl) FOR Bl TlIE BALL WILI. I:VI:NTUAI 1;Y STOP II‘ IS TRAPPED INSIDE (WELID) AND IT WILL S I’OP FLYING IZT ONE OF (SEGMENTll) Action Sequence to see if it is possible to bc moving outside the places that comprise the well. Often a collision between two balls can be ruled out because the two balls arc never in the same PI.ACE, as determined by examining their Scqucncc Graphs. With the Action Sequence description of motion it is possible to compute exactly where and when two balls collide if they do at all. Figure 7 contains the answers given by the progratn to collision questions in a simple situation. Fig. 7. Collision Problems ->>(collide? f g) (I’OSSIRI 1:. AT SEGMEN l-13 SREGlONl .,,) ->>(rannot-bc-at f segment31) (SIiGhlENT31) Ul’Dh I‘ING ASSUMPT10NS FOR (>> INlTlAl,-STATE 17) ClII:CKING l’ATli 01:MOIION ACAINSTASSUMPTIONS ->>(collide? f g) NO VI liclation to Other Work The focus of this work is very different from that of [4]!5][6], which arc mainly concerned with modcling students solving textbook physics problems. All of the problems dealt with in these programs were static. and the rcprcscntation of gcomctry expressed connectivity rather than free sp~c. lssucs such as gerting algebraic solutions and doing only the minimal amount of work rcquircd to answer a particular question about a situation were ignored here in order to bcttcr deal with the questions of spatial reasoning and the semantics of motion. The process of formalizing common sense knowledge is much in the spirit of the Naive Physics effort of Hayes (described in [7]). The Action Sequence, for example, may bc vicwcd as the history for a ball since it contains explicit spatial and temporal limits. However, this work is conccrncd with computational issues as well as issues of rcprcscntation. Unlike this work, Hayes (see [8]) explicitly avoids the use of metric rcprcscntations for space. 1 suspect that a metric representation will bc required to make the concept of a history useful, in that to compare lhcm rcquircs having a common coordinate frame. The Metric Dingrdm has much in common with the dcrcriptions used as targets for langu,tgc translation of [I] and the imagery theory of [9]. Arguments against the traditional “pure relational” gcomctric representations used in A1 and the “naive analog” rcprcscntations used by [lc)],[l l] may bc found in [12]. The concept of envisioning was first introduced in [3] as a technique for answering simple questions about a sccnc directly and as a planning device for algebraic solutions. ‘I’hc inclusion of dissipative forces, a true IWO dimensional domain, interactions of more than one moving object, and its USC in assimilation of global constraints on motion arc the envisioning advances incorporated in this work. vu Biblioprauhy [I] Waltz, D. and Boggess, L. “Visual Analog Representations for Natural Language Understanding” in Proc. IJCAI-79 Tokyo, Japan, August 1979 -- 121 Steele, G and Sussman. G “Constraints” Memo No. 502. MIT AI Iab. Cambridge, Massachusetts, November 1978 [3] deKleer, Johan “Qualitative and Quantitative Knowledge in Classical Mechanics” Technical Report 352. MIT AI lab, Cambridge, Massachusetts. 1975 [4) Bundy, A. et. al “MECllO:Year One” Research Report No. 22, Department of Artificial Intelligence. FAinburgh, 1976 (51 Novak, G “Computer Understanding of Physics Problems Stated in Natural Language” Technical Report NL-30. Computer Science Department, The University of Texas at Austin, 1976 [6] McDermott, J and Larkin, J. “Re-representing Textbook Physics Problems” in Proc of the 2nd National Conference of the Canadian Society for Computational Studies of Intelhgence, Toronto 1978 (71 llayes, Patrick J “The Naive Physics Manifesto” unpublished, May 1978 [8] Hayes, Patnck J “Narve Physln 1:Ontology for Liquids” unpublished, August 1978 191 Hinton. G “Some Demonstrations of the Effects of Structural kcription~ in Mental Imagery” Conmtive Science, Vol 3. No. 3, July-September 1979 [JO] Funt, B V. “Wl1ISPER:A Computer implementation Using Analogues in Reasoning” PI1.D. Thesis, Umvcrslty of British Columbia, 1976 [ll] Kosslyn & Schwartz, “A Slmuiation of Visual Imagery” Connitive Science. Vol. 1, No. 3, July 1977 [12] Forbus. K. “A Study of @ualitative and Geometric Knowledge in Reasoning about Motion” MIT Al Lab T~IJJJC~ report, in preparation. 173

1980

44

39

COMPUTER INTERPRETATION OF HUMAN STICK FIGURES Martin Herman Department of Computer Science Carnegie-Mellon University Pittsburgh, PA 15213 ABSTRACT A computer program which generates context-sensitive descriptions of human stick figures is described. Three categories of knowledge important for the task are discussed: (1) the 3-D description of the figures, (2) the conceptual description of the scene, and (3) heuristic rules used to generate the above two descriptions. The program’s representation for these descriptions is also discussed. 1. Introduction This paper describes a computer program, called SKELETUN, which generates context-sensitive descriptions of 2-D, static, human stick figures. The motivating interest is to study the process of extracting information communicated by body postures. Stick figures have been chosen to approximate the human form because they eliminate the problems involved in processing fleshed-out human figures (e.g., extracting them from the image, identifying and labeling body parts), yet they maintain the overall form conveyed by gross body posture. SKELETUN currently operates in two domains, emotions (figures may be sad, happy, depressed, etc.) and baseball (batting, catching, etc.) Its knowledge of baseball is much more complete, however. It can accept any figure and interpret it in terms of the following baseball activi ties: (1) batting, (2) throwing, (3) running, (4) catching a high ball with one or both arms, (5) catching a ball at torso height with one or both arms, (6) fielding a grounder with one or both arms. An example of how a figure is interpreted in the baseball domain is shown in Fig. 1, where hand-generated English Q 1. Two arms catching torso-high ball 2. (very good confidence) Batting (fair confidence) 3. Two arms fielding grounder (poor confidence) Fig. la PHYSICAL DESCRIPTION The figure is in a vertical orientation with the feet below the head. The figure is facing left and the face is pointing left. The torso is bent forward. The elbow of arm1 is in- middle and down. It can be considered either as partly or half bent. The elbow of arm2 is in- middle and down. It can be considered either as partly or half bent. The knee of leg1 is forward and partly bent. The knee of leg2 is down and partly bent. The lower body is in a configuration similar to “feet well planted.” The figure can also be consider- ed in a diagonal orientation with the feet to the lower right of the head (but with lower confidence than vertical). In this case, it is facing lower left with the face pointing low- er left, The following then changes from the previous description: the elbow of arm1 can be considered as either down or forward. The knee of leg2 is forward. MEANING-BASED DESCRIPTION The figure is catching a ball at torso height with two arms, with very good confidence. It may also be viewed as batting, but with only fair confidence. Finally, it may be fielding a grounder with two arms, but with only poor confidence. Fig. lb ((vertical orientation) verygood) ((feet to bottom of head) verygood) ((facing left) good) ((face pointing left) good) ((torso bent forward) good) ((elbow1 is in-middle) good) ((elbow1 is down) verygood) ((elbow1 partly bent) good) ((elbow1 half bent) good) ((elbow2 is in-middle) good) ((elbow2 is down) verygood) ((elbow2 partly bent) good) ((elbow2 half bent) good) ((kneel is forward) verygooc ((kneel partly bent) good) ((knee2 is down) verygood) ((knee2 partly bent) good) ((both legs “feet well planted” cfig) good) ((diagonal orientation) good) ((feet to lowerright of head) good) ((facing lowerleft) fair) ((face pointing lowerleft) fair) ((elbow1 is down) good) ((elbow1 is forward) good) ((knee2 is forward) verygood) ((two-arms-catching-torso high-ball) verygood) ((batting) fair) ((two-arms-fielding- grounder) poor) 174 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. descripticns are shown alongside the computer-generated output. SKELETUN’s primary purpose is to generate a description of what is communicated by body posture - the “meaning-based” description. In the process of generating this description, it also provides the 3-D configuration of the figures - the physical descripticn. Briefly, the notation in the example is as follows. If a figure is viewed from the front or back, each elbow or knee can be either QI& from the torso, b to the torso (i.e., crossing the torso) or If the figure is in-middle (i.e., along the same line as the torso). viewed from the side, each elbow or knee can be either u, forward. backward, or back-uo (i.e., backward and up). All assertions in the dessriptlons have discrete confidence values. The input to SKELETUN is a hand-encoding of the x, y coordtnates of the end points of the line segments of each figure, plus the center of the circle representing the head. SKELETUN assumes that all figures are complete and valid, and that no objects other than figures are in the scene (a scene may have two figures). This paper gives an overview of the types of information conveyed by gross body postures, SKELETUN’s representation for this information, and some inference rules used to generate this information from 2-D scenes. See [7] for details. 1 .l Backaround This work views vision as a medium of communication, recognizing that an important goal of the visual process is to provide the viewer with a “meaning” description of the external world. Most scene analysis systems are primarily concerned with identifying objects and other entities in a scene and specifying the spatial configuration of these entities [6, 3, 8, 41. Given a scene with human figures, such systems would tend to identify the individual figures, their body parts, and other objects, and then specify the spatial relationships of these entities [9, 10, 11. SKELETUN goes one step further in the interpretation process. It tries to determine what the people are doing, and perhaps why they are doing it. Although some previous work has taken the point of view of vision as communication [2, 14, 151, their primary purpose was to analyze and describe motion scenes, rather than to study how body posture conveys information. 2. Knowledqe cateqories Five categories of knowledge have been identified as important in the process of generating descriptions of 2-D scenes of stick figures. The first three represent important levels at which the scene should be described. o Two-Dimensional Descriotion - a low-level description involving the direction of each body part (each is a straight line segment), the angle of each joint (in the 2-D plane), and body parts which overlap (required for establishing touching relationships). e Phvsical Soace Descriotion - a 3-D description of the physical configurations of the figures. o Meanina Soace Descriotion - a description in terms of the information communicated by the figures (e.g., running, fighting, crying). The concepts here are said to be in Meaning Space since “meaning” (or “conceptual” information) is extracted from the scene. The next two categories involve knowledge used to extract the physical and meaning space descriptions from the 2-D description. @ Human Phvsical Structure - information dealing with the various parts of the stick figure body and components of these parts. Q Inference Rules - heuristic rules used to obtain the -- 3-D configuration of the figures from the 2-D scene, and to determine what the figures are doing based on the 2-D and 3-D configurations of the limbs. The following sections will further discuss the 2nd, 3rd, and 5th categories. More details than can be provided here on all of the categories may be found in [7]. 3. Phvsical Soace Descriotion In order to infer what is being communicated by a figure’s body posture, there must be knowledge of at least part of its 3-D configuration, for it is a 2-D figure interpreted as being in 3-D space to which meaning is applied. It is convenient to have two different levels of physical space descriptions. One, called the lower level phvsical soace descrintion, deals with the 3-D positions of the individual body parts. The second, called the hiaher level phvsical soace clescriotion, deals with frequently occurring positions of groups of body parts. Only the first description will be discussed in this paper (see [7] for more details). Although a figure’s 3-D configuration may be represented many ways, the representation to be described next was chosen for two reasons: 1. Its purpose is to describe the figure in a manner useful for generating meaning-based interpretations. If the resolution is too fine (as in [lo]), it will contain much information not significant for the task, thus burdening the system. If the resolution is too coarse, it will not contain enough information to perform the task. 2. It is convenient for SKELETUN to be able to express a figure’s 3-D configuration in a manner easily understood by humans. The current representation makes this kind of information explicit. 3.1 Descriotions relative to the torso The 3-D descriptions in SKELETUN are object-centered, as opposed to viewer-centered. That is, locations and directions of parts of the figure are indicated with respect to the figure, rather than the viewer. A viewer-centered description depends not only on the figure being described, but also on its orientation. An object-centered description, however, depends only on the figure being described, resulting in a smaller set of possible descriptions [111. Accordingly, the positions of the upper arms and legs are represented relative to the torso, and the shape of the torso is represented relative to the overall orientation of the figure. SKELETUN uses the predicates OUT, IN-MIDDLE, and IN to describe the position of each elbow or knee as viewed from the 175 front, and UP, FORWARD, DOWN, BACKWARD, and BACK-UP to describe the positions as viewed from the side. These predicates are adequate to completely specify (within the resolution of the representation) the 3-D position of any elbow or knee (i.e., upper arm or leg). SKELETUN uses the predicates BENT-FORWARD and BENT-BACKWARD to specify how the torso joints are bent. 3.2 Hierarchv of obiect-centered c&.criotions The positions of the lower arms and legs are represented relative to the upper arms and legs, respectively. Note that a representation of the lower limbs relative to the torso would result in a much larger set of possible descriptions than a representation relative to the upper limbs, since a different description of the lower limb would be required for each position of the upper limb relative to the torso, even if the position of the lower relative to the upper limb were to remain constant. Since similar arguments apply to describing positions of other body parts, such as hands, fingers, feet, etc., we conclude that each body part should be represented relative to the part it is connected to, resulting in a hierarchy of descriptions [lo]. SKELETUN represents the positions of the lower arms and legs by specifying the 3-D angle of the elbow and knee joints. The predicates used are PARTLY-BENT, HALF-BENT, FULLY-BENT, and NOT- BENT. 3.3 Orientation relative to viewer --- Thus far, all descriptions have been relative to parts of the figure. The whole figure must also be placed in 3-D space, relative to the viewer. The predicate ORIENTATION describes the overall orientation of the figure either as vertical, horizontal, or diagonal. Given one of these orientations, the predicate DIR-OF-FEET-TO-HEAD specifies the direction of the feet relative to the head. Finally, the predicates DIR-FACING and DIR-FACE-IS-POINTING specify the direction the figure is facing and the direction the face is pointing. 3.4 Phvsical space inference rules These rules generate the physical space description. They are domain-independent, for they depend only on the 3-D configuration of the figures. As an example of the knowledge in these rules, consider how SKELETUN determines the overall orientation of the figure. A figure is horizontal if both feet are east or west of the head (as in lying). A figure is diagonal if both feet are southeast, southwest, northeast, or northwest of the head. There are two types of vertical orientations, upright and upside-down. (SKELETUN currently cannot handle upside-down figures.) Fig. 2 shows three extremes of upright figures. In Fig. 2a, both feet are south of the head. In Fig. 2b, both feet are not south of the head; the point midway between the feet is south of the head. In Fig. 2c, the midway point is not south of the head; only one foot is south of the head. Rules which determine whether a figure is upright must examine these three types of cases. For more details on these and other inference rules, see [7]. 4. Meaninq soace descriotion 4.1 Reoresentation Meaning space concepts in SKELETUN are not represented explicitly in terms of simpler concepts and relationships between them (as in Conceptual Dependency [13]), since SKELETUN’s (a) (5) Cc) Three upright stick figures. 2 Fig. concern is not to extract all the details of each concept. Instead, they are represented as labels (RUNNING, CRYING,WALKING, etc.), where the meaning is represented implicitly in terms of the inference rules which may assert the concept and those which may use the concept to assert other concepts. This is because SKELETUN’s concern is to discover and make use of relationships among concepts [12]. Two important classes cf information that can be extracted from the body postures of stick figures deal with (1) the physical states of the figures (running, walking, throwing, standing, etc.) and (2) the mental or emotional states of the figures (weeping, happy, thinking, etc.). Two types of physical states can be distinguished, active and passive. Active physical states involve activities requiring motion, such as running, dancing, or hitting. Passive physical states involve no motion; examples are standing, pointing, and watching. Mental-emotional states can also be divided into two categories, negative and positive. Negative states generally involve feelings or tendencies such as painful excitement, destruction, dullness, loneliness, discomfort, tension, incompetence, dissatisfaction, and helplessness (e.g., anger, sadness, apathy, panic, hate, grief, disgust). Positive states generally involve feelings or tendencies such as vitality, empathy toward others, comfort, and self-confidence (e.g., cheerfulness, enjoyment, happiness, hope, love, pride) [5]. The negative and positive states can each be further subdivided into passive and active. These will not be pursued here (see [7]). 4.2 Meaninq soace inference rules These rules generate the meaning space description. They tend to be domain-dependent, since most meaning-space concepts are applicable only in limited domains. As an example of the knowledge in these rules, consider how SKELETUN determines that a figure is fielding a grounder (assuming that the domain is baseball). (See Fig. 3 for exampies.) First, one or both arms must be in a “fielding grounder” configuration (a higher level physical configuration described in [7]). In addition, the lower body should be in a configuration similar to “kneeling on one knee” (Fig. 3b), “kneeling on both knees”, “feet well planted” (Fig. 3c), or “crouching” (Fig. 3d) [7] and the figure should be vertical. If the figure’s orientation is diagonal, its lower body should be in a “crouching” configuration and it must be facing either lower-left or lower-right. Finally, if both arms are in a 176 “fielding grounder” configuration and the figure is running, it is also fielding a grounder, i.e., running after a ground ball (Fig. 3a). Acknowledoement This research is part of the author’s Ph.D. thesis done at the University of Maryland, under the guidance of Chuck Rieger and Azriel Rosenfeld. The support of the National Science Foundation under Grant MCS-76-23763 is gratefully acknowledged, as is Mike Shneier for valuable comments, and Ernie Harris for help in preparing this paper. > (a) (b) Cd) Each figure is fielding a grounder. Fig. 3 References 1. Adler, M. Computer interpretation of Peanuts cartoons. Proc. SIJCAI, Cambridge, MA, 1977. 2. Badler, N. I. Temporal scene analysis: conceptual descriptions of object movements. Tech. Rept. 80, Dept. of Computer Science, University of Toronto, 1975. 3. Bajcsy, R., and Joshi, A. K. A partially ordered world model and natural outdoor scenes. In Computer Vision Systems, Hanson and Riseman, Ed.,Academic Press, 1978. 4. Barrow, H. G., and Tenenbaum, J. M. MSYS: A system for reasoning about scenes. Artificial Intelligence Center Technical Note 121, Stanford Research Institute, 1976. 5. Davitz, 1969. J. R. The Language of Emotion. Academic Press, 7. Herman, M. Understanding body postures of human stick figures. Tech. Rept. 836, Computer Science Center, University of Maryland, College Park, MD, 1979. 8. Levine, M. D. A knowledge-based computer vision system. In Computer Vision Systems, Hanson and Riseman, Ed.,Academic Press, 1978. 9. Marr, D., and Nishihara, H. K. Spatial disposition of axes in a generalized cylinder representation of objects that do not encompass the viewer. AIM 341, MIT, 1975. 10. Marr, D., and Nishihara, H. K. Representation and recognition of the spatial organization of three-dimensional shapes. AIM 416, MIT, 1977. 11. Nishihara, H. K. Intensity, visible surface, and volumetric representations. Workshop on the Representation of Three-Dimensional Oblects, Univ. of Pennsylvania, Philadelphia, PA, 1979. 12. Rieger, C. Five aspects of a full-scale story comprehension model. In Associative Networks: The Representation and Use of Know/edge in Computers, N. Findler, Ed.,Academic Press, 1978. 13. Schank, R. C. Identification of conceptualizations underlying natural language. In Computer Models of Thought and Language, Schank and Colby, Ed.,W. H. Freeman and Co., 1973. 14. Tsuji, S., Morizono, A., and Kuroda, S. Understanding a simple cartoon film by a computer vision system. Proc. HJCAI, Cambridge, MA, 1977. 15. Weir, S. The perception of motion: actions, motives, and feelings. Progress in Perception Research Report No. 13, Dept. of Artificial Intelligence, University of Edinburgh, 1975. 6. Hanson, A. R., and Riseman, E. M. VISIONS: a computer system for interpreting scenes. In Computer Vision Systems, Hanson and Riseman, Ed.,Academic Press, 1978. 177

1980

45

40

RESEARCH ON EXPERT PROBLEM SOLVING IN PHYSICS Gordon S. Novak Jr. and Agustin A. Araya Computer Science Department University of Texas at Austin Austin, Texas 78712 ABSTRACT Physics problems cannot in general be solved by methods of deductive search in which the laws of physics are stated as axioms. In solving a real physics problem, it is necessary to treat the problem as a "nearly decomposable system" and to design a method of analysis which accounts for the salient factors in the problem while ignoring insignificant factors. The analysis method which is chosen will depend not only on the objects in the problem and their interactions, but also on the context, the accuracy needed, the factors which are known, the factors which are desired, and the magnitudes of certain quantities. Expert problem solvers are able to recognize many frequently occurring problem types and use analysis methods which solve such problems efficiently. Methods by which a program might learn such expertise through practice are discussed. I INTRODUCTION We are investigating the cognitive processes and knowledge structures needed for expert-level problem solving in physics. We believe that physics is a particularly fruitful area for the investigation of general issues of problem solving, for several reasons. The laws of physics are well formalized, and there is a large set of textbook physics problems (often with answers and example solutions) available for analysis and for testing a problem solving program. The application of physical laws is considered to be well defined, so that both the method of analysis of a problem and the final answer can be judged as either correct or incorrect. At the same time, physics is considered to be a difficult subject; students find problem solving especially difficult, even when the equations which express the physical laws are available for reference. Although the laws of physics are "well known", nobody has yet produced a program which can approach expert-level problem solving in physics. Such a program would potentially have great value for applications, since the types of reasoning used in computer science and engineering are closely related to those used in solving physics problems. Such a -------------------- * This research was supported by NSF award No. SED-7912803 in the Joint National Institute of Education - National Science Foundation Program of Research on Cognitive Processes and the Structure of Knowledge in Science and Mathematics. PB SCHWM PRGE 25 NUMBER 19 [THE FOOT OF R LRDDER RESTS RCAINST fl VERTICRL WFlLL RN0 ON fl HORIZONTAL FLOOR) (THE TOP OF THE LROOER IS SUPPORTEO FROM THE WLL BY 17 HORIZONTRL ROPE 30 FT LONG) (THE LfXKlER IS SO FT LONG . WEIGHS 100 LB WITH ITS CENTER OF GRAVITY 20 FT FROM THE FOOT . f4NO F1 150 L8 MAN IS 10 FT FROM THE TOPl[DETERMINE THE TENSION IN THE ROPE1 FINSWER : 120.00000 LB Figure 1: Example of Output of ISAAC Program program could also be of value in education, because many of the crucial skills used in solving physics problems are now taught only implicitly, by example; students who are unable to infer the skills from the examples do poorly in physics. The first author has previously written a program which can solve physics problems stated in English in the limited area of rigid body statics [1,21; an example of its output is shown in Figure 1. This program, which uses a general formulation of the laws of rigid body statics (similar to the form of the laws presented in modern textbooks), produces between nine and fifteen equations for simple textbook problems for which human problem solvers generate only one or two equations. This somewhat surprising result indicates that the expert human problem solver does not slavishly apply the general forms of physical laws as taught in textbooks, but instead recognizes that a 178 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. particular problem can be solved by applying a special case of the more general law and writes only the equations appropriate for the special case. By doing so, the expert greatly reduces the algebraic complexity of the problem. II THE NATURE OF PROBLEM SOLVING IN PHYSICS -__I__----- Most people, even experts, tend to identify the "content" of physics with the equations which express the physical laws. Bundy [3] has written PROLOG programs in which the laws of physics are expressed as Horn clauses and deductive search is used to find answers to problems posed as sets of predicates. As noted by Larkin, McDermott, Simon, and Simon r41, novice problem solvers do tend to use an "equation-driven search", working backwards from the desired quantity until the equations they have invoked can be solved; experts, however, usually work "forward" from the given information until the desired unknown has been found. Experts do not often verbalize the equations that they are using, but usually only verbalize intermediate "answers". The identification of physics knowledge with the equations which express physical laws and the notion that search is the primary mechanism used in solving physics problems are unsatisfying because they fail to account for several observed phenomena. Why is physics hard? If physics were "only" the equations for the laws, these equations could be collected in a reference book (along with tables of integrals and such), and the physics course could be dispensed with. Why does practice help? What is expertise in physics, i.e., what does the expert have that the novice lacks that enables the expert to perform so much better? Superior algebraic skills alone cannot account for the difference. What is the intellectual content of a physics course? Without continued practice, students forget the equations soon after taking the course; what is it that they retain that makes taking the course worthwhile? We believe that methods which employ deductive search and express the laws of physics directly as predicate calculus clauses (01: the equivalent) cannot account for expert-level problem solving ability when a variety of physical principles are involved (say, the principles covered in a first-year college physics course). Indeed, the best ways of solving certain problems are self-contradictory if examined too closely. Consider the following problem (from [5], p. 67): A rifle with a muzzle velocity of 1500 ft/s shoots a bullet at a target 150 ft away. How high above the target must the rifle be aimed so that the bullet will hit the target? An "expert" solution to this problem might proceed as follows : "It takes the bullet 0.1 second to reach the target. During this time, the bullet falls distance d = (1/2)*g*t**2 or (1/2)*32*70.1)**2 ft, or 0.16 ft. So we aim'up by that amount to cancel the fall." In this solution, the "expert" has made several conflicting assumptions: first, that the bullet travels in a straight line; second, that the bullet falls from that path as it travels; and third, that the bullet is aimed upward to cancel the fall. Each succeeding assumption invalidates previous assumptions upon which it is based. In fact, the final answer is not exactly right; however, it differs from a more careful calculation for the parabolic path actually followed by the bullet by only about one part in a million. The "expert" has not solved this problem precisely, using all the applicable physical laws, but instead has treated the problem as a "nearly decomposable system" [61. Using qualitative knowledge that bullets move approximately in a straight line, the expert has decomposed the motion of the bullet into the dominant straight-line motion and the much smaller fall and upward motion. If we look harder, other decomposition assumptions can be found, viz. that air friction is negligible and that the earth is flat (i.e., that gravity is "straight down"). In fact, the laws of physics cannot be used directly in a deductive fashion to solve problems. For example, Newton's law, which we write compactly as "F = ma", relates the acceleration of a body (relative to an inertial reference frame) to the vector sum of all forces on the body; however, there are infinitely many such forces, and the frame of reference (e.go, the earth) isn't really inertial. Fortunately, in real problems most of the forces on the body are small and can be ignored, and the frame of reference is nearly inertial; using an appropriate decomposition of the problem, a good approximation of the answer to the problem can be found. Thus, solution of a real physics problem always involves treating a nearly decomposable system as if it were actually decomposable. Programs which use deduction to solve physics problems in a "microworld" are able to do so only because the decomposition decisions have been made, by the choice of the microworld and/or by the form in which the problem to be solved is encoded; however, this limits the extension of such a program to wider problem domains where other decompositions of "similar" problems are required. This view of problem solving in physics suggests answers to the questions posed earlier. Physics is hard because it is necessary to learn not only the equations, but also ways of decomposing actual problems so that application of the equations is tractable. The expert can solve problems better than the novice because the expert recognizes that a given (sub) problem is an instance of a class which can be decomposed in a particular way; this knowledge of how to decompose a real-world problem along lines suggested by the "fundamental concepts" of physics is a large part of what is sometimes called "physical intuition" r41* The knowledge of how to decompose problems may be retained by the student even though the formulas have been forgotten, and allows problems to be solved correctly even though the formulas have to be looked up again. The expert works forwards rather than backwards because the first 179 order of business is not to deduce the answer from the given information (which is likely to be grossly inadequate at.the beginning), but to find an appropriate decomposition or way of modeling the interactions of the objects in the problem. Once an appropriate decomposition has been found, solution of the problem is often straightforward. III RESEARCH ON PROBLEM SOLVING AND --~- LEARNING TO BE EXPERT --- We are currently writing a program to solve physics problems involving a variety of physical principles. Our work on this program is concentrating on the representation of problems, recognition of (sub) problem types which can be decomposed in particular ways, and learning of problem solving expertise through experience. Each of these areas is discussed briefly below. To insure that the problem solver is not told how to solve the problem by the manner in which the problem is stated to it, we assume that the input will be English or a formal language which could reasonably be expected as the output of an English parser such as the parser in [1,2]. For example, a car in a problem will be presented to the problem solver as "a car"; whether the car should be viewed as a location, a point mass, a rigid body, an energy conversion machine, etc. must be decided by the problem solver. We are developing a representation language which will allow multiple views of objects and variable levels of detail. For example, a block sliding on an inclined plane may be viewed as a weight, a participant in a frictional contact relation, and a location. A car might be viewed simply as a point mass, or as an object with its own geometry and complex internal structure, depending on the needs of a particular problem. The expert problem solver does not have a single, general representation of each physical law, but instead is able to recognize a number of special cases of each physical principle and use a rich set of redundant methods for dealing with them. For example, in addition to the general rigid body problem, the expert recognizes special cases such as a pivoted linear rigid body acted on by forces perpendicular to its axis. Such special cases often allow a desired unknown to be found using a single equation rather than many, or simplify analysis, e.g. by approximating a nonlinear equation with a linear one. Recognition of special cases is based on context, on what information is known, and on what answers are desired, as well as on the type of object or interaction. Our approach to recognition of special cases is to use a discrimination net, in which tests of features of the problem alternate with construction of new views of objects and collection of information into the appropriate form for a "schema" or "frame" representation of the view. Such a discrimination net can also be viewed as a hierarchical production system, or as a generalization of an Augmented Transition Network [71* If the recognition of the type of a (sub) problem is done by means of a discrimination net, special cases of a particular kind of physical system can be added to the net by adding discriminating tests for the special cases "above" the point in the net at which the more general case is recognized. We are investigating ways in which knowledge for handling such special cases could be learned automatically from experience in solving problems. One method of initiating such learning is data flow analysis of solutions using the more general solution. For example, if a problem involving a pivoted lever were solved using the general rigid body laws, data flow analysis would indicate that with a particular choice of the point about which to sum moments (the pivot), the "sum of forces" equations would not play a part in reaching the solution. A special case method could then be constructed from the general method by adding tests for the special case to the discrimination net and adding a corresponding "action" part which writes only the moment equation. Other opportunities for special case learning include elimination of zero terms (rather than eliminating them later algebraically), elimination of forces which turn out to be small when calculated, linearization of "almost linear" equations, use of small-angle approximations, and selection of simplified views of objects under appropriate conditions. REFERENCES 1. 2. 3. 4. 5. 6. 7. Novak, G. "Computer Understandi .ng of Physics Problems S tated in Na tural Language", American Journal of Computational Linguistics, Microfiche 53, 1976. Novak, G. "Representations of Knowledge in a Program for Solving Physics Problems", Proc. 5th IJCAI, Cambridge, Mass., Aug. 1977. Bundy, A., Byrd, L., Luger, G., Mellish, C. and Palmer, M. "Solving Mechanics Problems Using Meta-Level Inference", Proc. 6th IJCAI, Tokyo, 1979. Larkin, J., McDermott, J., Simon, D. P., and Simon, H. A. "Expert and Novice Performance in Solving Physics Problems", Science, vol. 208, No. 4450 (20 June 1980). Halliday, D. and Resnick, 5. Physics. ~- New York: Wiley, 1978. Simon, H. A. The Sciences of the Artificial, M.I.T. Press, 1969. - Woods, W. A. "Transition Network Grammars for Natural Language Analysis", comm. &J, vol. 13, no. 10 (Oct. 19701, PP. 591-606. 180

1980

46

41

KNOWLEDGE-BASED SIMULATION Philip Klahr and William S. Faught The Rand Corporation Santa Monica, California 90406 ABSTRACT Knowledge engineering has been successfully applied in many domains to create knowledge-based "expert" systems. We have applied this technology to the area of large-scale simulation and have implemented ROSS, a Rule-Oriented Simulation System, that simulates military air battles. Alternative decision-making behaviors have been extracted from experts and encoded as object- oriented rules. Browsing of the knowledge and explanation of events occur at various levels of abstraction. I. INTRODUCTION - Large-scale simulators have been plagued with problems of intelligibility (hidden embedded assumptions, limited descriptive power), modifiability (behaviors and rules buried in code), credibility (minimal explanation facilities), and speed (slow to build, to run, to interpret). The area of large-scale simulation provides a rich environment for the application and development of artificial intelligence techniques, as well as for the discovery of new ones. The field of knowledge engineering [2] is developing tools for use in building intelligent knowledge-based expert systems. A human expert communicates his expertise about a particular domain in terms of simple English-like IF-THEN rules which are then incorporated into a computer- based expert system. The rules are understandable, modifiable, and self-documenting. Knowledge-based systems provide explanation facilities, efficient knowledge structuring and sharing, and interfaces that are amiable for system building and knowledge refinement. Our approach to simulation views a decision- based simulator as a knowledge-based system. The behaviors and interactions of objects, the decision-making rules, the communiciation channels are all pieces of knowledge that can be made explicit, understandable, modifiable, and can be used to explain simulation results. For our work in simulation, we chose the domain of military air battles. Current large- scale simulators in this domain exhibit exactly the simulation problems we discussed above and thus provide a good area in which to demonstrate the feasibility and potential of knowledge-based simulation. II KNOWLEDGE REPRESENTATION -- .--- - Our domain experts typically centered their discussions of military knowledge around the domain objects. For example, a particular type of aircraft has certain attributes associated with it such as maximum velocity, altitude ranges, time needed to refuel, etc. Similarly, individual planes have given positions, speeds, altitudes, routes, etc. In addition, experts defined the behaviors of objects relative to object types or catagories. For example, they defined what actions aircraft take when they enter radar coverages, what ground radars do when they detect new aircraft (who they notify, what they communicate), etc. It became clear that an object-oriented (Simula-like [l]) programming language would provide a natural environment in which to encode such descriptions. We chose Director [6] for our initial programming language. Director is an object- oriented message-passing system that has been used primarily for computer graphics and animation. It offered considerable promise for our use in simulation, both in how knowledge is structured and how it is executed. Each object (class or individual) has its own data base containing its properties and behaviors. Objects are defined and organized hierarchically, allowing knowledge to be inherited, i.e., offsprings of objects automatically assume (unless otherwise modified) the properties and behaviors of their parents. In Director, as in other message-passing systems (e.g., Smalltalk [3] and Plasma [5]), objects communicate with each other by sending messages. The Director format for defining behaviors is (ask <object> do when receiving <message-pattern> <actions>), i.e., when the object receives a message matching the pattern, it performs the associated actions. In ROSS, we have added the capability of specifying IF-THEN rules of the form (IF <conditions> THEN <actions> ELSE <actions>) as part of an object's behavior. The conditions typically test for values (numeric, boolean) of data items while actions change data items or send messages to objects. Since Director is written in Maclisp, one may insert any Lisp s-expression as a condition or action. The following behavioral rule contains all of these options: 181 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. (ask RADAR do when receiving (IN RADAR RANGE ?AC) (SCRIPT (IF (lessp (length (ask MYSELF recall your OBJECTS-IN-RANGE)) (ask MYSELF recall your CAPACITY)) THEN (ask (ask (ask (ask ELSE (ask >>I HISTORIAN at ,STIME MYSELF detects ,AC) MYSELF add ,AC to your list of OBJECTS-IN-RANGE) ,(ask MYSELF recall your SUPERIOR) MYSELF detects ,AC) MYSELF monitor ,AC while in coverage) HISTORIAN at ,STIME MYSELF doesn't detect ,AC) This rule is activated when a radar receives a message that an aircraft (AC) is in its radar range. The radar tests whether the number of objects currently in its radar coverage is less that its capacity. If its capacity is not full, then the radar tells the historian that it detects the aircraft at time STIME (the current simulation time), it records the aircraft in its log, it notifies its superior that it detected the aircraft, and it continues to monitor the aircraft through its radar coverage. III -* BEHAVIORAL DESCRIPTIONS A difficult problem with large-scale simulators is that they contain complex behaviors that are hard to describe and understand. Each object has many potential actions it can take, and there may be hundreds of objects whose behavior the user may wish to examine and summarize at differing levels of generality and detail. To alleviate this problem, we organized the simulator's behavioral descriptions along two lines: 1. Static descriptions: descriptions of the major events simulated and the rules governing each object's potential actions. The events and rules are organized so that users can quickly peruse and understand the knowledge. 2. Dynamic descriptions: descriptions of each object's behavior as the simulator runs. Events are organized so the user is not overwhelmed by a mass of event reports. Events are reported as patterns with detail eliminated and selected events highlighted. To organize the descriptions, we constructed scenarios representing sequential event chains. Each major event in ROSS has an associated llevent descriptor" (ED). EDs are collected into chains, where each chain is a linear list of EDs. Each ED is causally associated with its immediate neighbors in the list: a preceding ED is necessary to cause its successor in that chain, but not necessarily sufficient. ([7] discusses the use of such chains in constructing proofs.) ED chains are further organized into Ilactivities," e.g., radar detection of an aircraft. Each activity is a tree of EDs. The root of the tree corresponds to the event that starts the activity. Each path from the root to a leaf is an ED chain. Logically, each ED chain corresponds to one possible scenario of events that could occur in a simulation. The scenario structure is used for both static and dynamic behavior descriptions. !!* BROWSING KNOWLEDGE Static behavior descriptions are given by ROSS's "browse" function, an on-line interactive facility with which users can examine ROSS's knowledge base. The user is initially given a list of all activities. He then selects an activity, and the browse function prints a list of the names of all EDs in that activity. The user can ask for a more complete description, in which case the browse function prints a tree of the EDs. The user can then select a particular ED to examine, and the browse function prints a simplified description of the event. If the user asks for a more complete description, the browse function prints the actual code for the corresponding behavior (as in the example above). At any point the user can select the next ED, or go up or down activity/event levels. The composition of ED chains, i.e., which EDs are members of which chains, is selected by the system developers for "naturalness" or appeal to a user's intuitive structure of the simulator. The system itself constructs the actual ED chains from the simulator's source code and a list of start and end points for each chain. The facility appears to be quite useful in our domain where the objects have interdependent yet self-determined behavior. B. EVENT REPORTING -- ROSS contains several objects that have been defined to organize, select, and report events to users. The Historian receives event reports from other objects (as exemplified in the behavioral rule above) and, upon request, supplies a history of any particular object, i.e., a list of events involving the object up to the current simulation time. The Historian also sends event reports to the Reporter, who selectively reports events to the user on his terminal as the simulator is running. The user can request to see all reports or only those involving a particular object or objects of a particular class. In addition, a Statistician accumulates statistics about particular fixed events (e.g., the total number of radar detections). (We have also interfaced ROSS to a color graphics system which visually displays simulation runs. The graphics facility has been an indispensable tool for understanding ROSS and the simulations it produces and for debugging.) C. EXPLANATION USING SCENARIOS - To explain "why" certain results occur, traditional rule-based systems typically show the 182 rules that were used, one by one, backchaining from the conclusions. We have taken a different approach to explanation in ROSS. Rather than displaying individual rules to explain events, ROSS presents higher-level behavioral descriptions of what happened, akin to the activity and event descriptions used for browsing. It is often the case that an event occurring in a simulaton run can be explained by specifying the chain of events leading up to the current event. This is accomplished simply by comparing event histories (gathered by the Historian) to the ED trees described above. The user can then browse the events and activities specified to obtain the applicable rules. What is perhaps more interesting in simulation is explaining why certain events do not occur. Often times simulations are run with expectations and the user is particularly interested in those cases where the expectations are violated. ([41 describes how such expectations can be used to drive a learning mechanism.) We have developed an initial capability for such "expectation-based" explanation. Explanations are given relative to specified ED chains. One event chain is designated as the "expected" chain. An "analyzer" reports deviations from this chain, i.e., it determines the point (event) at which the ED chain no lcnger matches the simulation events. Typical responses from the analyzer are: aircraft in radar range but not detected; radar sent message to superior but it was not received; command center requested aircraft assignment but none available. Such analysis can occur at any time within a simulation run to determine the current status of expected event chains. It is important to note that expectations need not be specifed prior to a simulation run (although this could focus the simulator's reporting activity). Users can analyze events with respect to any number of existing scenarios. Each analysis provides a simplified description and explanation of the simulator's operation from a different point of view (e.g., from radar's view, from aircraft's view, from a decision maker's view). This feature has also been extremely useful in debugging ROSS's knowledge base. IV* ---- SUMMARY AND FUTURE RESEARCH - ROSS currently embodies approximately 75 behavioral rules, 10 object types, and has been run with up to 250 individual objects. To show ROSS's flexibility, we have developed a set of alternative rule sets which encompass various strategies and tactics. Simulation runs using alternative rules show quite different behaviors and results. Future research will include scaling ROSS up both in complexity and in numbers of objects. Our goal is to turn ROSS into a realistic, usable tool. A more user-oriented English-like rule-based language will be required for users to express behaviors and strategies. We are looking toward ROSIE PI, or a hybrid ROSIE object-oriented language for this purpose. Scaling-up will necessitate enhancements in speed. We plan to explore parallel processing, abstraction (e.g., aggregating objects, adaptive precision), sampling, and focusing on user queries to avoid irrelevant processing. In summary, we have applied knowledge engineering to large-scale simulation and implemented ROSS, an interactive knowledge-based system that simulates the interactions, communications, and decision-making behavior within the domain of military air battles and command and control. We have shown the feasibility and payoff of this approach and hope to apply it to other domains in the future. ACKNOWLEDGMENTS We wish to thank Ed Feigenbaum, Rick Hayes-Roth, Ken Kahn, Alan Kay, Doug Lenat, Gary Martins, Raj Reddy, and Stan Rosenschein for their helpful discussions and suggestions during ROSS's development. We thank our domain experts Walter Matyskiela and Carolyn Huber for their continuing help and patience, and William Giarla and Dan Gorlin for their work on graphics and event reporting. REFERENCES 1. Dahl, O-J. and Nygaard, K. Simula -- an Algol-based simulation language. Communications ACM, 9 (9), 1966, 671-678. 2. Feigenbaum, E. A. The art of artificial intelligence: themes and case studies in knowledge engineering. Proc. IJCAI-77, MIT, 1977, 1014-1049. - -- 3. Goldberg, A. and Kay, A. Smalltalk- Instruction Manual, SSL 76-6, Xerox Palo Alto Research Center, 1976. 4. Hayes-Roth, F., Klahr, P., and Mostow, D. J. Knowledge acquisition, knowledge programming, and knowledge refinement. R-2540-NSF, Rand Corporation, Santa Monica, 1980. 5. Hewitt, C. Viewing control structures as patterns of passing messages. Artificial Intelligence, 8 (3), 1977, 323-364. 6. Kahn, K. M. Director Guide, AI Memo 482B, Arti .ficial Intel1 .igence Lab, MIT, 1979. 7. Klahr, P. Planning techniques for rule selection in deductive question-answering. In Pattern-Directed Inference Systems, D. A. Waterman -____ and F. Hayes-Roth (Eds.), Academic Press, New York, 1978, 223-239. 8. Waterman, D. A., Anderson, R. H., Hayes-Roth, F Klahr, P., Martins, G., and Rosenschein S. J. DeAign of a rule-oriented system for impleienting expertise. N-1158-l-ARPA, Rand Corporation, Santa Monica, 1979. 183

1980

47

42

I NTERACTI VE FRAME I NSTANTIATION Carl Engelman Ethan A. Scar1 Charles H. Berg* ABSTRACT This paper discusses the requirements that interactive frame instantiation imposes on constraint Verification. The representations and algorithms of an implemented software solution are presented. INTRODUCTION A number of frame representation languages or data access packages, seven of which are discussed in [STEFIK], have been developed as LISP extensions. In most applications of these languages, frame instantiation -- the creation of a new frame which represents an "instance", i.e., a more specific example, of a given "generic" frame -- occurs as a major theme. Yet, these languages do not really provide control structures sufficient to support interactive frame instantiation. Most of this paper will be concerned with constraint verification. A frame representation language typically provides the programmer with a way of attaching a constraint as a "facet" Of a given slot. It will reject any proposed values for the slot which fail that constraint, screaming a bit to the user. Such constraints attached to a slot in some generic frame also obtain automatically for slots of the same name occurring within its progeny. That is, they are "inherited". And that's about it. To explain why we felt the need for more control of constraint verification and, in fact, of the whole dynamics of interactive frame instantiation, we must present just a bit of our application. THE APPLICATION The KNOBS project [ENGELMAN] is directed towards the development of experimental consultant systems for tactical air command and control. We chose to focus first on what seemed to be a very simple type of aid. Imagine an Air Force officer is trying to plan a mission to strike some particular target. We are providing a program which interactively accepts the target, the airbase from which to fly the mission, the type of plane, the time of take-off, etc., and checks the input for inconsistencies and oversights. Such missions are stereotypes which itCurrent affiliation, AUTOMATIX Inc. The MITRE Corporation P.O.Box 208 Bedford, MA 01730 are represented naturally as frames and the checks are constraints among the possible slot values in such frames. DATA BASE/LANGUAGE SETTING k We first translated FRL [ROBERTSJULY] ROBERTSSEPT) f rom MACLISP to INTERLISP ERICSON]. Data bases of targets and resources ave been implemented as nets of individual and generic frames. An individual target frame, for example, contains information, e.g., location, specific to a particular target, while a generic target frame contains information true about classes of targets, for instance, the type of radar normally associated with a particular kind of surface-to-air missile. In all, the data base currently contains some 1400 frames. We have introduced several upwards compatible extensions to FRL, e.g., programmer controlled parallel inheritance along paths defined by a specified set of slot names and the controlled automatic invocation of "$IF-NEEDED" procedures during attempts to retrieve missing data. We split the concept of generic frame: those which we continue to refer to as "generic frames" contain information (defaults, attached demons, etc.) applicable to all their instances. Their slots are not necessarily in correspondence with those of their instances. What we refer to as a "template", on the other hand, is a prototypical representation of an instance of the associated generic frame. Its slots correspond to those to be found in an instance, but contain "$IF-NEEDED" procedures where the instance would contain values. It also contains constraints on the values that may appear in an instantiation. We also differentiate frames representing fully specified objects from those representing subclasses. The former is referred to as "instance" call it in [sTEFIK] and [SZOLOVITS], and we "individual". Such a frame is identified by having an "AIO" (An-Individual-Of) slot pointing back to its generic frame. AI0 corresponds to set membership, while AK0 (A-Kind-Of) corresponds to set inclusion. While the frame instantiation procedures are designed for use at any level, we have thus far employed them only in the creation of individual frames. DESIDERATA Our goal is a demonstration which, like a good shortstop, makes it look easy. Some requirements are: 184 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. 1. The system must know what information is Constraints needed and how-to ask for it. It must also know when the instantiation is complete and what to do then. 2. The user must be able to enter a value for'any slot in any frame at any time and the system must know what must be checked or rechecked and what remains to be done. 3. There should be a general facility to suggest choices for a slot which are consistent with the current values of the other slots. 4. The system should complain as soon as the slot values become inconsistent. It must show dynamic discretion in explaining enough, but not too much, of the difficulty. 5. The user must be able to ask questions about the data base and about the status of the instantiation. The form of a constraint is: (name domain expr), where "name" is an atom used to index a user-oriented explanation of the constraint, "domain" (the terminology is suggested in {STANSFIELD]) is the list of slot names whose interrelations are tested by the constraint, and "expr" is a predicate to be satisfied. "Expr" can refer to the current values of slots being instantiated simply by reference to their slot names. We define a bucket as an unordered list of constraints. The contents of the CONSTRAINTS slot in the template is a list of buckets, ordered to express priority. All constraints in the same bucket are of equal priority. The attachment of constraints to the template at the slot level, rather than the usual attachment to slots at the facet level, reflects our view that all the action is in the interaction of the slots and that it is presumptive to make a decision -- especially a static one -- as to which slot is "bad". 6. The instantiation of one frame must be able to initiate the instantiation of another. 7. Constraint satisfaction must be maintained after the original instantiation whenever a slot value is changed or a template is changed. EXECUTION Assuming a sequence of values is being suggested for the instantiated slots, the algorithm is as follows: Initially all constraints are unmarked. REPRESENTATION Templates Templates are represented by frames with slots whose names will be replicated in the instantiated frames and which contain either $IF-NEEDED or $VALUE facets. The $IF-NEEDED procedures are responsible for deciding whether the value can be computed or is to be requested from the user. It is our normal practice to have the $IF-NEEDED procedures also perform type checking. The presence of a SVALDE facet causes a recursive call to the instantiator. The template also contains two special slots, named CONSTRAINTS (discussed below) and BOOKEEEP. The BOOKKEEP slot contains procedures to be run upon completion of the frame instantiation, a sort of "IF-ADDED" mechanism at the frame level. The interpreter either steps through the template filling slots or fills those commanded by the user. Moreover, the user may interact at any time wi th LISP or with a natural language Q/A system for retrieving facts from the data base, including those inferred through "inheritance". The latter is implemented as an ATN parser, whose syntax and semantics are intimately related to the structure of the frame net. In add ition, there are a number of amenities: spelling correctors, synonym and word truncation recognizers, and facilities for viewing the current status of the instantiation process values, i.e. , a presentation of the current slot disti nguishing with explanation those At any time, there is a current slot name, the one for which a value has been most recently proposed. A constraint is considered timely if its domain includes the current slot name and if all the other slot names in its domain have values already assigned. The interpreter passes through the buckets in decreasing priority until it discovers a timely constraint. If the constraint fails, the interpreter marks the constraint and traps the slots in its domain, i.e., renders their values unknown to constraints in lower priority buckets, which are not tested. If these lower priority constraints are already marked, they become unmarked since they are no longer timely. Other failed constraints in the current bucket are marked and their domains trapped. If a previously marked constraint now succeeds, then it is unmarked. If all the constraints in a bucket having a given slot name become unmarked, the slot name is released, i.e., pushed down to lower priority buckets along with the current slot name. The process normally terminates when all the slots are filled and none of the constraints are marked. CONSISTENCY MAINTENANCE Should a slot value in an instantiated frame or a constraint in a template be changed, the system makes appropriate checks. COMPLEX CONSTRAINTS The discussion above deals with constraints on the related slots in a given frame. Such constraints are called simple in [STANSFIELD]. What he calls complex constraints, those which violate constraints. 185 involving slots in ditrerent frames, are of great importance to us. For example, we must be concerned with the timing constraints needed to synchronize a primary mission with its support missions (refueling, defense suppression, etc.). We are currently engaged in the design and implementation of suitable representations and algorithms. We believe that a purely recursive (depth first) sequence of frame instantiations is not acceptable, and that we shall have to provide flexible control of interleaved "co-instantiations". CHOICE GENERATION When cueing the user for a slot-value, we would often like to present a list of values consistent with those already chosen for other slots. This is, in general, computationally impossible. It turns out, however, to be in the nature of our application that we frequently can produce a list of consistent values. Furthermore, we can do this by a fairly general method, generating the choices, in fact, from the constraints. The key point -- and this is obviously application dependent -- is that many of our constraints are of the form (name (A Bl -- Bn) (MEMBER A (f00 Bl -- Bn))), where, if the code is to make sense, (foo Bl -- Bn) is a computable, finite list. So, for example, one constraint might mean: (1) The airbase is one of our airbases in Europe. and another might mean: the (2) The chosen chosen airbase. fighter wing is located at We sav a constraint enumerates a slot. S. + I I absolutely iff it is of the form: (name (S)(MEMBER S----)). Constraint (l), above, enumerates airbases absolutely. A constraint To make choice generation more efficient, we optimize the constraints within the current context by collecting those which are timely and "compiling" them into a function of only the current slot, essentially by pre-evaluating all subexpressions which do not contain this slot. CRITICISM AND FUTURE DIRECTIONS 1) We need to design and implement a comparable system for the complex constraints. 2) The only relative priorities we can express between constraints are static. This might, someday, prove inadequate. 3) There is danger of an existential trap. The simplest example occurs when the first slot filled is A and the candidates value satisfies every constraint whose domain is (A). There may, however, be a constraint whose domain is (A B) which cannot be satisfied with the proposed value of A and any value of B. Our interpreter does not see this until B is selected. The choice generation scheme discussed above could also be employed to test (perhaps very expensively), for such traps. ACKNOWLEDGEMENTS This work was supported by the Rome Air Development Center under Air Force contract F19628-80-C-0001. The FRL system was originally created at MIT by R. Bruce Roberts and Ira P. Goldstein. Roberts assisted us by defining and creating an "export nucleus" of FRL. The work reported here is built directly on the INTERLISP version of FRL translated and extended by Lars W. Ericson, while he was at MITRE. We are deeply indebted to him. REFERENCES LENGELMAN] Engelman, C., Berg, Charles H., and Bischoff, Miriam, "KNOBS: An Experimental Knowledge Based Tactical Air Mission Planning System and a Rule Based Aircraft Identification Simulation Facility", Proc. Sixth Inter. Joint Conf. Artificial Intelligence, Tokyo, 1979, pp. 247-249. [ERICSON] Ericson, Lars W., "Translation of Programs from MACLISP to INTERLISP", MTR-3874, The MITRE Corporation9 Bedford, MA, Nov. 1979. [ROBERTSJULY] Roberts, R. Bruce, and Goldstein, Ira P., "The FRL Primer", MIT AI Lab. Memo 408, July 1977. [ROBERTSSEPT) Roberts, R. Bruce, and Goldstein, Ira P., "The FRL Manual", MIT AI Lab. Memo 409, September 1977. [STANSFIELD] Stansfield, James L., "Developing Support Systems for Information Analysis", in Artificial Intelligence, An MIT Perspective, Winston, P. H., and Brown, R. H., (Eds.), The MIT Press, Cambridge, MA, 1979. [STEFIK] Stefik, Mark, "An Examination of a Frame-Structured System", Proc. Sixth Inter. Joint Conf. on Artificial Intelligence, Tokyo, 1979, pp. 845-852. [SZOLOVITS] Szolovits, P., Hawkinson, L. B., Martin, W. A., "An Overview of Owl, A Language for Knowledge Representation", MIT/LCS/TM-86, MIT, Cambridge, MA, June, 1977. 186

1980

48

43

DESCRIPTIONS FOR A PROGRAMMING ENVIRONMENT Ira P. Goldstein and Daniel G. Bobrow Xerox Palo Alto Research Center Palo Alto, California 94304, U.S.A Abstract PIE is an experimental personal information environment implemented in Smalltalk that uses a description language to support the interactive development of programs. PIE contains a network of nodes, each of which can be assigned several perspectives. Each perspective describes a different aspect of the program structure represented by the node, and provides specialized actions from that point of view. Contracts can be created that monitor nodes describing different parts of a program’s description. Contractual agreements are expressible as formal constraints, or, to make the system failsoft, as English text interpretable by the user. Contexts and layers are used to represent alternative designs for programs described in the network. The layered network database also facilitates cooperative program design by a wow, and coordinated, structured documentation. Int reduction In most programming environments, there is support for the text editing of program specifications, and support for building the program in bits and pieces. However, there is usually no way of linking these interrelated descriptions into a single integrated structure. The English descriptions of the program, its rationale, general structure, and tradeoffs are second class citizens at best, kept in separate files, on scraps of paper next to the terminal, or, for a while, in the back of the implementor’s head. Furthermore, as the software evolves, there is no way of noting the history of changes, except in some primitive fashion, such as the history list of Interlisp [lo]. A history list provides little support for recording the purpose of a change other than supplying a comment. But such comments are inadequate to describe the rationale for coordinated sets of changes that are part of some overall plan for modifying a system. Yet recording such rationales is necessary if a programmer is to be able to come to a system and understand the basis for its present form. Developing programs involves the exploration of alternative designs. But most programming environments provide little support for switching between alternative designs or comparing their similarities and d<fferences. They do not allow alternative definition,s of procedures and data structures to exist simultaneously in the programming environment: nor do they provide a representation for the evolution of a particular set of definitions across time. In this paper we argue that by making descriptions first class objects in’s programming environment, one can make life easier for the programmer through the life cycle of a piece of software. Our argument is based on our experience with PIE, a description-based programming environment that supports the design, development, and documentation of Smalltalk programs. Networks The PIE environment is based on a network of nodes which describe different types of entities. We believe such networks provide a better basis for describing systems than files. Nodes provide a uniform way of describing entities of many sizes, from small pieces such as a single procedure to much larger conceptual entities. In our programming environment, nodes are used to describe code in individual methods, classes, categories of classes, and configurations of the system to do a particular job. Sharing structures between configurations is made natural and efficient by sharing regions of the network. Nodes are also used to describe the specifications for different parts of the system. The programmer and designer work in the same environment, and the network links elements of the program to elements of the design and specification. The documentation on how to use the system is embedded in the network also. Using the network allows multiple views of the documentation. For example, a primer and a reference manual can share many of the same nodes while using different organizations suited to their different purposes. In applying networks to the description of software, we are following a tradition of employing semantic networks for knowledge representation. Nodes in our network have the usual characteristics that we have come to expect in a representation language--for example, defaults, constraints, multiple perspectives, and context-sensitive value assignments. There is one respect in which the representation machinery developed in PIE is novel: it is implemented in an object-oriented language. Most representation research has been done in Lisp. Two advantages derive from this change of soil. The first is that there is a smaller gap between the primitives of the representation language and the primitives of the implementation language. Objects are closer to nodes (frames, units) than lists. This simplifies the implementation and gains some advantages in space and time costs. The second is that the goal of representing software is simplified. Software is built of objects whose resemblance to frames makes them natural to describe in a frame-based knowledge representation. 187 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Perspectives Attributes of nodes are grouped into perspectives. Each perspective reflects a different view of the entity represented by the node. For example, one view of a Smalltalk class provides a definition of the structure of each instance, specifying the fields it must contain; another describes a hierarchical organization of the methods of the class; a third specifes various external methods called from the class; a fourth contains user documentation of the behavior of the class. The attribute names of each perspective are local to the perspective. Originally, this was not the case. Perspectives accessed a common pool of attributes attached to the node. However, this conflicted with an important property that design environments should have, namely, that different agents can create perspectives independently. Since one agent cannot know the names chosen by another, we were led to make the name space of each perspective on a node independent. Perspectives may provide partial views which are not necessarily independent. For example, the organization perspective that categorizes the methods of a class and the documentation perspective that describes the public messages of a class are interdependent. Attached procedures are used to maintain consistency between such perspectives. Each perspective supplies a set of specialized actions appropriate to its point of view. For example, the print action of the structure perspective of a class knows how to prettyprint its fields and class variables, whereas the organization perspective knows how to prettyprint the methods of the class. These actions are implemented directly through messages understood by the Smalltalk classes defining the perspective. Messages understood by perspectives represent one of the advantages obtained from developing a knowledge representation language within an object-oriented environment. In most knowledge representation languages, procedures can be attached to attributes. Messages constitute a generalization: they are attached to the perspective as a whole. Furthermore, the machinery of the object language allows these messages to be defined locally for the perspective. Lisp would insist on global functions names. Contexts and Layers All values of attributes of a perspective are relative to a context. Context as we use the term derives from Conniver [9]. When one retrieves the values of attributes of a node, one does so in a particular context, and only the values assigned in that context are visible. Therefore it is natural to create alternative contexts in which different values are stored for attributes in a number of nodes. The user can then examine these alternative designs, or compare them without leaving the design environment. Since there is an explicit model of the differences between contexts, PlEc can highlight differendes between designs. PIE also provides tools for the user to choose or create appropriate values for merging two designs. Design involves more than the consideration of alternatives. It also involves the incremental development of a single alternative. A context is structured as a sequence of layers. It is these layers that allow the state of a context to evolve. The assignment of a value to *a property is done in a particular layer. Thus the assertion that a particular procedure has a certain source code definition is made in a layer. Retrieval from a context is done by looking up the value of an attribute, layer by layer. If a value is asserted for the attribute in the first layer of the context, then this value is returned. If not, the next layer is examined. This process is repeated until the layers are exhausted. Extending a context by creating a new layer is an operation that is sometimes done by the system, and sometimes by the user. The current PIE system adds a layer to a context the first time the context is modified in a new session. Thus, a user can easily back up to the state of a design during a previous working session. The user can create layers at will. This may be done when he or she feels that a given groups of changes should be coordinated. Typically, the user will group dependent changes in the same layer. Layers and contexts are themselves nodes in the network. Describing layers in the network allows the user to build a description of the rationale for the set of coordinated changes stored in the layer in the same fashion as he builds descriptions for any other node in the network. Contexts provide a way of grouping the incremental changes, and describing the rationale for the group as a whole. Describing contexts in the network also allows the layers of a context to themselves be asserted in a context sensitive fashion (since all descriptions in the network are context-sensitive). As a result, super-contexts can be created that act as big switches for altering designs by altering the layers of many sub-contexts. Contracts and Constraints In any system, there are dependencies between different elements of the system. If one changes, the other should change in some corresponding way. We employ contracts between nodes to describe these dependencies. Implementing contracts raises issues involving 1) the knowledge of which elements are dependent; 2) the way of specifying the agreement; 3) the method of enforcement of the agreement; 4) the time when the agreement is to be enforced. PIE provides a number of different mechanisms for expressing and implementing contracts. At the implementation level, the user can attach a procedure to any attribute of a perspective, (see [2] for a fuller discussion of attached procedures): this allows change of one attribute to update corresponding values of others. At a higher level, one can write simple constraints in the description language (e.g. two attributes should always have identical values), specifying the dependent attributes. The system creates attached procedures that maintain the constraint. There are constraints and contracts which cannot now be expressed in any formal language. Hence, we want to be able to express that a set of participants are interdependent, but not be required to give a formal predicate specifying the contract. PIE allows us to do this. Attached procedures are created for such contracts that notify the user if any of the participants change, but which do not take any action on their own to maintain consistency. Text can be attached to such informal contracts that is displayed to the user when the contract is triggered. This provides a useful inter-programmer means of communication and preserves a failsoft quality of the environment when formal descriptions are not available. Ordinarily such non-formal contracts would be of little interest in artificial intelligence. They are, after all, outside the comprehension of a reasoning program. However, our thrust has been to build towards an artificially intelligent system through succcessive stages of man-machine symbiosis. This 188 approach has the advantage that it allows us to observe human reasoning in the controlled setting of interacting with the system. Furthermore, it allows us to investigate a direction generally not taken in Al applications: namely the design of memory-support rather than reasoning-support systems. An issue in contract maintenance is deciding when to allow a contract to interrupt the user or to propagate consistency modifications. We use the closure of a layer as the time when contracts are checked. The notion is that a layer is intended to contain a set of consistent values. While the user is working within a layer, the system is generally in an inconsistent state. Closing a layer is an operation that declares that the layer is complete. After contracts are checked, a closed layer is immutable. Subsequent changes must be made in new layers appended to the appropraiate contexts. Coordinating designs So far we have emphasized that aspect of design which consists of a single individual manipulating alternatives. A complementary facet of the design process involves merging two partial designs. This task inevitably arises when the design process is undertaken by a team rather than an individual. To coordinate partial designs, one needs an environment in which potentially overlapping partial designs can be examined without overwriting one another. This is accomplished by the convention that different designers place their contributions in separate layers. Thus, where an overlap occurred, the divergent values for some common attributes are in distinct layers. Merging two designs is accomplished by creating a new layer into which are placed the desired values for attributes as selected from two or more competing contexts. For complex designs, the merge process is, of course, non-trivial. We do not, and indeed cannot, claim that PIE eliminates this complexity. What it does provides is a more finely grained descriptive structure than files in which to manipulate the pieces of the design. Layers created by a merger have associated descriptions in the network specifying the contexts participating in the merger and the basis for the merger. Meta-description Nodes can be assigned meta-nodes whose purpose is to describe defaults, constraints, and other information about their object node. Information in the meta-node is used to resolve ambiguities when a command is sent to a node having multiple perspectives. One situation in which ambiguity frequently arises is when the PIE interface is employed by a user to browse through the network. When the user selects a node for inspection, the interface examines the meta-node to determine which information should be automatically displayed for the user. By appropriate use of meta-information, we have made the default display of the PIE browser identical to one used in Smalltalk. (Smalltalk code is organized into a simple four-level heirarchy, and the Smalltalk browser allows examination and modification of Smalltalk code using this taxonomy.) As a result, a novice PIE user finds the environment similar to the standard Smalltalk programming environment which he has already learned. Simplifying the presentation and manipulation of the layered network underlying the PIE environment remains an important research goal, if the programming environment supported by PIE is to be useful as well as powerful. We have found use of a meta-level of descriptions to guide the presentation of the network to be a powerful device to achieve this utility. Conclusion PIE has been used to describe itself, and to aid in its own development. Specialized perspectives have been developed to aid in the description of Smalltalk code, and for PIE perspectives themselves. On-line documentation is integrated into the descriptive network. The implementors find this network-based approach to developing and documenting programs superior to the present Smalltalk programming environment. A small number of other people have begun to use the system. This paper presents only a sketch of PIE from a single perspective. The PIE description language is the result of transplanting the ideas of KRL [2] and FRL [6] into the object oriented programming environment of Smalltalk [8], [7]. A more extensive discussion of the system in terms of the design process can be found in [l], and [4]. A view of the PIE description language as an extension of the object oriented programming metaphor can be found in [5]. Finally, the use of PIE as a prototype office information system is described in [31. References ill PI [31 [41 151 161 [71 WI WI [lOI Bobrow, D.G. and Goldstein, I.P. “Representing Design Alternatives”, Proceedings of the A/S6 Conference, Amsterdam, 1980. Bobrow, D.G. and Winograd, T. “An overview of KRL, a knowledge representation language”, Cognifive Science 1, 1 1977. Goldstein, I.P. “PIE: A network-based personal information environment”, Proceedings of the Office Semantics Workshop, Chatham, Mass., June, 1980. Goldstein, I.P. and Bobrow, D.G., “A layered approach to software design “, Xerox Palo Alto Research Center CSL- 80-5. 1980a. Goldstein, I.P. and Bobrow, D.G., “Extending Object Oriented Programming in Smalltalk”, Proceedings of the Lisp Conference. Stanford University, 1980b. Goldstein, I.P. and Roberts, R.B. “NUDGE, A knowledge- based scheduling program”, Proceedings of the Fifth International Joint Conference 0” Artificial Intelligence, Cambridge: 1977, 257-263. Ingalls, Daniel H., “The Smalltalk- Programming System: Design and Implementation,” Conference Record of the Fifth Annual ACM Symposium on Principles of Programming Languages, Tucson, Arizona, January 1978, pp 9-16. Kay, A. and Goldberg, A. “Personal Dynamic Media” /EEE Computer, March, 1977. Sussman, G., & McDermott, D. “From PLANNER to CONNIVER -- A genetic approach”. Fall Joint Computer Conference. Montvale, N. J.: AFIPS Press, 1972. Teitelman, W., The lnterlisp Manual, Xerox Palo Alto Research Center, 1978. 189

1980

49

44

A BASIS FOR A THEORY OF PROGRAM SYNTHESIS’ P.A.Subrahmanyam USC/Information Sciences Institute and Cepartment of Computer Science University of Utah, Salt Lake City, Utah 84112 1. Introduction and Summary In order to obtain a quantum jump in the quality and reliability of software, it is imperative to have a coherent theory of program synthesis which can serve as the basis for a sophisticated (interactive) software development tool. We argue that viewing the problem of (automatic) program synthesis as that (automatically) synthesizing implementations of abstract data types provides a viable basis for a general theory of program syntheeis. We brlefly describe the salient features of such a theory [5, Sj, and conclude by listing some of the applications of the theory. 1.1. Roquiromonts for en Accrptable Theory of Program Synthesis. We view some of the essential requirements of en acceptable theory of program synthesis to be the following: - the theory should be general; - it must adhere to Q coherent set of underlying principles, and not be based on an ad hoa collection of heuristics; - it must be based on a sound mathematical framework; - it must account for the “state of the art” of program synthesis; in particular, it must allow for the generation of “efficient” programs. Further, if a theory is to be useful, we desire that it possess the following additional attributes: - it should serve as the basis for a program development system which can generate provably correct non-trivial programs; - it should possess adequate flexibility to admit being tailored to specific application tasks1 - it should provide new and insightful perspectives into the nature of programming and problem solving, 1 This work warn rupportrd in prrt by l n IBM Followskip With these requirements in mind, we now examino the nature of the programming process in an attempt to characterize the basic principles that underly the conttructlon of “good” programs. 2 A basis for a Theory of Program Synthesis ---_ Intuitively, the abstraction of a problem can be viewed as consisting of an appropriate set of functions to be performed on an associated set of objects. Such a collection of objects and functions is an “abstract data typo” and has the important advantage of providing a representation independent characterization of a problem. Although the illustrations that most readily come to mind are commonly employed data structures such as a stack, a file, a queue, a symbol table, etc., any partial recursive function can be presented as an abstract data type. Programming involves representing the abstractions of objects and operations relevant to a given problem domain using primitives that are presumed to be already available; ultimately, such primitives are those that are provided by the available hardware. Various programming methodologies advocate ways of achieving “good” organizations of layers of such representations, in attempting to provide an effective means of coping with the complexity of programs. _ there exists, therefore, compelling evidence in favor of viewing the process of program synthesis as one of obtaining an implementation for the data type correponding to the problem of interest (the “type of interest”) in terms of another data type that corresponds to some representation (the “target type.“) This perspective is’ further supported by the following basic principles that we think should underly the synthesis process (if reliable programs are to be produced consistently): 1. The programming process should essentially be one of program synthesis proceeding from the spocificrtions of a problem, rather than being primarily analytic (e.g. constructing a program and then verifying it) or empirical (e.g. constructtng a program and then testing it). 74 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. 2. The specification of a problem should be representation independent. This serves to guarantee complete freedom in the program synthesis process, in that no particular program is excluded a priori due to overspecification of the problem caused by representation dependencies. 3. The synthesis should be guided primarily by the semantics of the problem specification. 4. The level of reasoning used by the synthesis paradigm should be appropriate to “human reasoning,” rather than being machine oriented (see [6]). In addition to making the paradigm computationally more feasible, this has two major advantages: a. existing paradigms of programming such as “stepwise refinement” can be viewed in a mathematical framework b. user interaction with the system becomes more viable, since the level of reasoning is now “visible” to the user. The above principles led us to adopt an algebraic formulation for the development of our theory [l, 21, [3, 43. An important consequence of this decision was that the synthesis paradigm is independent of any assumptions relating to the nature of the underlying hardware. In fact, it can even point to target types suited to particular problems of interest i.e. trrget machine architectures which aid efficient implementations. 3. the Proposed Paradigm for Prowam Synthesis We adopt the view that any object representing an instance of a type is completely characterized by its “externally observable behavior”. The notion of an implementation of one data type (the type of interest) in terms of another (the target type) is then defined as a map between the functions and objects of the two types which preserves the observable behavior of the type of interest. The objective, then, is to develop methods to automate the synthesis of such implementations based on the specifications of the type of interest and the target type. Intuitively, the crux of the proposed paradigm lies in “mathematically” incorporating the principle of stepwise refinement into automatic programming. This is done by appropriately interpreting both the syntactic and semantic structure inherent in a problem. An important distinction from most transformation based systems is that the refinem@rnt is guided by the semantics of the functions define on the type of interest, rather than by a fixed set of rules (e.g. [7]>. An formal characterization of some of the pivotal steps in the synthesis process is provided, and an attempt is made to pin-point those stages where there is leeway for making alternative choices based upon externally imposed requirements. (An example of such a requirement is the relative efficiency desired for the implementations of different functions depending upon their relative frequency of use.) This separation of the constraints imposed by (a) the structure inherent in the problem specification, (b) the requirements demanded by the context of use, and (c) the interface of these two, serves to further subdivide the complexity of the synthesis task -- it becomes possible now to seek to build modules which attempt to aid in each of these tasks in a relatively independent manner. In summary, our goal was to seek, in as far as is possible, a mathematically sound and computationally feasible theory of program synthesis. The formal mathematical framework underlying our theory is algebraic. The programs synthesized are primarily applicative in nature; they are provably correct, and are obtained without the use of backtracking. There is adequate leeway in the underlying formalism that allows for the incorporation of different “environment dependent” criteria relating to the “efficiency” of implementations. The objectives of the theory include that conventional programs be admitted as valid outcomes of the proposed theory, This is in consonance with our belief that any truly viable theory of synthesis should approximate as a limiting case already existing empirical data relevant to its domain. 4. An Example: The Synthesis of Block Structured Symbol Table Using an Indexed Array To illustrate some aspects of the paradigm for program synthesis we outline the synthesis of a block-structured SymbolTable using an indexed array as a target type (cf. [4].) The sorts of objects involved are instances of SymbolTable, Identifier, Attributes, Boolean, etc.; our primary interest here is on the manipulation of instances of SymbolTables. The functions that are defined for manipulating a SymbolTable include: NEWST (spawn a new instance of a symbol table for the outermost scope,) ENTERBLOCK (enter a new local naming scope,) ADDID (add an identifier and associated attributes to the symbol table,) LEAVEBLOCK (discard the identifier entries from the most current scope, re-establish the next outer scope,) ISINBLOCK (test to see if an identifier has already been declared in the current block,) and RETRIEVE (retrieve the attributes associated with the most recent definition of an Identifier.) The formal specifications may be found in [4] (see also [6].) Although implementations for more complex definitions of SymbolTables have been generated (which include tests for “Global” Identifiers), we have chosen this definition because of its familiarity, The overall synthesis proceeds by first categorizing the functions defined on the type of interest (here, the SymbolTable) into one of the flowing three categories: 0) Base 75 Constructor functions that serve to spawn new instances of the type (e.g. NEWST); (ii) Canstructor functions that serve to generate new instances from existing ones (e.g. ADDID, ENTERBLOCK, LEAVEBLOCKh and (iii) E&&or functions that return instances of types other than SymbolTable (e.g. RETRIEVE, ISINBLOCK). The next step is to identify a subset of these functions (termed kernel functions) which serve to generate all instances of SymbolTables: these are NEWST, ADOID, and ENTERBLOCK. A major step in obtaining an implementation for the TOI is to provide an implementation for the kernel functions. Since no model for the kernel functions is explicit in the specification of a type, a suitable model must be inferred from the behavior of the functions defined on the type. Such an inference follows from an examination of the axioms defining the extraction functions. Specifically, the domain of the terms of type SymbolTable is partitioned into its) equivalence classes by the extractors defined on the type -- and this is precisely what an implementation is attempting to capture. The defining equations of each function indicate how it “contributes” towards this partitioning, and therefore how this “semantic structure” imposed upon the terms of the SymbolTable is related to the syntactic structure of the underlying terms, Due to lack of space, we omit the details of how this is done. One of the implemetations generated (automatically) is shown in figure 1, wherein 0 denotes the “implementation map. We note that an auxiliary data type which is almost isomorphic to a Stack (of integers) was (automaticlly) defined in course of the implementation; this Stack can, in turn, be synthesized in terms of an indexed Array by a recursive invocation of the synthesis procedures. Other Implementations that are generated for the Symbol Table include an implementation using a “Block Mark” to identify the application Of the function ENTERBLOCK, and an implementation similar to a “hash table” implementation is suggested upon examining the semantics of the functions defined on the Symbol Table. --l----------cII--------------------~~~==~=-------~~~~~~~”~~-- We list below one of the implementations generated for a SymbolTable, using an Array as the initially specified target type. The final representation consists of the triple <Array,lnteger,AOT1*1 the integer represents the current index into the array, whereas AOTl (for &diary Eata lype-1) is introduced in course of the synthesis process, and is isomorphic to a Stack that records the index-values corresponding to each ENTERBLOCK performed on a particular instance of a SymbolTable.) We denote this by-writing t#s) - <a,i,edtl>. informally, ENTERBLOCK.ADTl serves to “push” the current index value onto the stack adtl, LEAVEBLOCK.ADTl serves to “POP” the stack, and D.ADTl returns the -topmost element in the Stack, returning a zero if the stack is empty. WCC and PREO are the successor and menus functions on Integers. B(NEWST) E <NEWARRAY,ZERO,NEWAOTl’ B(AOOIO(s,id,al)) - <ASSIGN(a,SUCC(i),<id,al>),SUCC(i),adt 1’ g(ENTERBLOCK(s)) - <a,& ENTERBLOCK.AOTl(adtl,i)> aLEAVEBLOCK(s)) - <a,D.AOTl(adt 1), LEAVEBLOCK.ADTl (adtl)> BIISINBLOCK(s,id 1)) = ISINBLOCKTT(<a,i,adt l>,id 1) bKRETRIEVE(s,idl)) - RETRIEVETT(<a,i,adt l>,ldl) lSlNE3LOCKTT and RETRlEVETT are defined as follows: ISINBLOCKTT(~a,i,adt 1~’ id 1) - if i - ZERO then FALSE else if D.ADTl(adt 1) < i then if proj(l, OATA = idl, then TRUE else ISINBLOCKTT(<a,PRED(i),adt l*,id 1) else FALSE RETRIEVETT(<a,i,adt l>,id 1) = if i - ZERO then UNDEFINED else if proj(l,OATA)a,i)) - id1 then proj(2, OATA(a,i)) else RETRIEVETT(<a, PRED(i),adt I>, idl) Here, proj(i,exl..xn”) - Xi. Figure 1. A Symbol Table lmplemen!a!iOn Othw Examples of Applications of the Synthesis Paradigm. Several Programs have been synthesized by direct applications of the synthesis algorithms developed so far. These include implementations for a Stsck, a Queue, a Oeque, a Block Strut tured SymbolTable, an interactive line-orlented Text-Editor, o text formatter, a hidden surface elimination algorithm for graphical dispteys, and an execution engine for a data driven machine, eferencso [l] J.Goguen, J.Thatcher, E.Wagner, J.Wright. Initial Algebra Semantics and Continuous Algebras. JACM 24:68-95, 1977. (23 J.Goguen, J.Thatcher, E.Wagner. An Initial Algebra Approach to the Specification, Correctness, and implementation of Abstract Data Types, in Current Trends in Programming Methodology, Vol IV, Ed. R.Yeh, Prentice-Hall, NJ, 1979, pages 80-149. [33 J.Guttag, E.Horowitz, O.Musser. The Design of Data Type Specifications, in Current Trends in Programming Methodology, Vol IV, Ed. R.Yeh, Prentice-Hall, N.J., 1979. [43 J.Guttag, E.Horowitz, O.Musser. Abstract Data Types and Software Validation. CACM 21:1048-64, 1978. [53 P.A.Subrahmanyam. Towards a Theory of Program Synthesis: Automating Implementations of Abstract Data Types. PhO thesis, Dept. of Comp. SC., State University of New York a! Stony Brook, August, 1979. 16 J P.A.Subrahmanyam. A Basis for a Theory of Program Synthesis. Technical Report, Dept. of Computer Science, University of Utah, February, 1980. [7 3 OBarstow. Knowledge-Based Program Construction, Elsevier North-Holland Inc., NY., 1979. 76

1980

5

45

ABSTRACT Rule-Based Inference In Large Knowledge Bases * William Mark USC/Information Sciences Institute Having galned some experience with knowledge-based systems (e.g., [a), [9], [ 1 l]), our aspirations are growing. Future systems (for VLSI design, office automation, etc.) will have to model more of the knowledge of their domains and do more Interesting things with It. This means larger, more structured knowledge bases and inference mechanisms capabl? of manipulating the structures these knowledge bases contain. The necessarily large Investment In building these systems, and the very nature of some of the applications (e.g., data base query, cooperative interactive systems), also require these systems to be more adaptable than before to new domains within their purview (e.g., a new data base, a new Interactive tool). II RULE-BASED INFERENCE The need for adaptability argues strongly for perspicuity and modularity in the inference engine: the adapter must be able to see what must be changed (or added) and be able to make the alterations quickly. Inference mechanisms based on rules have these characterlstlcs. Unfortunately, most rule-based * Thus research was supported in part by the Defense Advanced Research Projects Agency under Contract No. DAHClS 72 C 0308, ARPA Older NO. 2223, and tn part by General Motors Research Laboratories. Views and conclustons contained in this paper are the author’s and should not be interpreted as representing the official opinion or policy of DARPA, the U.S. Qovsrnment or any person or agency connected with them. approaches rely on small, simply structured system knowledge bases (e.g., the rule-based formalism used In [lo] and [la] is dependent on representation In terms of triples). As rule-based systems grow to encompass a large number of rules, and as they are forced to work on complex knowledge structures to keep pace with modern knowledge base organizations, two major problems arise: o The inference mechanism becomes inefficient: It is hard to find the right rule to apply If there are many possibilities. o Rules begin to lose their properties of modularity and perspicuity: the “meaning” of a rule, especially In the sense of how It affects overall system behavior, becomes lost if the rule can interact with many other rules In unstructured ways. The remainder of this paper describes an approach to solving these problems based on a philosophy of doing inference that Is closely coupled with a principle of rule base organization. This approach is discussed in the context of two Implementation technologies. Ill PHILOSOPHY The inference methodology described here is a reaction to the use of rules as pattern/action pairs in a rule base that is not intimately related to the program’s knowledge baso. The basic philosophy [6] is to treat expert system inference as a process of redescription: the system starts with some kind of problem specification (usually a singlo user request) and redescribes It until It fits a known solution pattern (a data base access command or an implemented program function in the two examples described below). The control structure for this redescription depends on a knowledge representation that explicitly includes the Inference rules in the knowledge base. The redoscriptlon effort proceeds In two modes: a narrowing-down process that simply uses any applicable rules to redescribe the input as something more amenable to the program’s expertise; and a m-b process that 190 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. takes the description remaining from the first process (I.e., v&en no more applicable rules can be found), finds the most closely related solutlon pattern, and then uses that solution as a goal. In thls homlng-in mode, the inference procedure uses consequent rules to resolve the differences between the current description and the desked end. The narrowing-down phase focuses solely on the Input and transforms It lnto something closer to the system’s solution model. Depending on the input, the result of narrowlng-down might be an actual solution (meaning the Input request was quite close to the system expectations), or something that the system has no idea of how to deal with (meaning that it was very far from what was expected). Given that the user has some Idea of what the system can do, and that he wants to make himself understood, we can assume that the result of narrowing-down will often be only a minor perturbatlon of a system solution. This enables the homing-in process to find a closely related solution to use as a goal. Very speclflc consequent rules can then be used to resolve the few remalnlng differences. Assuming rule-based inference and a structured system knowledge base, the above phllosophy can be restated as follows: o Inference Is the transformation of one significant knowledge base structure into another. o The organization of the knowledge base organ&es the rule base. o The control structure supports both straight rule application anywhere in the knowledge base and consequent rule application In the context of a well defined goal. This approach makes rule application more efficient, even in a large knowledge base, because highly specific rules are not used--are not even looked for--until they have a good chance of success. That Is, the system does not have to look through everything it knows every time it must apply a rule. This efficiency is further enhanced by the fact that inference is modeled as redescriptlon and rules are tied closely to the system knowledge base. The inference mechanism will not even look at rules that do not relate directly to the kind of knowledge structures found in the description to be transformed. This makes the problem of finding the right rule to apply far less dependent on the number of rules In the system. The highly structured and detailed nature of the knowledge base now works for efficiency of rule application rather than against it. Separating the Inference process into modes and tying rules directly to the structures of the knowledge base also enhances the modularity and perspicuity of the rule Thls methodology has been used to Implement the inference components of two knowledge-based systems using quite different technologies. The first, a pat tern-match design, Is slmply sketched, while the second, a network based scheme, Is presented in more detail. . Inference In Asklt -- The Asklt system currently being developed at General Motors Research Laboratories (173, [S]) is a natural language data base query facility designed to be adaptable to new domains (e.g., a new project scheduling data base or a new inventory management data base). The inference task is to translate the user’s query Into the appropriate set of data base commands to provide the neodcd data. Askit’s knowledge base consists of case structures representing user request types, data base commands, individual English words, etc. Rules are expressed as transformations between system case frames. The condition part of the rule is a partially instantiated case frame that is treated as a pattern to be further Instantiated (Le., “matched”) by a description (a fully Instantiated case frame) from the program’s current state. The conclusion is a case frame to be filled on the basis of the instantiation of the condition part. When the rwlo is applied, the instantiating description is replaced In tho current state by the new structure generated from the conclusion part of the rule. Rules therefore represent allowable redescriptions of case frames for certain choices of case fillers. For example, the following is a rule for redescribing restrictions in certain user requests as SUBSET commands to the database system: 191 (REOUEST (OBJECT <table-data>:objI (RESTRICTION oobj (ccolumn-data>:property :relation :vaiue)) -> tSUBSt?T =obj WHERE -property w-elation =value) That is, the condition of the rule applies to any request which deals with something that is classified as “table data” in the system knowledge base, and which expresses a restriction of that table data in terms of “column data”. The conclusion states that such a request can be redescribed as a SUBSET command. When the rule is applied, the SUBSET command replaces the request In Askit’s current state. Rules are organized into “packets” based on the case frames In the knowledge base: all rules whose conditions aro partial instantiations of the same case frame are grouped in the same packet. For example, all rules that deal with restrictions In user requests would form a packet. The packet is represented by a packet pattern which states the common case structure dealt with by the rules of the packet, I.e., a generalization of their condition parts. The packet pattern for the ‘@restriction in rwquests” packet would be: (REClUEST (OBJECT :obj) (RESTRICTION nobj :rstr)) Packets play a key role in Askit’s rule application process. The process begins by matching a description in Asklt’s current state agains’t the packet patterns known to the system. If a match is found, the individual rules of the packet are tried. If one of the rules is successfully matched, the input structure is redescribed and placed back in the current state. If no rule matches, Askit goes into homing-in mode. It posits the discrepant part of a not-fully-matched rule from the packet as a new description in the current state, and looks for consequent rules to match this description. Consequent rules are also organized into packets (but based on their conclusions rather than their conditions). If Askit finds a matching rule in the consequent packet, and if the condition part of the consequent rule matches other descriptions in the current state, the discrepancy between the original description and the partially matched rule is considered to be resolved, and processing continues. Otherwise, other possible discrepancies or other partially matched rules are posited, and homing-in Is tried agaln. Thus, the narrowing-down process in this system Is represented by normal packet application, in which an lnltial description is matched against packets and successively redescribed until it can be seen as an instantiation of one or more system case structures representing data base commands. If this process breaks down, i.e., if a packet pattern is matched but no rule In tho packet can be matched, the input description is troated as a perturbation of one of the case structures rapresented by rule conditions in the packet. Consequent rules are then used to home-in the discrepant description on the deslred case structure. Successful resolution of the discrepancies allows normal processing (i.e., nerrowlng-down) to continue. t3, Inference in Consul -- The Consul system is being designed to support cooperative interaction between users and a set of online tools (for text manipulation, message handling, network file transmission, etc.). This “cooperative” interaction Includes natural language requests for system action and explanation of system activities. Since user requests may differ radically from the Input form expected by the tool designer, Consul’s inference task is to translate the user’s natural form of requesting system action into the system’s required input for actually performing that action. In the Consul system, the dependence of rule ropresentation and organization on the system knowledge base is carried much further than in Askit. On the other hand, the control structure does not have to be as complex. Consul’s knowledge base is a KL-ONE [2] topresentation of both tool-independent and tool-dependent knowledge. The major organizational framework of the knowledge base is set by the tool-independent knowledge, with the tool-dependent elements instantrating it. inference in Consul is a process of taking an exlstlng description in the knowledge base, redescribing it In accordance with an inference rule, and then reclassifying * It In the knowledge base. If the description can be classified as the invocation of an actual tool function In the system, then the description is “executable”, and inference is complete. A key aspect of Consul inference is that inference rules are also rspresented in KL-ONE in the knowledge base; the same process of classification that determines the status of a description also determines the applicability of a rule (compare 153). For example, let us examine Consul’s treatment of the user request “Show me a list of messages.” Parsing (using the PSI-KLONE system [l J) results In the structure headed by ShowAct. in figure 1. Consul classifies thls structure in its knowledge base, finding, as shown in figure 1, that it is a subconcept of the condition of Rulcl. This means that Consul can redescribe the request ShowAct. as a call to a “display operation” according to the conclusion of Rulel. The result is a new description, DisplayOperationlnvocationl .l, which Consul then classifies in Its knowledge base. If, via this classification process, the new descrlptlon Is found to be B subconcept of an executable function, inference Is “c The classification algorithm was written by Tom Lipkis. 192 complete--DlsplayOperationlnvocationl .I can simply be invoked by the Consul interpreter. Otherwise Consul will have to use additional rules to further refine the description until it can be seen as an actual call on 8ome other function or functions. (ab]ect) (mclpiant) Figure 1: Rule Application in Consul Figure 2 shows the classlflcatlon of DlsplayOperatlonlnvocatlonl .I in Consul’s knowledge base. Unfortunately, the system does not know very much about it In this posltion: it is not a subconcept of an executable function (shown shaded), nor Is It subconcept of any rule condition. Since no applicable function or rule can handle the current description, the system seeks a “closely related” function description to use as a mapping target. The deflnltlon of Wosely related” Is that the two descriptions must share a common ancestor that is a @*basic” concept On Consul% actual ., the ancestor cannot be rule Itself, nor newly generated description. the function d P)lsplayMessageSummeryl ocation is closely related to the current description because they share an appropriate lsplayOperationlnvocation. Once a related tool function description Is found, It Is used as a goal for consequent reasoning. First, Consul must find the discrepancies between the current description and the desired result. These discrepancies re simply the differences that escrlption from being classified as desired description: In this example, DisplayOperationlnwocationI .I that prevent It from instantiating DlsplayMcssageSummerylwwoc ThW discrepancy is that the “input” role of the current ption is filled wlth list of messages ageLIst. f ), while the executable function yMessageSummerylnwocation requires list of mmarles (Summaryhist). Figure 2: Flndlng B Closely Related Function Description y to redes If It hope iatlon of P executable function. ere are two ways the current description: rules and executable functions (since functions produce new descriptions via output and side-effects). Rules are preferable because they save tool execution time; Consul therefore looks for rules first. In this case, there is a quite general rule that produces the desired effect. Users frequently ask to see a list of things (messages, files, etc.) when they really want to see a list of summaries of these things (surveys, directories, etc.). Consul therefore contains a rule to make this transformation, if necessary. The rule, shown as Rule2 in figure 3, says that if the current description is a display operation to be invoked on list of “summarizable objects”, then it can be redescribed as a display operation on a list of summaries of those objects. Consul finds this rule by looking In the knowledge base for rules (or executable functions) that produce the “target” part of the discrepancies found earlier. As shown in figure 3, Rule2 can be found by looking “up” the generalization hierarchy (not all of which is shown) from SummaryLIst. Consul must next be sure that the condition part of the rule is met in the current state of the knowledge base (this Includes the current description, descriptions left by previous rule applications and tool function executions, and the original Information in the knowledge base). if the condition is not satisfied, Consul will try to produce the needed state through further rule application and function 193 application of the execution--i.e., through recursive consequent reasoning process. In this example, the condition of Rule2 is entirely met in the current description (see figure 3). Furthermore, the conclusion of Rule2 resolves the entire discrepancy at hand. In general, a combination of rules and functions is needed to resolve discrepancies. Therefore, with the application of Rule2, Consul has successfully redescribed the initial user request as an executable function. The Inference process passes the resulting description on to the interpreter for execution, and the user’s request is fulfilled. Narrowing-down and homing-in are much the same In Consul as they are in Askit. In Consul, however, relationships between rules and notions such as “perturbation” and Iccfosely related structure” come directly from the existing knowledge base representation; a superimposed packet structure is not necessary. The control structure of rule application therefore need not consider the separate issues of matching packets, setting up packet environments, matching rules, etc. Instead, classification includes matching, as applicable rules are found “automaticallyl* when a newly generated description Is put in its proper place in the knowledge base. Finding related knowledge structures and consequent rules are aI90 classification problems, representing only different use of the classification algorithms. a slightly b DisplryOporatlon Invocatlonl.1 summarized object) Rusty Bobrow and Bonnie Webber, “F’SI-KLONE: Parsing and Semantic Interpretation in the BBN Natural Language Understanding System,” in Proceedings of the 1980 Conference of the Canndlan Sociely for Computational Studies of /nte//igonco, CSCSI/SCEiO, 1980. Ronald Brachmen, A Structural Paradigm for Reprsscntlng Knowledge, Bolt, Beranek, and Newman, Inc., Technical Report, 1978. Bruca Buchanan, of al., Heuristic DENDRAL: A Program For Generating Explanatory Hypotheses In Organ/c Chemistry, Edinburgh University Press, 1960. Ranclall Davis, Applications of Meta Level Knowledge to the Construction, Maintenance and Use of Large Know/edge Bases, Stanford Artificial Intelligence Laboratory, Technical Report, 1976. Richard Fikes and Gary Hen&ix, “A Network-Based Knowledge Representation and it:; Natural Deduction System,” in Proceedings of the Fifth lnternatlonal Joint Conference on Artificial Intelligence, IJCAI, 1977. William Mark, The Reformulation Model of Expertise, MIT Laboratory for Computer Science, Technical Report, 1976. William Mark, The “Asklt” English Database Query Facility, General Motors Research Laboratories, Technical Report GMR-2977, Jim3 1!379. William Mark, A Rule-Based Inference System for Natural Language Database Query, General Motors Research Laboratories, Technical Report GMR-3290, May 1980. William Martin and Richard Fateman, “The MACSYMA System,” Proceedings of the Second Symposium on Symbolic and Algebraic Manipulation, 197 1. Edward Shortliffe, MYCIN: Computer-Based Medical Consultations, American Elsevier, 1976. William Swartout, A Digitalis Therapy Advisor with Explanations, MIT Laboratory for Computer Science, Technical Report, February 1977. Figure 3: Using Consequent Reasoning 194

1980

50

46

N PROCESS FOR EVALUATING TREE-CONSISTENCY John L. Goodson Departments of Psychology and Computer Science Rutgers University, New Brunswick, N.J. 08903 ABSTRACT General knowledge about conceptual classes represented in a concept hierarchy can provide a basis for various types of inferences about an individual. However the various sources of inference may not lead to a consistent set of conclusions about the individual. This paper provides a brief glimpse at how we represent beliefs about specific individuals and conceptual knowledge, discusses some of the sources of inference we have defined, and describes procedures and structures that ' can be used to evaluate agreement among sources whose conclusions can be viewed as advocating various values in a tree partition of alternate values. * I. INTRODUCTION Recent work by several investigators; [II, C31, C4l and C7l; has focused on the importance of augmenting deductive problem solvers with default knowledge. Their work provides some of the logical foundations for using such knowledge to make non- deductive inferences and for dealing with the side effects of such inferences. Currently, we are pursuing how conceptual knowledge about general classes of persons, locations, objects, and their corresponding properties can be represented and used by a planning process and a plan recognition process to make deductive and non-deductive inferences about particular persons, objects and locations in an incompletely specified situation (see C51 and C61). General knowledge about conceptual classes represented as a concept hierarchy provides a basis for various types of inferences about individuals. The definition of a conceptual class might be believed to hold for an individual, xl, that is believed to be a member of that class (Definitional Inference). The definitions of concept classes which include the class of which xl is a member might be believed to hold for x 1 (Inheritance Inference). The definition of some class that is a subset of the class to which xl belongs might be -------- * This research is supported by Grant RROO643-08 from the Division of Research Resources, BRP, NIH to the Laboratory for Computer Science Research, Rutgers University. used as a source of potential inferences (a kind of Plausible Inference). Additionally, information might be stored directly about xl (Memory Inference) and there may be other inference types based on different strategies of deductive or plausible inference (for a more detailed discussion see [61). However, these sources of inference may not lead to a consistent set of conclusions about the individual. For default theories in general, Reiter [4] has shown that when a default theory is used to extend a set of beliefs, determining the consistency of the extensions is an intractable problem. In this paper we are concerned with a very local and focused subproblem involved in evaluating the agreement or consistency of a set of conclusions and with a strategy for dealing with belief inconsistency. The focus arises from considering these issues in the context of a concept class hierarchy where the classes form a tree partition. Before we discuss this restricted case of belief consistency, we provide a brief glimpse at how we represent beliefs about specific individuals and at several types of class hierarchies used to represent concepts. II. REPRESENTATION a BELIEFS AND CONCEPTS Beliefs about specific objects, persons, etc. are represented as binary relations of the form ((x r y).CT F or Ql) where : r is a relation defined between two basic classes of entities X and Y; x is an instance of X; and y is either an instance of Y or is a concept that is part of the concept hierarchy with Y as its root. An example of the former relation is ((DON LOC NYC).F) where DON is an instance of PERSON and NYC is an instance of LOCATION. The latter form is exemplified by ((DON AGEIS YOUNG1.T) where YOUNG is a concept that is part of an AGE hierarchy. T, F or Q represents the truth value in the current situation. Concepts are organized into inclusion hierarchies which have as their root one of several basic classes, such as PERSON, AGE, LOCATION, OBJECT, which may have instances in a specific situation. A particular individual may be an instance of several basic classes, e.g. a person DON may be viewed as an OBJECT or a PERSON. Two simplified hierarchies are given below in graph form. Note that no claim is being made here about the adequacy or naturalness of the knowledge represented in these examples. 195 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. PERSON III. CONSISTENCY m INFERENCE SOURCES AGE I I ------------------- ------- We now consider structures that can be used t.o I I f t determine the consistency of a set of sources STUDENT ATHLETE YOUNG OLD contributing beliefs relevant to a proposition I I I I about an individual. The following paradigm -----mm---- ---------- m-w- ---- provides a more specific context in which to I I I I I I I I I I discuss the problems and mechanisms we have HIGH SCHOOL COLLEGE BASEBALL FOOTBALL I5 25 40 80 considered. Assume that the task specification is: STUDENT STUDENT PLAYER PLAYER YR YR YR YR 1) A goal proposition is given whose truth value is Hiearchies may contain subtree partitions as might desired by a higher level process. The goal is be the case for the PERSON hierarchy shown above. a statement about a particular individual A particular person may be a high school student represented as a binary relation between the &a football player. Some hierarchies, individual and a concept, called the goal- especially those representing properties, may form target. The goal-target is a member of a tree tree partitions, as is the case for the AGE partition of concepts called the target-tree. hierarchy shown above. This '1family11 of concepts is the set of potential targets defined for the relation There are several types of information that occurring in the goal. It should be noted that can be associated with a concept. One type, the there may be several trees rooted in the same concept definition, provides an intensional basic class. For now we limit our consideration characterization of the elements of the conceptual to the case where there is only one tree. class. A definition is a conjunctive set of descriptions of the form (<basic 2) Several sources of inference are consulted for class> (rl yl)(r2 zl)...) that are the necessary beliefs relevant to the goal, that is, beliefs and sufficient conditions for an instance of the relating the individual to concepts in the basic class to be considered an element of the target-tree. conceptual class. The relation that represents the hierarchical structure among concepts is termed 3) Since the target-tree is a partition, the "tree- COVERS and appears in the above example as a line consistency" of the beliefs can be evaluated and joining pairs of concepts. This relation implies used to determine a truth value for the goal that for each description in the higher level (one of T for true, F for false or Q for concept definition (e.g. (rl yl)), there is a question, determinate T or F not assignable). corresponding description -(e.g. (1-1 ~2)) in the This evaluation amounts to assessing the lower level concept definition and either yl=y2 or agreement among the sources on a truth value for (~1 COVERS ~2) in the Y (basic class) hierarchy and the goal. the lower or more specific concept definition must contain at least one description that is more 4) The structures created in 3 can be used to specific (where (yl COVERS ~2)) than the record the sources of inference drawn upon in corresponding description in the more general the attempt to achieve a conclusion about the concept definition. As an example, consider these goal. possible definitions for the concept ATHLETE and Many of the structures and procedures discussed in and a more specific concept, FOOTBALL-PLAYER. relation to this paradigm are implemented in the knowledge representation system AIMDS (see [ 83). (ATHLETE DEF I (PERSON (PHYSICALSTATE SOUND) The target-tree can be represented as a Truth I (PLAYSON TEAM)) Value Tree (TVT) where each node represents a COVERS1 concept in the target-tree. Each node has a slot, (FOOkBALL-PLAYER DEF TV, for one of the determinate truth values, T or F, and slots for two lists, a true-list and a (PERSON (PHYSICALSTATE SOUND> false-list. These lists consist of two inner (PLAYSON FOOTBALL-TEAM)) lists. The true-list contains one list for recording the sources that support the truth value T and one for recording nodes that require the node An example of a type of plausible inference to have the truth value T. The false-list has the can be given using these definitions. If the same structure and records the information relevant beliefs in memory about DON satisfy the to the truth value F. An example will serve to descriptions in the definition of ATHLETE, then clarify how this structure is used to evaluate there ' basis for believing (DON PLAYS;: FOOTBiLL-TEAM) with truth value T. tree-consistency. However, it is not the case that this inferred Assume that the goal is (DON AGEIS 40YR). The belief is consistent with those in memory. Thus TVT for the goal represents the AGE tree presented two sources of information about DON, memory and in a previous example. Initially each node in the plausible inference, may not agree. TVT has three empty slots shown below. 196 slots (TV NIL) Truth Value (T 1 True-list (F Ii 1 False-list supported required by by If ( (DON AGEIS YOUNG) .T) is contributed by memory, the following cycle of actions takes place. 1) The truth value slot of each node is made NIL. 2) T is entered as the TV of the node representing YOUNG. 3) Tree-consistency rules propagate truth value requirements to other nodes. Note that. these rules depend on interpreting the AGE hierarchy as a tree partition of a finite and closed set of values. One rule propagates T up the tree ( to all ancestors), thus the AGE node receives the truth value T. Another rule gives the node OLD the truth value F since the relation AGEIS has been defined such that only one path (leaf to root) of targets in the tree can be true for an individual at any given time, and all others must be false. Finally , a third rule propagates F down the tree (to all descendents) , thus the nodes 40YR and 80YR receive the truth value F. In the general case, these rules are looped through until no additional nodes can be given a truth value. Two other rules not applicable here are: propogate T from a parent to a daughter if all other siblings are F; and if all daughters are F, propogate F to the parent. 4) The source of the truth value T for the node YOUNG, i .e. Memory Inference (MI), is registered as support on its true-list. For each of the remaining nodes with a non-null TV, YOUNG is registered as a requirement on the true-list or false-list depending on the truth value required by YOUNG being T. At the end of this cycle, the nodes in the AGE TVT have the following form. YOUNG OLD TV T TV F (T (MI)()) (T 00) (F 00) (F O(YOUNG)) 15YR 25YR 4OYR 80~~ TV NIL TV NIL TV F TV F (T (10) (T 00) (T 00) (T 00) (F 00) (F 00) (F O(YOUNG)) (F O(YOUNG)) This cycle can be repeated for each of the relevant beliefs that have been contributed. Note that a cycle generates the deductive consequences of a truth value assignment, T or F, to a single node. We are not concerned with the propagating changes to previously assigned truth values based on a new assignment from a second relevant belief. Such a mechanism would be required for updating a model and might utilize antecedent and consequent propagation proposed by London [ 21. The TVT created by this cycle is inspected for a truth value for the goal. In our example, the node 40YR represents the goal-target and it has an entry for a single truth value, F. Thus F is returned as t,he truth value of the goal. However, the assignment of a truth value to the goal can be made contingent on the target-tree. Tree- consistency and thus agreement among the sources is easily determined. If any node has entries in both the true and false lists, then an inconsistency exists and the value Q (indeterminate truth value) should be returned even if a determinate truth value is indicated for the goal -target. Conversely, if no such node can be found, then the tree, and thus the set of beliefs contributed, are consistent and the truth value indicated for the goal-target may be returned. One way to find a det,erminate truth value for the goal when the tree is inconsistent is to seek a subtree of TVT such that : I) the subtree is rooted in the top node; 2) the subtree is tree-consistent; 3) the subtree contains the node representing the goal-target; and 4) this node has a true-list or false-list entry, then a determinate value can be assigned to the goal. An intuitive example of this case is where you can be fairly sure that Don is young even though you have conflicting information about whether he is ten or fifteen years old. If the consistent subtree does not indicate a determinate truth value for the node representing the goal-target, the tree-consistency requirement can be relaxed in order to find a determinate truth value. The consistent subtree can be extended sue h that the resulting subtree is tree-consistent, each node has a determinate trut,h value and the set of truth values is maximally supported by the sources contributing information. One extension procedure involves the following steps : 1) For each node in the consistent subtree assign the truth value indicated by its non-empty list. 2) From the set of nodes without a truth value (TV NIL), select the node with the maximum support for a truth value that is tree-consistent with the current subtree. 3) Assign this node the truth value indicated and apply the tree-consistency rules to further extend the consistent subtree. 4) Continue at 2 until all nodes have a truth value or the basis for deciding 2 does not exist (some 197 nodes may have empty support and cannot be assigned a determinate truth value by the rules). Several issues must be addressed in carrying out step 2. First, since ties are possible, a decision procedure must be provided. Second, the degree to which a node is required to have a particular truth value might be taken as a measure of indirect support for that truth value. Since the set of nodes that could possibly require a node to have a truth value is partially dependent on its position in the tree, positional bias must be taken into account in deciding degree of indirect support. This procedure can be applied even when a consistent subtree cannot be found. In this case the top node is given the truth value T to provide a trivial consistent subtree from which to extend. The staged relaxation of the tree-consistency constraint is particularly important when all of the beliefs relevant to a goal are drawn from default knowledge (arrived at through non-deductive inferences). There may be little basis for expecting this knowledge to be consistent yet it may be rich enough to suggest that one truth value is more plausible than the other. ACKNOWLEDGEMENTS. My collaborators, C.F. Schmidt and N.S. Sridharan, have contributed to the evolution and implementation of the ideas expressed here. REFERENCES. Cl] Doyle, J. A Truth Maintenance System. Artificial Intelligence 12 (1979) 90-96. [2] London, P. A Dependency-based Modelling Mechanism for Problem Solving. AFIPS Proc. Vol. X NCC-78, Anaheim, Ca., (June, 1978) 263-274. [3] McDermott, D. & Doyle, J. Non-Monotonic Logic I. Proc of the Fourth Workshoo on Automated Deduction. Austin, Texas, (Feb., 1979) 26-35. [4] Reiter, R. A Logic for Default Reasoning. Technical Report 79-8, Dept. of Computer Science, University of British Columbia, (July, 1979). 151 Schmidt, C.F. The Role of Object Knowledge in Human Planning. Report CBM-TM-87, Dept. of Computer Science, Rutgers University, (June, 1980). t.71 [81 c91 Shrobe, H.E. Explicit Control of Reasoning in the Programmer's Apprentice. Proc - oft&& Fourth WorkshoD on Automated Deduction. Austin, Texas, (Feb., 1979) 97-102. Sridharan, N.S. Representational Facilities of AIMDS: A Sampling. Report CBM-TM-86, Dept. of Computer Science, Rutgers University, (May, 1980). Sridharan, N-S., Schmidt, C.F., & Goodson, J.L. The Role of World Knowledge in Planning. Proc 1980). AISB-80. Amsterdam, Holland, (July, C63 Schmidt, C-F., Sridharan, N.S. & Goodson, J.L. Plausible Inference in a Structured Concept Space. Report CBM-TR-108, Dept. of Computer Science, Rutgers University, (May, lg8o). 198

1980

51

47

Reasoning About Change in Knowledgeable Office Systems Gerald R. Barber Room 800b Massachusem Ittslitute of Techtlologv 54.5 Techttologv Square Cambridge, Mass. 02139 (617) 253-5857 ABSTRACT Managing of and reasoning about dynamic processes is a central aspect of much activity in the office. WC present a brief dcscrintion of our view of office systems and why change is of central importance in the office. A description system used to describe the structure of the office and office activity is discussed. A viewpoint mechanism within the description system is prcscntcd and we discuss how this mechanism is used to dcscribc and reason about change in the offke. A gcncral scenario .is dcscribcd in which viewpoints arc illuslratcd as a meiins of describing change. Previous technologies for accommadating change in knowlcclgc embedding languages are charnctcrizcd. WC contrast the approach using viewpoints with previous technologies whcrc change is propagated by pushing and pulling information bctwccn slots of data structures. I. Int reduction The computer has been used in the office cnvironmcnt for many years with its application mainly limited to highly structured and repetitive tasks in a non-interactive mode. With the cvcr-dccrcasing cost of hardware computers can bc potentially used in the future to aid office workers in a wider variety of tasks. Indeed, the computer based office system is today seen as both a motivation for achieving a new understanding of office work and as a medium within which to intcgratc new tools and knowledge into a cohcrcnt system. This has led to the realisntidn that there is enormous potential in the use of the computer in the office in novel anti as yet unforcscen ways. These new uses will impact the way office work is done in fundamental ways demanding new ideas about how to manage information in an office and a new conceptualization of what off& work is in the presence of powerful computational capabilities. As a step toward developing computer systems that may effcctivcly support office workers in their tasks we employ two paradigms from Artificial Intclligencc, those of knowledge embedding and problem solving. We arc dcvcloping ;I descriptions system called OMEGA [IIcwitt SO] to bc used to cmbcd knowlcdgc about the structure of the office ilnd office work in an oflicc system. One of our objcctivcs is to support the problem solving activities of individuals in an office. Much of the problem solving activity within an office concerns reasoning about change. WC hnvc dcvclopcd mechanisms in OMEGA to dcscribc changing situations. In the following section we present a short description of our model of the office. Following this WC discuss the importance of ch,mgc in the office and the mechanism within OMEGA to deal with change. ‘I’hc approach WC take is compared with other approaches to the problem of accommodating and managing change. II. The Knowledgeable Office System We view an office system in terms of the two dominant structures in the office, the applicalion s(nlcfure and the organiza~iunal struclure. l’hc application structure of an office system concerns the subject domain of the office. It includes the rules and objects that compose the intrinsic functions of a particular office system. As an example, in an office concerned with loans the application structure includes such cntitics as loan applications. credit ratings and such rules as criteria for accepting or rejecting loans. In an insurance company the application structure is concerned with insurance politics, claims and actuarial tables. ‘Ihc application structure explains the scope of the functionality an offrcc system has on a subject domain as well as providing a model by which those functions arc charncteriir.cd. OVCI tly, the application structure is the primary reason for the existence of the office system. In contrast to the application structure is the social structure of the office system as an organi/.ation [Katz 781. Our concern with this aspect of an office system stems from the fact that the activity in the application domain of an ofl?cc system is IcalLed by people coopcrating in a social system. ‘I’hc structure of this social system involves such aspects of an organization as the roles of the individual participants. the interaction of roles, the social norms of the office and the various subsystems that make up the organization. WC bicw the office system as a functioning organism in an environment from which it extracts resources of various kinds and to which it dclivcrs the products of its mechanisms. OMEGA’s descriptions arc the hlndamental entities upon which the Knowlcdgcnblc Office System is based. ‘I’hc emphasis of our approach is on a description manipulation system for cmbcdding knowledge as opposed to a forms manipulation system. Descriptions are used to express the relationships bctwcen objects in the Knowledgcablc Office System. Descriptions are more fimdamental than electronic forms, in particular, electronic forms arc a way of viewing descriptions, a visual manifestations of descriptions. One of the goals of our work is to support office workers in their problem solving activity. Problem solving is a pervasive aspect of office work that has been ncglcctcd until recently [Wynn 79, Suchman 791. Office work is naturally characteri/.cd as goal oriented activity. ‘The office proccdurc is mcrcly a suggested way by which to accomplish a particular goal. WC believe that this is one reason why it has proved to bc difficult to dcscribc office work from a procedural point of view. ‘I‘hc formalism WC are developing allows us to describe and reason about the application and organizational structures of office systems as well as the inlcraction bctwccn these structures. ‘I’hc major bcncfits of OMEGA with relevance to our discussion here arc that a computational system can support problem solving in dynamic cnvironrncnts that are weakly structured, and knowledge rich. OMEGA also provides a prccisc Lmgu,igc within which to characterize the static and dynamic aspects of office systems. A central problem in an office system is reasoning about and managing change. This is a rccul rent thcmc at several lcvcls of 199 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. abstractions. Viewing the organization in relation to its environment, the organization must cvolvc in order to adapt to the changing cnvironmcnt. For example. an accounting off~cc must adapt to new tax laws or an office must adapt to new technology. Viewing the organization as producer of some product, the organization must adjust its production output to the demand for t.hc product which it produces in light of the rcsourccs available to the organitation and the constraints under which it must operate. The individuals that make up an organization arc faced with such tasks as reasoning about processes that have produced anomalous results, maintaining system constraints as the state of the constrained parts change and analyzing the implications of hypothesized processes. OMEGA has a viewpoint mechanism that is used to describe and reason about change. ‘I’hc viewpoint mechanism provides a means to present time varying proccsscs to office workers for analysis, bc these proccsscs historical, in progress or postulated. Cltanging cnvironmcntal dcpcndcncics and changing aspects of the organization can be captured in descriptions using the viewpoint mechanism. In the remainder of this pnpcr WC dcscribc OMEGA and its viewpoint mechanism. IV. The Viewpoint Mechanism OMEGA is a system with which a structure of descriptions is built. ‘1%~ system is dcsigncd to bc incremental; new knowlcdgc can be incorporated into the system as it is discovered or as the need for it arises. There is no minimal amount of information nccdcd bcforc the system is usable. The system is monotonic in the sense that nothing is lost when new information is added. As is cxplaincd in the following paragraphs, knowlcdgc is rclntivir.cd to viewpoints, information that is inconsistent with information in a particular viewpoint can be placed in a different viewpoint. ‘This ac&nmodntcs aspects of non-monotonic systems [McDermott 79]--where new information may invalidate previously held beliefs--without the need for a notion of global consistency. OMEGA’s fundamental rule of inference is mergitzg; new descriptions are merged with previous descriptions. Any new deductions as a result of the new information arc carried out during the merging operation. OMEGA is used to build, maintain and reason over a lattice of descriptions. Descriptions are related via an inhcritancc relation called the is relation. ‘I’he is relation is rclativized to a viewpoint that indicates the conditions under which the is rcl,ltion holds. Intuitively a viewpoint represents the conditions ulider which the inheritance relation holds. In this rcspcct it is rcminisccnt of McCarthy’s situational calculus [McCarthy 691 and the contexts of QA4 [Rulifson 721. A major differcncc between these approaches and viewpoints is that viewpoints arc descriptions and thus are subject to the full descriptive power of OMEGA. Viewpoints may be cmbcddcd in structures expressing complex inheritance relationships relating viewpoints to one another. Other aspects of OMEGA include higher order capabilities such as the ability to describe properties like transitivity for relations in the system and meta-description capabilities to talk about the parts of descriptions. Ill. Dealing With Change A key property-of viewpoints is that information is only added to them and is never changed. Consider, for example, a description which is the underlying representation of a form. I’hc description is rclativi7cd to a kicwpoint and information is added to this description increasing its specificity. JIcscriptions may contain constraints bctwecn attributes, as information is added further information may be deduced. Should the information in a field of a form be changed then the following scenario might occur: 1. A new viewpoint is created successor to the old viewpoint. and dcscribcd as being a 2. All information that was not derived from information is copied to the new viewpoint. the changed 3. The new information is added in the new viewpoint, deductions resulting from this information arc made. 4. ‘J’hc description in the new recent contents of the form. viewpoint rcprcscnt the most In this case the new viewpoint inherits all but the changed information and the information deduced from the changed information from the old viewpoint. What actions arc taken when information in a viewpoint is changed is specified via mcta-descriptions. Previous approaches to the problem of accommodating changing information have been to perform updntcs to the data structures in question. System based on property lists such as J,lSP have used pur and gel operations to update and read database information. ‘I’hcsc hate the disadvantage that deductions based on updated infonnation must bc handled explicitly leading to unacccptablc complexity and modularity problems. J.angungcs like J:RJ* [Goldstein 771 use triggers on data structure slots to propagate changes. ‘Ilrc disadvantage here is that thcrc is no support for keeping track of what was deduced and why. ‘This makes changes difficult because information dependencies arc not recorded. ‘lhc language K RI, has been used to implement a knowledge-based personal assistant called OlIYSSJ<Y [Fikes X0]. ODYSSlXY aids a user in the planning of trips. In this system pushers and pullers arc used to propagate deductions as a result of updates and to mnkc deductions on reads. A simple dependency mechanism is used to record information dcpcndcncics. In this cast it is necessary to be very careful ilbout the order in which triggers fire for as updates arc made there is both new and old information in the database making it difficult to prcvcnt anamolous results due to inconsistencies. OMEGA separates new and old information into diffcrcnt viewpoints. Information consistency is maintained within viewpoints. ‘l’hc propagation of information bctwccn viewpoints is controlled via mcta-description. An advantage of the approach using viewpoints is that the system has a historical character. This is an important step toward our goal of aiding office workers in problem solving about dynamic proccsscs. Viewpoints can bc used as historical records of past processes, as an aid in tracking ongoing proccsscs and as an aid to dctcrminc the implications of postulated actions. 200 VI. Conclusion We have presented the viewpoint mechanism of the descriptions system OMEGA along with some examples of its use to describe a changing form in an accounting office. The viewpoint mechanism has proved useful in describing objects whose properties vary with time as well as a means with which to interpret statements about the system’s description structure. The viewpoint mechanism presented here is related to that in IYHER [Kornfcld 791 and to the layers of the PIE system [Goldstein 801. Viewpoints arc a powerful unifying mechanism which combine aspects of McCarthy’s situational tags [McCarthy 691 and the contexts of QA4 [Rulifson 721. They serve as a replacement for update and pusher-puller mechanisms. Omega is monolottic using merging of descriptions as a fundamental rule of inference. It uses viewpoinls to keep track of different possibilities. This aspect causes it to differ substantially from systems based on property lists [IPL, Lisp, etc.] which arc based on operations to yul and get attributions in data structures. These differences carry over to more recent systems [SIR, SIMUI,A, FRL, KRL, etc.] based on record structures with attached procedures that execute when a plr( (rrytkt&) or ger (rettcl) operation is performed. References \Fikes 801 Richard Fikes. Odyssey: A knowledge-Based Assistant. To appear in Artificial Intelligence. [Goldstein 771 Goldstein, I. P. and Roberts, R.B. NUDGE, a Knowledge-Based Scheduling Program. Proceedings of the Fifth In~erna(ional Joint Cotlference on Artificial Intelligence. [Goldstein 801 Goldstein, Ira. PIE: A Network-Based Personal Information Environment. Presented at the Off~cc Semantics Workshop, Chatham, Mass. June 15-18 [Hewitt 801 Hewitt, C., Attardi, G., and Simi, M. Knowledge lkbedding wirh a Descriplion System. AI Memo, MIT, August, 1980. to appear [Katz 781 Katz, D. and Kahn, R. The Social Psychology of Organizations. John Wiley and Sons, 1978. [McCarthy 691 McCarthy, J. and Hayes, P. J. Some Philosophical Problems from the Standpoint of Artificial Intelligence. In Machine Inrelligence 4, pages 463-502. Edinburgh University Press, 1969. [McDermott 791 McDermott, D. and Doyle, J. Non-Alonoronic I,ogic I. AI Memo 486b, MIT, July, 1979. [Rulifson 721 Rulifson. J., Derksen, J. and Waldinger, R. QA4: A Procedural Calculus for lnluilive Reasoning. Artificial lntclligcncc Center Technical Note 73, Stanford Research Institute, November, 1972. [Suchman 791 Suchman, L. Office Procedures as Praclical .4ction: A Case Sludy. Technical Report, XEROX PARC, Scptcmber, 1979. [Wynn 791 Wynn, E. Office Conversaliott as an Infortnaiion Medium. Phi) thesis, Department of Anthropology, University of California, Berkeley, 1979. [Kornfeld 791 Kornfeld, W. Using Parallel Processing for Problem Solving. AI Memo 561, MIT, December, 1979.

1980

52

48

On Supporting the Use of Procedures in Office Work Richard E. Fikes and D. Austin Henderson, Jr. Systems Sciences Laboratory Xerox Palo Alto Research Center 3333 Coyote Hill Road Palo Alto, California 94302 Abstract In this paper, we discuss the utility of AI techniques in the construction of computer-based systems that support the specification and use of procedures in office work. We begin by arguing that the real work of carrying out office procedures is different in kind from the standard computer science notions of procedure “execution”. Specifically, office work often requires planning and problem solving in particular situations to determine what is to be done. This planning is based on the goals of the tasks with which the procedures are associated and takes place in the context of an inherently open-ended body of world knowledge. We explore some of the ways in which a system can provide support for such work and discuss the requirements that the nature of the work places on such support systems. We argue that the AI research fields of planning and knowledge representation provide useful paradigms and techniques for meeting those requirements, and that the requirements, in turn, present new research problems in those fields. Finally, we advocate an approach to designing such office systems that emphasizes a symbiotic relationship between system and office worker. Introduction We are interested in developing office systems that would make use of a knowledge base describing whal tasks are to be done, who is to do them, and how they are to be done. Such descriptions specify the functions of an office and how it is organized to achieve that functionality. We claim that such a knowledge base can form the basis for a broad range of system support in an office. In this paper, we discuss some of the ways in which AI paradigms and techniques are relevant to the support of office work by such computer-based systems. We begin by describing some of the support functions we have in mind, and then address what we consider to be the primary issue; namely: what is the nature and structure of the information in such a knowledge base? We are guided in addressing that issue by considering the nature of the work that occurs in an offrce and how such information is used in that work. We first argue that the work involved in carrying out office procedures is different in kind from the “execution” of a procedure that one might expect by drawing analogies with the behdvior of a computer executing a progmm. We illustrate and support this claim by presenting a typical case of office work and analyzing the actions that take place there. From this argument we derive a requirement for systems which support office work: namely, that they be flexible enough to support the variety of behavior occasioned by the unpredictable details of particular situations. We then turn to the relevance of AI for achieving this functionality. We develop the idea that the paradigms from the AI literature for automatic planning and execution monitoring of plans provide a more suitable alternative to the procedure execution model of office work: and furthermore that the demands of supporting office work require extensions to those paradigms. Second, we argue that the knowledge representation problems presented by the open-ended office domain are unsolved and challenging. We suggest that they can be attacked by the use of specialization-based representations and facilities for storing “semi-formal” structures in which uninterpreted text is intermixed with data whose semantics is understood by the system. Finally, we argue that the whole enterprise of supporting office work can only hope to succeed if we regard the office systems as functioning in a partnership with the office workers. Due to the open-endedness of the domain, the system cannot hope to “understand” the full import of the information which it is handling, and so must rely on human aid. Furthermore, to fully support the users, the system must be able to represent, although not necessarily understand, any of the infcrmation in the domain. We conclude by advocating an approach of “symbiotic processing” between system and office worker and the use of AI techniques in constructing systems to support office work. Supporting the Production and Use of Procedural Descriptions We begin by considering some of the whys in which computer-based office systems could facilitate the effective production and use of descriptions of what tasks are to be done, who is to do them, and how they are to bc done. There are two groups of people whom an office system dealing with such descriptions can support: the producers and the users. However, the production and LW phases are often tightly interwoven, with the same people often involved in both (despite what managers may choose to think). The producers of these “what, who, and how” specifications <typically managers and planners) are engaged in a process of organizing the work in the office so that the office’s goals and commitments will be met. That process involves defining the tasks to be done, designing procedures for doing those tasks, and assigning individuals to carry out those procedures. A system can support these specification processes by providing a descriptive namework in which to express the specifications and by helping to manage the complexity that arises from the interactions of the tasks. constraints, procedures, and policies being specified. The descriptive. framework would provide a guide as to what information needs to be specified (based on the intended purpose and uses of the specifications) and a terminology for expressing that information. For example, the system might provide a template for describing a task that would include fields for the task’s goals, inputs, outputs, responsible agent, activation conditions, etc., and a description language for filling those fields. The system could also indicate direct implications of a description, such as the subtasks implied by a task description of recognizing the task’s activation events, obtaining the task’s inputs, or communicating its outputs. The system could aid in managing the complexity of the specifications primarily by monitoring interface requirements atnong interacting components to help assure that those interfaces are well specified and the requirements are met. For example, if the description of a task included the source of each of the task’s inputs and the destination of each of its outputs, then the system could alert the specifier when those input-output connections between tasks are inconsistent (e.g., when some input is not an output of the 202 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. source task), and could input-outputs prompt for a specification are to be communicated. of how The other grotto that the system can support is the users of the “what, who. how” specifications. -. That support would include facilities for accessing the specifications in useful ways, for adding infarmal notes to the specifications, for monitoring the carrying out of tasks, and for doing some of the procedural steps. One useful way in which the system would act as an information source is in providing “how-to” information to a person who has a task to do and doesn’t know how to carry it out. For example, when a secretary is given a task that he is not familiar with (such as “order some business cards for me” or “obtain a consulting contract and arrange the travel plans for this person”), the system could provide him with a description of what needs to be done, how to do it, and who the people are who will play a role in getting it done. One could amplify the system’s usefulness in this role as a how-to resource by allowing its users to add informal notes to the descriptions. Then the system also becomes a repository for the accumulated societal wisdom concerning the realitics of actually carrying out the tasks. The functionality we have discussed thus far has only required knowledge of the procedures in general. The system’s usefulness can be further enhanced by providing it with the capability of knowing about specific instances. With this capability the system could participate in the work in one of two ways: by tracking the progress of tasks, and by carrying out tasks itself. A task tracking facility would allow the system to: * be a source of information regarding the task’s status, history, and plan: * send requests to the agents who are to do the next steps, and make available to them a description of what they are to do what has been done, and pojnters to the relevant documents: * send out reminders and alerts when expected completion times of task steps have passed; and * ask for intervention by the appropriate agent when problems arise. A system which is tracking tasks in this way is participating as a partner in doing the work. Once that symbiotic relationship has been established between system and office worker, there are many steps in office procedures that the system could do itself. Such tasks would certainly include communication activities (e.g., using electronic mail), and maintenance of consistency in structured information bases (e.g., automatically filling in fields of electronic forms, see [Fikes], 1980). Office Work and Office Procedures With this class of intended systems in mind, we now turn to the question of the nature of the office work that we hope to support. In so doing, our goal is to determine the nature and structure of the information needed by our intended systems to support that work. A Procedure Execution Model A common model of office work considers an office worker to be a processor with a collection of tasks to be done and a procedure for doing each task. The work, in this model, involves executing each procedure and “time sharing” among them. However, studies of office work reveal a complexity and diversity of behavior far beyond what would be predicted by this model (For example, see [Suchman], [Wynn], and [Zimmerman]). In this section we explore, the nature of this apparent discrepency as a way of exposing1 characteristics of office work that we think have importam implications in the design of systems to support that work. {Note: The potential usefulness of this discrepancy was suggested to us by [Suchman].} tasks assigned to him. His work involves the planning, scheduling, and context switching associated with time sharing among those tasks. However, he can exercise options in carrying out his scheduling task that are not available to the scheduler in a computerized time sharing system. In particular. he can modify tasks themselves. For example, the worker may choose to * ignore some of the requirements of a task, * reni:gotiate the requirements of a task, * get someone else to do a task, * create and follow a new procedure for doing a task. Iicnce, office work includes, in addition to the carrying out of tasks, the determination of when a task should be done, how the task is to be done, and whether the task will be done at all. {Note: The office worker also has goals other than the completion of assigned tasks. For example, he has career goals (try to get ahead in the company), company goals (maximize profit), personal goals (keep from being bored), social goals (be regarded as good company), and societal goals (be honest).] Second, we take it as obvious that the domain with which office systems must deal is open-ended: truly anything may become relevant to the workings of offices at one time or another. This fact inplies that a procedure which implements a task is necessarily an inadequate description of all the actions which must be done to achieve the task’s goals in all the various situations that can (and inevitably will) occur. That is, at the time the procedure is defined which implements a task, one cannot predict ‘either the range of situations that will be encountered in an office or the extent of the knowledge, activities, and considerations that will be needed to carry out the task. Hence, for any given procedure, situations may occur in which the procedure does not indicate what is to be done, or in which what is indicated in the procedure cannot be done. For example, situations may occur in which: a case analysis in the procedure does not include the current situation, assumptions about the form and availability of inputs for a step are not met. resources required to do a step are not available, the action described in a step will not have the intended effects. The procedures associated with each task serve as a guide in that they indicate one way of doing the task under a particular set of assumptions. The office worker has the responsibility of deciding in each particular situation whether the procedure’s assumptions are satisfied and whether he wants to carry out the task in the way specified by the procedure. Third. the office worker has the problem of interpreting abstract specifications of what is to be done. For example, it is not uncommon in procedure specifications to find phrases like “include any other pertinent data”, “send forms approximately six weeks in advance of the deadline”, and “arrange for employee to receive benefit checks at home”. What is “any other pertinent data”, when is “approximately six weeks in advance of the deadline”, and how is one to “arrange for the employee to receive benefit checks at home”? The specification of the procedure doesn’t say. Hence, a necessary part of the work of following office procedures is determining what the abstract specification implies is to be done in each particular case. We conclude from these observations that the standard model of procedure cxccution is inadequate for describing office work. The original procedure specification serves only as a guide in this process and can be thought of as the first approximation to a plan for the particular task at hand. It is the responsibility of the office worker in each particular case to determine the suitability of the procedure, fill in missing details, and modify it where necessary to achieve the goals of the task. First, consider the office worker’s ongoing (meta-)task of determining how to allocate his resources among the collection of 203 An Example Of Office Work. To make these points more tangible. of the everyday work which goes _ _ let us now look at an example on in an office (This is an elaboration of an actual case of office work reported by [Wynn], p. 49). This example exhibits the problematic nature of the work, and the need for reflecting upon the specifications of the procedures. Xerox sells supplies for its copiers - paper, toner, and such. Customer orders for supplies are taken over the phone by a “customer order entry clerk” (COEC). The COEC talks to the customer and fills out a form which records the order. This order form is used by other clerks to bill the customer and to deliver the supplies. The form has a field for rccordir,g the address at which the copier is located, and there is an underlying assumption that this is the address to which the supplies are to be delivered. In the particular incident of interest, the customer informed the COEC ihat he could not supply an address for the copier because it was located on an ocean-going bargc(!j. This situation, of course, raised the question of what should be put into the address field of the order form. The clerk realized that the intended use of the address was to specify where the supplies were to be delivered, and that because the copier was on a barge that the needed address was dependent upon when the delivery of the supplies was to be made. Since he could not predict that date, he obtained from the customer a telephone number that could be called when the delivery was about to be made to obtain the address of the current location of the barge. He entered that telephone number into the field of the form and added a notation indicating how the number was to be used. The story continues: When the billing clerk was making up the bill, the question arose as to whether or not to charge California sales tax. The answer depends on whether or not the supplies were to be delivered out-of-state. The address field of the order form was examined, as per the usual procedure for answering the question, and of course no information about the state was available. What now? The billing clerk read the notation, called the telephone number, and ask the respondent whether the delivery was to be made in or out of California. Again, the date of the delivery was crucial in determining the answer. However, the billing clerk knew approximately when the supplies would be available, and therefore was able to determine from the person called that the delivery would be made in California, even though the precise delivery address was still not known. An addition was made to the information in the address field of the order form indicating that the delivery was to be made in Califo&a, and the bill was prepared and sent. Finally, the shipping clerk, with the supplies in hand, repeated the telephone call when preparing the shipping label. The address was then known, the address was added to the form, and the supplies were delivered. Analysis of This Example. What we have here is a case of a blown assumption. The iprocedures in which all three of these clerks were playing a role were designed on the assumption that copiers do not move and therefore have a fixed address. The particular case violated that ‘assumption. The COEC ‘was confronted with a problem because he could not carry out a step of a procedure (i.e., he could not fill in an address for the copier). There are several things hc could have done at that point, including ignoring the step or telling the customer that unless he provided an address that the order could not be taken. Instead, he chose to stop “executing” the procedure and to step back and reason about it. In particular, he considered what were the intended uses for the problematical address; i.e.. what was the goal of filling in the form’s address field. Using that information, he created a plan involving both himself and the shipping clerk that was within the spirit, although not within the letter, of the established procedures. That is, he devised an alternative that would satisfy the goals of the intended users of the address, as he perceived them. Hence, those goals were the crucial information that the COEC needed in order to determine suitable alternative actions when the unexpected situation occured. Note that the COEC was apparently not aware of the billing clerk’s use of the address field to determine state sales tax. Hence, the COEC’s alternative plan did not indicate how the billing clerk was to deal with this situation. The billing clerk, like the COEC, was confronted with a problem of not being able to carry out a step in a procedure (because the address field of the order form did not contain an address). Again, as was the case with the COEC, he did not ignore the problem or reject the situation as unacceptable. Instead, he attempted to find suitable alternative actions that would satisfy his task goals and allow the billing to proceed. His planning involved understanding the alternative procedure for the shipping clerk that had been formulated by the COEC, and realizing that he could use the telephone number included in that formulation to satisfy his goals. Consider the nature of the information involved in this example. Note the unpredictabiliry at the time the form was designed of the kinds of information that would be put on the fom:. Note also that the information on the form regarding the address was changing throughout the procedure. First there was a note describing a procedure for obtaining the address, then a parfial address containing only the state added to that note, and finally a complete description of the address. Another form of partial description that played a role in the example was approximafim; in particular, the clerks’ knowledge of the approximate delivery date. The strength and certainty of those approximations determined when and to what extent the delivery address was obtained. Supporting the Work Requires Flexibility We have presented the idea that the work that actually goes on in offIces is not routine. It consists of many particular cases of applying the given procedures to the details of those cases. This work involves dealing with unsatisfied assumptions, doing planning, understanding goals, and using information that is partial, approximate, and changing. The illusion that office workers execute procedures in a manner that is analogous to the way computers execute procedures ignores these realities of the situation. Given that picture of office work, we now turn our attention to the requirements placed on the design of computer-based systems to support such work. A primary design challenge is to find ways of providing the flexibility that is needed to allow for the application of established procedures to the circumstances of particular cases. With respect to information being supplied to the system by users, this flexibility involves dealing with cases where information is missing, information is provided in unexpected forms, and/or information in addition to what was expected is supplied. With respect to the procedural steps being carried out, this flexibility involves dealing with cases where steps are omitted, steps are done in different order, and/or additional steps are done. When office systems lack the flexibility to deal with these contingencies, they severely restrict the options of their users and thereby become yet another bureaucratic barrier to be overcome in “getting the work done”. Consider. for cxamplc. an electronic fcrms system for supporting the work of the COEC. When the “copier on a barge” problem arose, the COEC would hake ncedcd that system to be flexible CI;OL@ to allow entries other than addresses in the form’s address ficid. In particular, the COEC needed to be able to say to the sy%tcm, in effect., “I can’t give you an address in this c&c. Insicad, 171 give you a note for the shipping clerk.” If the syr.tem aI50 used its descriptions of the procedures being followed to provide instmctions to the clerks regarding what is to be done, then the system would need to bc able to accept the COECs decision to omit the step of 204 providing an address in the form’s address field, and to incorporate into the shipping clerk’s procedure an instruction to read the COEC’s note the first time the address was needed. In addition to being able to accept such alternative inputs, any of the system’s facilities for doing computations based on those inputs (e.g., to compute the state sales tax on customer orders using the address on the order form) must be designed to deal with cases in which those inputs have some unexpected structure or are not available at all. The challenge in the design of such processing facilities is to provide ways for the system, in cooperation with the office worker who is being supported, to overcome the difficulty posed by the failed computation so that work on the task can continue in a productive manner. One often hears the argument that this need for flexibility and variation from established procedures could be overcome by doing a more thorough analysis of the office tasks and thereby producing complete procedures that would cover all of the cases that could occur. Our claim is that because of the open-ended nature of the office domain, one cannot anticipate all of the situations that will occur in the carrying out of a given task, and therefore cannot totally characterize the inputs that will be available or the actions that might be taken to satisfy the task’s goals. The Relevance of AI to Supporting the Work An Alternative Model: Planning and Plan Execution The observations we have presented on the characteristics of office work have led us to seek an alternative to the procedure execution model to guide us in building a knowledge base for office support systems. We have found what we think to be a suitable alternative in the paradigms from the AI literature for automatic planning and execution monitoring of plans. That is, we take the viewpoint that we are confronted not so much with the problems of representing and supporting the execution of procedures, but with the problems of representing plans and supporting the monitoring and replanning that occurs during their execution. This viewpoint provides us with a conceptual framework for understanding the use of procedures in an office, an understanding that we feel is critical to dealing with the problems of designing systems to actually support that work. In the following paragraphs we present some of the key aspects of this point of view and discuss the ways in which it suggests that a system could provide useful support. The basic requirement on a data base describing these plans are that they provide the information needed to monitor a plan’s execution and to do whatever replanning might be required. What information is needed during those operations? By referring to the planning paradigm used in the STRIPS systems ([Fikes], et al, 1972), we obtain the suggestions that execution monitoring requires descriptions of the expected results of each operator, the intended use of each operator result, the preconditions of each operator, and the assumptions made by the planner about the world at each step of the plan. Planning involves the use of descriptions of the current state of the world, the operators available as potential plan steps, and the goals to be achieved by the plan. This planning paradigm characterizes some of the information that might be useful in the doing of office tasks and therefore suggests what to include in the description of office tasks and their associated procedures. In particular, it suggests the inclusion of information about task goals, intended uses of operator results, and precondition assumptions of operators. For example, the COEC employed information regarding the intended use of the address by the shipping clerk to determine an alternative plan when the address was not available. If the COEC had also known about the billing clerk’s intended use of the address, then he would have tried to obtain the information needed for that use (i.e., the state in which the delivery would be made) and, if successful, would have eliminated the difficulties that the billing clerk had in the example. One of the major ways we see for a system to provide support is by serving as an information source for office workers -buring the execution of plans and during any replanning that may be required. Hence, the planning p?radigm suggests what information to include in the system’s representation of the tasks and procedures, and provides us with a basis for characterizing the questions that the user may ask of the system. The paradigm of hierarchical planning (e.g., see [Sacerdoti]) also applies here and can be used in our characterization of office work. That paradigm would suggest that we consider each individual step in a plan as being a task with its own inputs, enabling conditions, goals, etc. There may or may not be a plan associated with any given step’s task. In the cases where there are, these plans form a tree and then get combined in various ways to form a planning network. Such a network represents a hierarchical plan, where the top of the hierarchy decribes a top level task and each sucessive level of the hierarchy describes increasingly detailed subtasks. In the standard non-hierarchical planning case, there is a plan for each step and each plan consists of a single operator; hence, there is a one-to-one correspondance between plan steps and operators. In the hierarchical planning paradigm and in the office, that one-to-one correspondance need not exist. Hierarchical planning networks appear to be an important device for representing office plans for several reasons. They are a useM structure for representing the task-subtask and goal-subgoal relationships that need to be known about during execution monitoring and replanning, and they provide the basic descriptive framework for indicating how the work is to be organized. Also, since there are effectively no primitive operators in the office, there is a need for describing office plans at varying levels of detail, depending on the specific needs of the describer and users of the descriptions. That flexibility in the level of detail of specification is therefore needed in an office system’s representation facilities. The system can then be involved in the office work at varying levels of detail. For example, the system may know that a travel request needs to be authorized, but know nothing about the subtasks involved in obtaining the authorization. Such flexibility is also an important tool for enabling the system to participate in situations that it does not understand. For example, if a plan that the system is helping to monitor fails, the user may not describe to the system the alternative plan he decides to use. However, the system knows about the goals of the original plan that the alternative must also satisfy and can therefore monitor the accomplishment of those goals, ev;;evz~gh it now has no model of how those goals are being An important way in which office work motivates extension of current AI planning paradigms is that office work is done in a multi-processor environment, That is, one can consider each agent (person or system) in an office to be a processor that is accomplishing tasks by carrying out plan5 that have either been given to it or created by it. Any given plan may involve the participation of several agents, each asent acting as an independent processor executing one or more of the plan’s steps. The processors make commitments to each other regarding the goals th.ey will achieve during the execution of a plan [Flores]. Therefore, in creating a plan to be carried out by multiple agents, the commitments that those agents will make to each other are a crucial part of a multi-processor plan. Furthermore, the commitments an agent has made and that have been made to him form for him a set of constraints within which he must work. In particular, these commitments form the context in which replanning takes place. in that any new plan must satisfy those commitments. However, the agent also has the options during replanning of renegotiating the commitments he has made or of ignoring them altogether. 205 Any system that is to participate in the replanning process needs to support this commitment-based view of planning and take them into consideration. In particular, a system could help an agent keep track of the commitments he is involved with that relate to a particular task and indicate his options for changing them during replanning. To support this tracking, the system’s plan representation needs to include for each plan step both the commitments made by and to the agent responsible for doing the step. The multi-processing nature of office work also implies that steps of a plan can be done in parallel. Hence, representations for office plans need to allow specification of partial orderings for plan steps. That requirement and the pervasiveness of replanning that occurs during plan execution suggest that task descriptions should include a set of necessary and sufficient “enabling conditions” for beginning the task so that it can be determined when the task can be begun irrespective of the order or nature of the steps that achieved those conditions. Up to this point we have not considered perhaps the most immediate question that arises out of looking at office work from a planning point of view: to what extent can we expect a system to automatically do the planning and replanning that is needed for office tasks? The primary limitation on such automatic planning seems to be the open-endedness of the office domain. That is, the extent to which there are considerations relevant to the formation of the plan of which the system has no understanding will limit the system’s effectiveness at determining an appropriate plan. For example, the system may not know about possible operators in the situation, the costs or chances for success of renegotiating or ignoring existing commitments, other goals that interact with the task’s goals, or the implications of an unexpected situation. These limitations have lead us to focus on a symbiotic relationship between system and office worker during planning and replanning, ,where the system plays primarily a role of supporting the planning /being done by the users by helping represent., manage, and communicate the resulting plans. In conclusion, then, we are claiming that * a multi-processor hierarchical planning model is B useful one for understanding office work and therefore for structuring an offlice system’s knowledge base, and * the demands of supporting offlce work motivate new research in multi-processor commitment-based planning. Knowledge Representation Challenges We turn now to the demands that supporting office work makes on the representation of the knowledge which systems have of the ‘office domain. We then discuss two techniques that arise from work in the AI community that provide particularly promising starting points for confronting those demands. The single most salient demand of such representation schemes is that they be able to respond to the need for change in conceptualization of the work. As we have seen, the domain of office work is inherently open-ended (e.g., before the barge case there was no notion of addresses being time-dependent: afterwards there was). Consequently there is no way to anticipate the full range of subject matter with which the system will have to deal. In consequence, the representation scheme must be able to handle any conceivable conceptualization which, over the course of the life of the office, the users of the office system choose to enter into the system. Furthermore, as time passes, this conceptualization will change to meet the changing understanding of the office domain which the users of the system have. Sometimes these changes will be small; at other times, there will be major “re-thinkings” of the information. The system must not only be able to represent this changing pattern of thought, but must also be able to simultaneously represent the pattern of changing thought: to support office work, it will have to be able to support the history of what has happened previously, and consequently will have to be able to hold simultaneously the old conceptualizations - for supporting the understanding of the past, and the new. The second demand placed by offlice systems on the mechanisms for representing the knowledge which is within them is to support partial knowledge of their domain. This incompleteness comes in at least three forms: the support of a subset of some expected body of knowledge (e.g., the state in which an address is may be known, but nothing more): the support of an abstraction of the knowledge (e.g., the supplies ordered are a paper product, but which one is unknown); and the support of an approximation (e.g,. the date of delivery is between mid-April and mid-May). In particular, this ability to support partial knowledge will permit the entry into the system of all that one knows, even though that may only be part of what is desired in a complete description. A resultant demand is that the mechanisms which access information must be prepared for the expected information not to be there (e.g.,. the state portion of an address is missing from the available information when the billing clerk tries to bill the barge- ownei). This preparation involves the representation scheme in at least being able to detect the absense, and further, in having some means of coping with the resulting problems. The third demand results from the fact that knowledge of things /often accumulates over time. Sometimes such an accumulation of partial descriptions can be reformed into a coherent whole. But more often the pieces are better retained as independent, un- coordinated facts. Indeed, rather than think about OX description of an entity, it is often useful to view the object as having mulfiple descriptions. Thus, for example, the knowledge about the address of the copier might at some point include three distinct descriptions of the address: as having California as its state portion, as being a changing thing, and as being something which can be fkthur determined by carrying out the procedure “call this number and ask”. The final demand arises from the expectation that the system should provide ‘a general model of office work. This model could be crafted by experts on the organizational structuring of oflices, and would then be available as a conceptual framework to support the description of more particular details. In fact, these concepts become the zerrns in which the details are not only described, but understood. Thus, for example, the concepts of task, goal, procedure, plan, agent, post, authorization, commitment, and data repository might be provided as a very general framework for modeling offices. Particular offices would have their own particular tasks, goals, etc. These demands pose challenging research problems in knowledge representation which we are not claiming to have solved. However, we discuss in the following paragraphs two starting points for confronting these problems that look particularly promising to us and that we are using in our work. Our first starting point for responding to the demands of supporting office work is the use of a specialization-based knowledge representation formalism (see, for example, [Brachman]). This, and similar, schemes for formally and precisely representing knowledge take as their goals the first three of our needs: representation of a changing support for conceptual structure, description, and multiple description. partial The major structuring principle for their representations is defnilion 6~ specialization: any concept in a knowledge base can be taken as a basis for defiriing a new concept which is a special case of the old one. Thus, describing the details of a particular office can be done in this formalism by specializing more general descriptions of offices. This specialization can be done in steps, thus permitting the tailoring of the conceptualization in various ways to produce progressively less abstract descriptions. In the end, the most specific descriptions are understood through their place in a taxonomic lattice (a concept can specialize more than one abstraction) of more abstract concepts. Not only does this enrich the understanding of the domain, for one can understand similarities between concepts in terms of common abstractions, but it also provides locations for attaching knowledge about abstractions which will immediately apply to all special cases of those abstractions. Our second response to the knowledge representation needs presented by the offlice domain is the use of semi-formal descriptions. That is, WC are using the formal knowledge representation mechanisms in a style which permits us to capture information that the system dots not “understand” in such a way that it can be uscfmly employed b:i human users of the system. For example, paragraphs of English prose or diagrams can be associated with concepts: the only use the system will be able to make of them will be to present them to a human user, and permit him to read and modify them. We view these mixtures of formal (the structure is understood by the system) and informal (the structure is understood only by humans) descriptions as an essential “escape valve” for our knowledge representation systems: if it were required that the system had to understand the conceptualization underlying all the informaion in its knowledge structure, then the cost of entering information which is currently beyond the system’s understanding would be very high. Instead, by escaping into informal description, the system can still be used as a repository for all of the information about the situation at hand, and yet permit the work to proceed. The system becomes primarily a communications . device for supporting inteiaction between people carrying out office work. However, because the informal descriptions are represented in the same description formalism as the formal descriptions, they can be integrated into the knowledge base in a consistent manner. An Approach to System Building: Symbiotic Processing In the account given above of our vision of the behavior and properties of a system to support the real work of offices, there has been repeated reference to interaction between the system and its human users. This will be’ an important aspect of successfully completing the planning which must be done in office work. Also in representing knowledge within the system, we have argued that a “semi-formal” mixture of information will be important for achieving practical systems. That is, it takes both man and machine to understand the information held within the system. Changing the emphasis, we prefer to think of the humans and the computer as cooperating processing engines carrying out the “computations” of the system in partnership. The idea here is that both processing engines, each with its own processing capabilities, knowledge, and memory structures. are essential to getting the task done. Neither could effectively do the work without the other. In biology, such interdependence is called symbiosis, and a system composed of two or more interdependent organisms is called a symbiolic system. We therefore use the same term to refer to the sort of offrce systems envisioned here: there is a symbiosis of human and machine. Why is it that this quite obvious co-operation between man and machine has not so far been the dominating pattern of computer use in office (as well as other) systems? Our theory is this: when batch processing was the only economically feasible form of business (and therefore office) system, such interaction was impossible - the human partners were simply not around when they were needed to help the computer in its tasks. However, the pervasiveness of the belief among designers of computer systems that ale procedures in offices were “routine” obscured the need for truly cooperative interaction. A result of this belief, and of the introduction and widespread use of batch processing systems in the business environment, has been the establishment and buttressing of the pow-accepted notion that there is something fundamental about partitioning the world into routine cases and exceptions. We view this distinction between routine and exception as quite artificiai: routine cases are those cases where no action is required of the human partner in the symbiosis: exceptions are everything else. And, in fact, it is even worse than that: the distinction breeds viewing the world this way, which in turn enhances the distinction. To get the proportion of routine cases up to the point where batch systems could be justified, many cases which require some small amount of human processing are handled by forcing the human processing to be done before the cases are “entered into” the system. This is often done at the expense of some capacity of the system (as a whole) to handle not-quite-standard cases: these cases are wedged into the mold of the “routine”. And when exceptions are not even permitted - when everything has to go through the system as a routine case, the mold can well become a straight-jacket. We propose that office systems need not make this distinction between the routine and the exceptional. Instead, it should be possible, and is desirable, to return to the “good old days” when all cases were processed in the same way, some with more effort than others. We believe that, armed with the understanding of ofliccs presented here, and supported by studies in the AI fields of automatic planning and knowledge representation, a modem version of the “good old world” can be achieved through systems built around the notion that all cases are handled by the poweri% symbiosis of humans and machines. Acknowledgements We would like to thank Lucy Suchman and Eleanor Wynn for early contributions to the ideas in this paper. In particular, they made available to us transcripts of their interviews with office workers and observations of office work, and helped open our eyes to what those interviews and observations had to say about offlice work. We would also like to thank Lucy for her insightful participation in the continuing research that has led to the observations in this paper. References Brachmap, R. et al. KLONE Reference Manual, BBN Report No. 3848, July 1978. Fikes, R. E., Hart, P. E., and Nilsson, N. J. “Learning and Executing Generalized Robot Plans”. Artificial Intelligence, 3(4), winter 1972, pp 251-288. Fikes, R. E. “Odyssey: A Knowledge-Based To appear in ArtiJicial Intelligence. Personal Assistant”. Flores, F. Univ. of California at Berkeley, personal communication. Sacerdoti, E. D. A Structure for Plans and American Elsevier, 1977. Behavior, New York: Suchman, L. A. “Office Procedures as Practical Action: A Case Study”. SSL Internal Report, Xerox Palo Alto Research Center, Sept. 1979. WY% Eleanor Herasimchuk “Office Conversation as an Information Medium”. Department of Anthropology, University of California at Berkeley, May 1979. Zimmerman, D. H. “The Practicalities of Rule Use”. in J. D. Douglas (Ed.), Underslanding Everyday Life, Aldine Publishing Company, Chicago, pp 221-237. 207

1980

53

49

Metaphors and Models Michael R. Gcnesereih Computer Science Department Stanford University Stanford, California 94305 1. Introduction Much of one’s knowledge of a task domain is in the form of simple facts and procedures. While these facts and procedures may vary from domain to domain, there is often substantial &nilarity in the “abstract structure” of the knowledge. For example, the notion of a hierarchy is ‘found in biological taxonomy, the geological classification of time, and the organization chart of a corporation. One advantage of recognizing such abstractions is that they can be used in selecting metaphors and models that are computationally very powerful and efficient. This power and efficiency can be used in evaluating plausible hypotheses about new domains and can thereby motivate the induction of abstractions even in the face of partial or inconsistent data. Furthermore, there is a seductive argument for how such information processing criteria can be used in characterizing “intuitive” thought and in explaining the cogency of causal arguments. The idea of large-scale, unified knowledge structures like abstractions is not a new one. The gestalt psychologists (e.g. [Kohler]) had the intuition decades ago, and recently Kuhn [Kuhn], Minsky [Minsky], and Schank [Schank & Abelson] have embodied similar intuitions in their notions of paradigms, frames, and scripts. (See also [Bobrow & Norman] and [Moore & Newell] for related ideas.) The novelty hcrc lies in the use of such structures to select g6od metaphoa and models and in the effects of the resulting power and efficiency on cognitive behavior. This paper describes a particular formalization of abstractions -in a knowlcdgc rcprcsentation system called hNAr.OG and shows how abstractions can be used in model building, understanding and generating analogies, and theory formation. The prcscntation here is necessarily brief and mentions only the highlights. The next section defines the notions of abstraction and simulation structure. Section 3 describes the use of abstractions in building computational models, and section 4 shows how abstractions can bc used to gain power as well as efficiency. 2. Abstrrrclions and Sirrwlalion Struclures Formally, an abslrucfion is a set of symbols for relations, functions, constants, and actions together with a set of axioms relating these symbols to each other. Abstractions include not only small, simple concepts like hierarchies but also more complex notions like concavity and convexity or particles and waves. A model for an abstraction is essentially an interpretation fcr the symbols that satisfies the associated axioms. Different task domains can bc mod& of the same abstraction (as biological taxonomy, geological time, and organization charts are instances of hierarchies); or, said the other way around, each abstraction car have a number of different models. Importantly, there are multipk computational models for most abstractions. In order to distinguish computer models from the task domains they are designed tc mimic, they are hereafter termed simulufion slrucfures, following Weyhrauch [Weyhrauch]. There is a strong relationship between abstractions and metaphors, or analogies. Many analogies are best understood as statements that the situations being compared share a common abstraction. For example, when one asserts that the organization chart of a corporation is like a tree or like the taxonomy of animals in biology, what he is saying is that they are all hierarchies. With this view, the problem of understanding an analogy becomes one of recognizing the shared abstraction. Of course there are an infinite number of abstractions. What gives the idea force is that the simulation structures for certain abstractions have representations that arc particularly economical, algorithms that are particularly efficient, or theorems that are particularly powerful, e.g. hierarchies, grids, partial orders, rings, groups, monoids. Consequently, there is advantage to bL gained from recognizing the applicability of one of these special abstractions rather than synthesizing a new one. Even when the applicability of such special abstractions and simulation structures cannot be determined with certainty (say, in the face of incomplete or faulty information), there is advantage in hypothesizing them. Until one is forced to switch abstractions due to incontrovertible data, one has an economical representation and powerful problem solving methods. By biasing the early choice of abstractions in this way, these criteria can have qualitative effcctc on theory formation. 3. Models The importance of abstractions and their associated measures of economy, efficiency, and power is clearest in the context of a concrete implementation like the ANALOG knowledge reprcscntation system. l’hc interesting feature of ANALOG is that it utilizes a variety of simulation structures for representing different portions of the knowledge of a task domain. This setup is graphically illustrated in figure 1. The user asserts facts in the system’s uniform, domain-independent formalism, and the system stores them by modifying the appropriate simulation structure. Facts for which no simulation structure is appropriate are simply filed away in the uniform representation. (ANALOG currently uses a semantic network representation called Dl3 [Gcnesereth 761. The formalism allows one to cncodc scntcnccs in the predicate calculus of any order and provides a rich meta-level vocabulary.) Descriptions of each of ANALOG'S abstractions and simulation structures arc also cncodcd within the DU rcprescntation. I ASS1 Figure 1 - An Overview of ANALQC This approach departs from the custom in knowledge representation systems of using uniform, domain independent formalisms. While there are advantages to uniformity, in many cases the representations are less economical than specialized data structures, and the associated general procedures (like resolution) are less efficient or less powerful than specialized algorithms. For 208 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. example, a set in a small universe can be efficiently represented as a bit vector in which the setting of each bit determines whether the corresponding object is in the set. Union and intersection computations in this representation can be done in a single machine cycle by hardware or microcoded boolean operations. By contrast, a frame-like representation of sets would consume more space, and the union and intersection algorithms would have running times linear or quadratic in the sizes of the sets. The distinction here is essentially that between “Fregcan” and “analogical” representations, as described by Balzer [Balzer]. Note that ANALOG'S approach is perfectly compatible with uniform knowledge representation systems like DB and RLL [Greiner & Lenat]. The addition of abstractions and simulation structures can be viewed as an incremental improvement to such systems, and their absence or inapplicability can be handled gracefully by using the uniform representation. It’s important to realize that ANALOG is not necessarily charged with inventing these clever representations and algorithms, only recognizing their applicability and applying them. The approach is very much in the spirit of the work done by Green, Barstow, Kant, and Low in that there is a knowledge base describing some of the best data representations and algorithms known to computer science. This knowledge base is used in selecting good data representations and efficient algorithms for computing in a new task domain. One difference with their approach is that in ANALOG there is a catchall representation for encoding assertions when no simulation structure is applicable. Other differences include an emerging theory of representation necessary in designing new simulation structures (see section 3.2) and the use of the criteria of economy, efficiency, and power in theory formation. ANALOG'S use of simulation structures is in a very real sense an instance of model building. Architects and ship designers use physical models to get answers that would bc too difficult or too expensive to obtain using purely formal methods. ANALOG uses simulation structures in much the same’ way. In fact, there is no essential reason why the simulation structures it uses couldn’t be physical models. Furthermore, as VLSI dissolves what John Backus calls the vonNeumann bottleneck, the number of abstractions with especially efficient simulation structures should grow dramatically. 3.1 Building a Model As an example of modeling, consider the problem of encoding the ofganization chart of a corporation: ‘[‘he first step in building a model for a new task domain is finding an appropriate abstraction and simulation structure. The knowledge engineer may directly name the abstraction or identify it with an analogy, or the system may be able to infer it from an examination of the data. In this case, the hierarchy abstraction is appropriate, and there are several appropriate simulation structures. One of these is shown in figure 2. Each object in the universe is represented as a “cons” cell in which the “car” points to the object’s parent. The relation (here called Rel) is just the transitive closure of the Car relation, and Nil is the root. For the purposes of this example, the “cdr” of each cell may be ignored. ,++$$yj&$$ Figure 2 - A Simulation Structure for the Hierarchy Abstraction An important requirement for a simulation structure is that it be modifiable. Therefore, it must include actions that the model builder can use in encoding knowledge of the task domain. Usually, this requires the ability to create new objects and to achieve relations among them. In this case, the Neons subroutine creates a new object, and Rplaca changes an object’s parent. Part of the task of finding an appropriate abstraction and simulation structure is setting it up for use in encoding knowledge of the task domain. This includes three kinds of information. The first is an index so that the system can determine the simulation structure appropria‘tt: to a new assertion. (This index is necessary since several domains and simulation structures may be in use simultaneously). Secondly, there must be a procedure for mapping each assertion into its corresponding assertion about the simulation structure. And, finally, the system must have information about how to achieve the new assertion. Once the simulation structure is chosen and set up, the system builder can begin to assert facts about the task domain, and the system will automatically modify the simulation structure. As an example of this procedure, consider how the system would handle the assertion of the fact (Boss-of Carleton Bertram). First, it would use its index to determine that the simulation structure of figure 2 is being used and to recover the mapping information. ‘I’hcn it would map the assertion into the simulation domain. In this cast, let’s say that Arthur is the boss of Bcrtrm and lhdrice while Carleton has been installed in the model as Ileatrice’s employee. Then the new assertion would bc (He1 ((Nil)) (Nil)), where the first argument is the object rcprcsenting Carleton and the second represents Bcrtram. By examining the mcta-level information about the simulation structure, the system retrieves a canned proccdurc (Rplaca) for achieving this fact and cxecutcs it, with the result that Carleton’s “car” is redirected from Bcittr!.ce to Bcrtram. An intcrcsting aspect of model building is that complete information is often required. For example, in adding a node to the simulation structure of figure 2, the system must know the object’s parent in order to succeed. (It has to put something in the “car” of the cell.) This problem can sometimes be handled by the addition of new objects and relations that capture the ambiguity. For example, one could add the special token Unknown as a place filler in the simulation structure above. (Of course, the resulting structure would no longer be a hierarchy.) Another example is using the concept of uncle as a union of father’s brother and mother’s brother. Unfortunately, this approach increases the size of the model and makes deductions more difficult. Unless there are strong properties associated with such disjunctive concepts, it is usually better to carry the ambiguity at the meta-level (i.e. outside the model, in the neutral language of the knowledge representation system) until the uncertainty is resolved. Another interesting aspect of the use of simulation structures is the automatic enforcement of the axioms of the abstraction. For example, in the simulation structure of figure 2, it is impossible to assert two parents for any node simply because a “cons” cell has one and only one “car”. Where this is not the case (as when a simulation structure is drawn from a more general abstraction), thy axioms can still be used to check the consistency and completeness of the assertions a system builder makes in describing his task domain. For example, if the system knew that a group of assertions was intended to describe a hierarchy, it could detect inconsistent data such as cycles and incomplete data such as nodes without parents. 3.2 Designing a Simulation Sfructure for an Abslraclion The only essential criteria for simulation structures are representational adequacy and structure appropriate to their abstractions. For every assertion about the task domain in the language of the abstraction, there must be an assertion about the simulation structure; and the structure must satisfy the axioms of the abstraction. In creating a simulation structure, one good heuristic is to try to set up a homomorphism. Sometimes, the objects of the simulation structure can be used directly, as in the case of using “cons” cells to represent nodes in a hierarchy. In the example above, the mapping of objects from the corporation domain into the domain of list structure was one-to-one, i.e. the corporate objects were all rcprcsentcd by distinct pieces of list structure, and 209 the relations and actions all mapped nicely into one another. Of course, this need not always be the case. Consider, for example, the state vector representation of the Blocks World proposed by McCarthy, in which the Supports relation bctwccn each pair of blocks is represented by a distinct bit in a bit vector. (Think of the vector as a matrix in which the Ci, jxh bit is on if and only if block i is on block J). In this representation the fact (Supports A B) would translate into something like (On Bit-AR Vector-l), and theie would be no distinct representations of the blocks A and B. In other cases, more complex objects may be ncccssary in order to provide enough relations. When a domain does not provide an adequate set of relations, it’s a good idea to synthesize complex structures from simpler ones. For example, a simulation structure’ for an abstraction with three’ binary relations could be built in the list world by representing objects as pairs of “cons” cells in which the “car” represents the first relation, the “cd? points to the second cell, the “cad? represents the second relation, and the,“cddr” represents the third. This approach is facilitated by programming languages with extensible data structures. Obviously, it pays to economize by using the prcdefined relations of the simulation domain where possible. For example, a good representation for a univariate polynomiat is a list of its coefficients, and one gets the degree of the polynomial for free (the length of the list minus 1). One advantage is representational economy; another is automatic enforcement of the abstraction’s axioms, as described in the last section. In order to use a simulation structure, it may be necessary to transform objects into a canonical form. For example, one can represent a univariate polynomial as a list of coefficients, but the polynomial must be in expanded form. There is a large body of literature on canonical forms for specific algebraic structures, while [Genesereth 791 gives a general but weak technique for inventing such forms directly from an abstraction’s axioms. In using a simulation structure, there is a tradeoff betweeh the amount of work done by the model and the amount done by the knowledge representation system. For example, in the simulation structure of figure 2, one must loop over the parent relation to determine whether two objects are related. This can be done either by the knowledge representation system or by a L'ISP procedure in the simulation structure. Obviously, it’s a good idea to have the simulation structure do as much work as possiblq, 3.3 Interfacing Simulation Stniclures For most interesting task domains, the chances are that a single simulation structure is not sufficient. In such cases, it is sometimes possible to piece together several different simulation structures. The simplest situation arises when the objects of the task domain form a hierarchy under the “part” relation. Then one can choose one representation for the “topmost” objects and a different representation for the parts. The spirit of this approach is very similar to that of “object-oriented” programming in which each object retains information about how it is to be processed.’ One disadvantage of this approach is that each object must have explicit “type” information stored with it. Barton, Genesereth, Moses, and Zippel have recently developed a scheme that eliminates this need by separating the processing information from each object and passing it around in a separate “tree” of operations. ANALOG uses this schcmc for encoding the operations associated with each simulation struct’ure. Task domains with several relations are sometimes decomposable into several abstractions, and these relations can then be represented independently. More often the relations are interdependent; and, when this is the case, the interdependence must be dealt with in the uniform representation, Even when a single abstraction would fit the task domain, it may be advisable to use several. Consider, fw .example, a partial order that nicely decomposes into two trees. Furthermore, there are often advantages to multiple representations of objects, as argued by Moses [Moses]. 4. Thinking With Abstractions and Simulation S@uclures The USC of specialized simulation structures gives ANALOG an economy and efficiency not possible with a uniform representation. The economy can be expressed in terms of the space saved by representing assertions in the simulation structure rather than the uniform representation. This economy derives from the elimination of the overhead inherent in uniform formalisms and the use of relations implicit in the simulation structure (as the length of a list reflects the dcgrce of the polynomial it represents). The efficiency refers to the time involved in doing dcd*:rtions and solving problems. This efficiency may be attributable to clever algorithms, or it may be the result of long familiarity with the domain from which the abstraction evolved (due to the memory of many special case heuristics). Lenat [Lenat et. al.] discusses how a computer might improve its own performance by self-monitoring. An interesting pos,*bility suggested by this economy and efficiency is for the program to use these criteria in evaluating plausible hypotheses about a new domain. In the face of incomplete or contradictory data, the program should favor the more economical abstraction. Clearly, there is some evidence for this sort of behavior in human cognition. Consider, for example, Mendeleev’s invention of the periodic table of the elements. He was convinced of the correctness of the format in spite of contradictory data, for reasons that can only be identified as simplicity. These criteria of economy and efficiency are also of use in characterizing why it is easier to solve problems from one point of view than another, e.g. proving a theorem using automata theory rather than formal grammars. Part of what makes causal arguments (see [deKleer] for example) so compelling is that they are easy to compute with. The reason for this is that a causal argument is an instance of a cognitively efficient abstraction, namely a diicctcd graph. One is tempted, therefore, to generalize dcKleer’s notion of causal envisionment as finding economical and efficient abstractions (perhaps identified with analogies) in which the desired conclusions are reached via simple computations. The idea can be carried a bit fUrther and generalized to include the criterion of problem solving power. In particular, one should favor an abstraction for its ability to solve a pending problem despite insufficient data. The obvious difficulty is that the assumption may bc wrong or there may be scvcral abstractions that are equally probable and useful. Consider, for example, the following arguments for determining the distance between the observer and the middle vertex of a Necker cube. “Well, the lines form a cube, and so the middle vertex must be closer to me than the top edge.” “No, not at all, the figure is concave, and so the middle vertex must be filrther away.” Both arguments are consistent with the data and refer to a single abstraction, 2nd in each case the conclusion is deductively related to that view. A second example is evident in the particulate-wave controversy. The particulate view is a simple abstraction that accounts for much of the data and allows one to solve outstanding problems. Of cours6, the same can be said for the wave view. Unfortunately, the predictions don’t agree. A similar argument explains the inferential leap a child makes in declaring that the wind is caused by the trees waving their leaves. When the child waves his hand, it makes a breeze; the trees wave when the wind blows; so they must have volition and motive power; and that would account for the wind. The reasoning in these examples is usually termed “analogical”. The key is the recognition of a known abstraction common to the situations being compared. This conception of analogy differs markedly from that of Hayes-Roth and Winston. In their view two situations are analogous if there is any match between the two that satisfies the facts of both worlds. If the match is good, the facts or heuristics of one world may be transferred to the other. The problem is that these facts may have nothing to do with the analogy. Just because two balls are big and plastic, one can’t infer because one ball is red that the other is also red. Abstractions are ways of capturing the necessary interdependence of facts. For example, the size and material of a ball do affect its mechanical behavior, and so the skills usefU1 for 210 bouncing one should be of value in bouncing the other. Also notL that the match need not be close in order for there to be a useful analogy. Linnaean taxonomy and organization charts have few superficial details in common, but the analogy is nonetheless compelling, and as a result the algorithms for reasoning about one can bc transferred to the other. The work of Hayes-Roth and Winston is, however, applicable where no abstractions exist yet. Their matching algorithms and the techniques of Buchanan, Mitchell, Dietterich and Michalski, and Lenat should be important in inducing new abstractions. An important consumer for these ideas is the field of computer-aided instruction. There is a current surge of interest in producing a “generative theory of cognitive bugs” (see [Brown], [Genesereth 80a], and [Matz]). The use of abstractions and the criteria of economy, efficiency, and power in theory formation is very seductive in this regard. Unfortunately, there is no reason to believe that the hardware of a vonNeumann computer in any way resembles the specialized capabilities of the human brain. (Indeed, psychologists are still debating whether there are any analogical processes in the brain at all. See, for example, [Kosslyn & Pomerantz], [Kosslyn & Schwartz], pylyshyn], and [Shepard & Metzler].) Thus, the idea at present is not so much a model for human cognitive behavior as a metaphor. 5. Conclusion The ANALOG system was dcvclopcd over a period of time to test the ideas presented here. One program accepts an analogy and infers the appropriate abstraction; another builds a model of the task domain as assertions are entered; and a third uses the model to answer questions. There is a sketchy implementation of the simulation structure designer, but no effort has been made to build the theory formation program. In summary, the key ideas are (1) the role of abstractions in understanding metaphors and selecting good models for task domains, (2) the use of’ models to acquire economy, efficiency, and problem solving power, and (3) the-importance of these criteria in theory formation. Abstractions and simulation structures make for a knowledge representation discipline that facilitates the construction of powerful, efficient AI programs. The approach suggests a program for much future work in AI and Computer Science, viz. the identification of us&11 abstractions and the implementation of corresponding simulation structures that take advantage of the spccia! computational characteristics of the vonNeumann machine and its successors. Acknowledgemenfs The content of this paper was substantially influenced by the author’s discussions with Bruce Buchanan, Rick Hayes-Roth, Doug Lenat, Earl Sacerdoti, and Mark Stefik, though they may no longer recognize the ideas. Jim Bennett, Paul Cohen, Russ Greiner, and Dave Smith read early drafts and made significant suggestions to improve ‘he presentation. grants from ARPA, NLM, and ONR. The work was supported in part by References Balzer. R. Automatic Programming, Institute Technical Memo, Southern California/ Information Sciences Institute, 1973. University of Barstow, D. R. Knowledge Bused Program Construction, Elsevier North-Holland, 1979. Brown, J. S. & vanlehn, K. forthcoming paper on learning. Buchanan, B. & Feigenbaum. E. A. Dendral and Meta-Dendral: Their Applications Dimension, Arfificial InteUigence, Vol. 11, 1978, pp 5-24. . deKleer, J. The Origin and of the Sixth International 197-203. Resolution of Ambiguities in Causal Arguments, Proc. Joint Conference on Artificial Intelligence, 1979. pp Hayes-Roth, F. & McDermott An Interference Matching Technique for Inducing Abstractions, Comm of rhe ACM, Vol. 21 No. 5. May 1978. pp 401-411. Genesereth. M. R. A Fast Inf&ence Algorithm for Semantic Networks, Memo 5, Mass. Inst. of Tech. Mathlab Group, 1976. Genesereth, M. R. The Canonicality of Rule Systems, Proc. of the 1979 Symposium on Symbolic and Algebraic Manipulation, Springer Verlag, 1979. Genesereth, M. R. The Role of Plans in Intelligent Teaching Systems, in Inrclligent Teaching Systems, D. Sleeman, ed. 1980. Genesereth, M. R. & Lenat, D. B. Self-Description and Self-Modification in a Knowledge Representation System, IIPP-880-10. Stanford University Computer Science Dept., 1980. Green, C. C. The Design of the PSI Program Synthesis System, Proc. of the Second International Conference on Software Engineering, Oct. 1976, pp 4-18. Greiner, R. D. & Lenat, D. B. A Representation Language Language, submitted for inclusion in the Proc. of the First Conference of the American Association for ArtiIicial Intelligence, Aug. 1979. Kant, E. Efficiency Considerations in Program Synthesis: A Knowledge Based Approach, doctoral dissertation, Stanford Univ. Computer Science Dept., 1979. K_ohler. W. Ges!nJc P.weholnPv: An Introduction to New Concepts in Modem Psychology, Liveright, 1947. Kosslyn, S. M. & Pomcrantz. J. R. Imagery, Propositions, and the Form of Internal Representations, Cognifive Psychofogy, Vol. 9 No. 1, 1977, pp 52-76. Kosslyn, S. M. & Schwartz, S. P. A Simulation of Visual Imagery, Cosnitive Science, Vol. 1, 1977, pp 265-295. Kuhn, T. The Structure of Scientific Revolutions, Univ. of Chicago Press, 1962. Lenat, D. B. Automated Theory Formation in Mathematics, Proc. of the Fifth International Joint Conference on Artificial Intelligence, 1977, pp 833-842. Lenat, D. B., Hayes-Roth, F., Klahr. P. Cognitive Economy in Artificial Intelligence Systems, Proc. of the Sixth International Joint Conference on Artificial Intelligence, 1979. pp 531-536. Low, J. R. Automatic Coding: Choice of Data Structures, ISR 15. Birkheuser Verlag, 1976. Matz, M. A Generative Theory of High School Algebra Errors, in Intefligenf Teaching Systems, D. Sleeman. ed. 1980. McCarthy, J. Finite State Search Problems, unpublished paper. Minsky, M. A Framework for Representing Knowledge, in The Psychology o/ Computer Vision, P. H. Winston, ed., McGraw-Hill, 1975. Mitchell, T. Version Spaces: A Candidate Elimination Approach to Rule Learning, Proc. of the Fifth International Joint Conference on Artificial Intelligence. 1977. Moore, J. & Newell, A. How can MERLIN Understand, in L. W. Gregg, ed. Knowledge and Cognition, Lawrence Erlbaum, 1974. Moses. J. Algebraic Simplification: A Guide for the Perplexed, Comm of rhe ACM, Vol. 14 No. 8, 1971, pp 527-537. Pylyshyn. Z. W. What the Mind’s Eye Tells the Mind’s Brain: A Critique of Mental Imagery, Psychological Bullerin, Vol. 80, 1973, pp l-24. Schank. R. & Abelson, R. Scripts, Plans, and Knowledge, Proc. of the Fourth Internatylonal Joint Conference on Artificial Intelligence, 1975, pp 151-157. Shepard, R. N. & Metzler, J. Mental Rotation of Three-dimensional Objects, Science. Vol. 171, 1977, pp 701-703. Thorndyke, P. W. & Hayes-Roth, B. The Use of Schemata in the Acquisition and Transfer of Knowledge, Cognifive Psyckology Vol. 11, 1979, pp 82-106. Weyhrauch, R. Prolegomena to a Theory of Formal Reasoning, STAN-CS-78-687, Stanford Univ. Computer Science Dept., Dec. 1978. Winston, P. H., Understanding Analogies, Mass. Inst. of Tech. Artificial Intelligence Laboratory, Apr. 1979. Dietterich, T. G. & Michalski, R. S. Learning and Generalization of Characteristic Descriptions: Evaluation Criteria and Comparative Review of Selected Methods, Proc. of the Sixth International Joint Conference on Artificial Intelligence. 1979, pp 223-231. 211

1980

54

50

EVERYTHING YOU ALWAYS WANTED TO KNOW ABOUT AUTHORITY STRUCTURES BUT WERE UNABLE TO REPRESENT James R. Meehan uept. of Information and Computer Science university of California lrvine CA 92717 If we're ever to get programs to reason intelligently about activities that are governed by laws, we’ll need a clear understanding of such concepts as obliga- tions, politics, disputes, and enforce- ment. Our work on this topic started when we tried to build a program that could understand some legal correspondence from l'i'th-century England, and its current application is in a program that simulates social interaction among small groups such as families and schools. While the long-range applications are in natural language processing, this work is primarily concerned with fundamental issues of representation. we first attempted to adapt the work by Schank and Carbonell [I,21 but later chose alternative definitions and representations, because we were looking at a wider variety of legal systems, where legal authority, for example, did not always imply the power of enforcement, and while they focused on the resolution of disputes, we were more interested in how aecisions are influenced. Authority structures can be found everywhere, in law courts, AI conferences, bridge games, friendships, and even restaurants. The language is used meta- phorically even in such domains as everyday physics. We define an authority structure in terms of a group of people who participate in some set of activities. There are many structures one could impose on a group for purposes of analysis, such as the sequencing of events, or the plans and goals that motivate the participants, and different contexts will provide different answers to the same question. For example, if we ask, "Why did John order a hamburger?" we might answer with '-because he was hungry" or "So that the waitress would tell the cook" or "In order to initiate a contract with the restaurant" depending on the context. An authority structure, then, is associated with a aroun to pick a neutral term. A group has a se; of participants, connected by a social set which specifies the attitudes they have about their acquaintances in the group. Every group has a set of normal Drocedures or activ- ities, in which the participants take certain roles. For our present purposes, it doesn't matter whether those activities are highly predictable, goal-driven, or structured in any particular way. Some of those acts change the social net ("social" acts); others involve the exchange of goods and services ("economic" acts); and others are acts of authority (to be defined shortly). An individual belongs to many groups at tne same time. In fact, a pair of individuals may Delong to two groups and relate to each other in different ways (e.g., role conflict). Any group-associated act may have a legal status, an indication of its conformance with the laws. We define 6 types of legality: 1 . An act (in the past, present, or future) is explicitly legal, requiring no permission. [Example: free speech.] 2. An act is legal only . permission has been given. [YiG need a license to practice medicine.] 3. An act is legal only if it is commanded. [A six-year old child taking medicine.] 4. An act is legal only if YOU are acting as someone's agent (for whom the act may or may not be legal). LA judge authorizes a police officer to search a house, even though ne is not allowed to searcn it himse1f.j 212 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. 5. An act is legally required. LIn Autnority structures are often bridge, YOU must follow suit if emoeaaea in one anotner, you can.1 and if a group has no rules governing certain actions, then it may inherit the rules from an 6. An act is explicitly forbidden. embeaaing structure. r'or example, a [You may not submit more than one Gontract paper to this conference.] is an embedded system specifying little more than mutual obligations. The embedding system takes care of the rest. embedding is not universal, however. You Of the many possible states, some are can't sue your friends if they fail to of concern to the law, such as ownership, show up for an invited dinner. legal responsibility, and right of claim (debt or injury). eisputes are questions In general, your power within a group either about the legal status of some act is measured by your ability to "make LAm 1 required to file a tax return by April lbtn if l'm getting a refund?] or things happen," to cause social, economic, legal, or other kinds of actions. You a bout the trutn value of some law-relatea incur a debt when someone acts to increase state LThe defendant pleaded not guiltyJ. your power, and it is expected that you will reply, though not t'primitivesll necessarily in 'The acts of authority kind. A bribe, for example, is an are: exchange of economic power (money) for legal power (position of authority). 1. to define or decide the legal status of an act, according to many activities call for decisions to the six types listed above, such be made. Some of these decisions are as commanding or obliging someone based solely on evidence and are to do something [Clean up your simply evaluations, but in other cases, real room] choices must be made. A crucial part of the description of any participant in a 2. to enforce a decision [Papers group is the ability he has to influenca received after the program aecisiqns by whatever means, and we treat committee meets will be rejected tnis as a special kind of power, automatically] categor- lzed bY tne kind of decision being influenced. Politics is defined as the 3. to create and revise the rules influencing of authority-related themselves [The voters approved decisions, i.e., whether (or in what the ERA1 manner> to perform one of the acts of authority, the laws or 4. to resolve disputes [The such as revising jury admitting a new member to the acquitted the defendant] group. Attempts to influence aconomiq decisions range from friendly advice ["Try the 5. to change the position of someone cheeseburgerl'] to hard-sell in the group Lkou're hiredJ advertising techniques. Finally, you might attempt to influence a decision about a social act, such as arranging a blind date for breaking tne rules may entail someone. punisnment; non-compliance may entail enforcement; but then again, may be not. In some groups, the rules cannot be in Tudor Gngland, if a british shipping changed, in which case attempts to merchant were unable to obtain payment for influence decisions are absurd. [I'll let goods he had delivered to a Dutch YOU capture my rook if you let me ignore merchant, he could ask the Court of the the fact that I’m in check for the next Admiralty for a "Letter of Reprisal," a few m0ves.J On tne other hand, in a document that permitted him to seize any highly reticulated authority Dutch structure, ship and to share in the worth of where the rules can all be changed, the ship and its cargo. Two aspects are politics are likely. The important simplest here. First, the Court in no political acts are those that attempt to way enforced its decision; it was not the change tne rules by following the British Navy that went out to capture a appropriate procedures Lworking witnin the Dutch ship, but the merchant himself (or estaolisnmentJ. Another method is to more likely, his agents). Second, he was behave as if the rules had changed in the permitted to commit what would otherwise way you seek Idefiance], which may mean be a highly illegal act. that you've committed an illegal act. You might decide not to fill a role in a group whose rules YOU disagree with [Boycott 213 Coors and Carl's Jr.], and YOU can also exploit the embedding of inconsistent authority structures [sit-ins at lunch counters]. APPLICATIONS Perhaps the clearest applications of this work would be in a natural language understanding program, and would be visible in at least three places. First, acts would be categorized by their legal status. Some of these would be explicit LJohnny is allowed to stay up until 10 o’clockJ, but most are quantified to some degree La11 personnel actions require tne manager's approvalJ. we would nave a more accurate Ji!uzini xepresentation than we now do for examples such as 88Mommy, can 1 go to the movies?" l!or kids (i.e., in the family-group), going to the movies might be directly labeled as requiring permission, but we might also be able to infer that from knowing that it involves leaving home and spending money, both of which require permission. Second, understanding an authority structure would enable a program to make the inferences needed to connect and explain events in a story. Example: "Mary forgot to renew a book from the library. They sent her a bill." Without some understanding of the library rules, the second sentence is difficult to explain. Third, we can use authority structures to make gredictions about people's actions. If Mary orders her son Johnny to go to bed, we can make a set of reasonaole predictions about Wnat he might do ana hOW Mary will respond. If Johnny's little sister orders Johnny to go to bed, tne predictions are quite different since the authority relationsnip is obviously different. 1 f' Sue loans Tom some money, thUS increasing his economic power, we can inf'er a state of indebtedness and expect him to repay her in some way. If a student slips a $20 bill in an examination book, he probably intends to induce a state of indebtedness on the part of his professor. I;ONCLUSIONS Our goal here been to organize general informat authority struct ures, providing framework in which we can specific cultura 1 instances and d relevant infere rices.. We envis authority struct ures as a necess of the represent ation for aroubs i has to on abo ut comm on descri be fine t he on usi ng w pa rt f pew le ACKNuWLoUGhnENTs my thanks to Lamar Hill, Gene Fisher, Bob bechtel, Dave Keirsey, and Steve Heiss for their constructive (and lively) aiscussions on this topic. 11 21 RCr'tinBNChS Jaime G. Carbonell. Sub.jective Understanding: Computer Models sf Belief Svstemg, PhD dissertation, kale University, 1979. Research Report 150, Yale Computer Science Department. Roger C. Schank and Jaime G. Carbonell. He: The Gettysburg Address: Representing social and political acts. nesearcn Report 127, kale Computer science Department, January 1970. whose behavior pattern is shared. 214

1980

55

51

REAL TIME CAUSAL MONITORS FOR COMPLEX PHYSICAL SITES* Chuck Rieger and Craig Stanfill Department of Computer Science University of Maryland College Park, MD 20742 ABSTRACT Some general and specific ideas are advanced. about the design and implementation of cau;llmn;ltorlng s P stems for complex sites such as a p ant or NASA missions control room. Such '*c,P:% monitors" and are interesting from l;idth ;zEl;heoretical engineering greatly improve viewpoints, existing man-machine interfaces to complex systems. INTRODUCTION Human understandin of physical facility, a ~NASA mission control large complex sue as a room or a nuclear power plant, is often a hapk;;a:d affair. Traditionally, once a site is knoyledge about. it resides ,in technical manual; $hlch tend to sit on the-she1 ) and in the minds the experts. As technlcal staff turnover occurs over an extended period of time, this knowledge tends to get misplaced, rearranged, or altogether forgotten, - at least - as knowledge. The result is da--to-day usua P working ly that no single individhl fully appreciates or understands the system as a whole; cope with it to the operators learn simply to the extent required for daily operations and maintenance. Of course, thisa;;;;; that when an emergency occurs, or when need to perform some unusual human ex ertise maneuver, the required the she1 P is often hard to locate. Manuals on are virtually worthless in most contexts, especially those in which time is critical. Even ~JF the expert.is on the site, it may take him too per orm the synthesis of a lar e set of parameters P to perceive the context of an emergency and necessary to diagnose the prob em. 9 Given this state of produces systems too deep1 high technology, which or broadly complex for an individual or reasona ly 41 sized group of individuals to -there is a clear need for "intelligent" comprehend, secondary systems for monitoring the primary systems. To develop such intelligent systems, we need (1) flexible representations for physical causality, (2) good human interfaces that accept descriptions of the physical world andrnk;;ii $szrlptions In tkese representations, (3) ComDrehension that are capable of relating the myriad * states of the site to the causal description, then passing along only relevant and ~~por%:t information to the human operators, and efficient symbolic computation environments to support items l-3 in real time. This paper brief1 CAM (Causal Monitor) Project, ?I desc;;kes*the w ose 1s to develop a framework for the construction of intelligent, causal1 -b;s;zs real-time physica P monitors for ar~;:;~;~ The project is under funding by NASA Goddard Space Flight Center, and will result in a causal monitor generator Our aim rototype ?%kEm*about K ere is to advance some specific the conceptual architecture of such a ~~~~~~*for man-made devices can be found in [6]. Background ideas on the concept of causal GOALS An ideal causal monitoring system does not replace humans, F,ut rather 'x;;$$&~~$,b,'oad~;; their abi:;j-;;s process enhances ability to collect and synthesize relevant information in time critical situations. The ideal system would ap ear to the human controllers as a small number o I; color CRT displays keyboard. The system would and a 1. 2. 3. 4. 5. 6. 7. continually sense and all in symbolically characterize sensors reflected their relative a way ttCi importance acceptable ranges in the current operating context continually verify that causally related sensor groups obey symbolic rules ex ressin P Fi the nature of their causal re ations ip be aware of the human onerator's exoressed intent (e.g.; "runn'ing check 32", preventative maintenance "executin mornin power-up") and adiust causal rela ions an f 8 ex ectations P and- lower-level to erances accordingly parametric have a knowledge of component and sensor failure modes and limit violations, their &&e;table precursors, probable their i..iicators, corrective proceduresc8%%'in the their automatic correction algorithms forma;: recommendations to the huma:) have a knowledge of standard "maneuvers", expressed action cog?irmable sequences with ste wise f: (for bot automatic consequences execution and as an archival reference for the human) continually s nthesize K all aspects of important t e system synthesis, and from t;k,s identify which aspects of system to display to the human controllers (in the absence of specific requests from the humans) decide on the most appropriate screen allocation piece of and displayu;;c,nique for each information the current context Most procedural knowledge about the would be on-line, primar and in a form meaningfu P t:YE:etmh the hu;nTeando:he corn uter. P The s stem would play the inte ligent ifl watt er continually monitorin H hundreds of sensors and refating them to the causa model for the cur:;zt o crating context. The system would summarize s P te s state via well-managed CRT displays, and would be capable not Otll pro able 4: of detecting irregularities and suggesting causes, consequences and corrective measures, but also of automatically carrying out both routin: ;;2a;m;rge:;z c;;;z;tive measures w",;; given the physical s !? - te would ge "self aware"c~~t~~~~efimited sense. * The research described here is funded by NASA. Their support is gratefully acknowledged. 215 ARCHITECTURE OF THE CAM GENERATOR SYSTEM --- In the CAM Project we are concerned not with modelin specific * but rather with develop ng z general-purp~~~e'model P that can be imported to any site. Once there, it will interact with the site experts as causal they define the site's structure system. From the v~~si.~t~nte;~cti;;~sframe-driven acquisition generated. phase, the specific site knowledge CAM is The CAM generator system consists of several pieces: 1. 2. 3. 4. 5. 6. A collection of frame-like knowledge uni;; that collectively virtually any hysica incorporate fj Actuator, ActionStep, Component, -* FailureMode; Dis laypacket, Maneuver, Operatingcontext, wil !f OperatorIntent. Others emerge as the project progresses. A collection of procedz;s for interact;;5 with site scientists will engineers interactively describe the site by filling in frames. interface is semi-active * in a methodical way from the site engineers. compiled by the The information frame-driven interface results in a collection of and production data objects rule-like knowledge about their interrelationships. A collection of display primitives and display handling techniques ,A,iollection of *primitive sensor-reading actuator-driving schemata (i.e., code schemata) A system for corn iling R the production rule description of t e site onto an efficient lattice-like data structure that obviates most run-time pattern matching A run-time maintenance for coordinatin system causal mode 9 the real time of the symbolic monitor. The frames and the nature of the knowledge acquisition the interface are described in [3]. Since real-time efficiency of the generated system is a critical issue, we devote the remainder of discussion here to the architecture of the Ek real-time monitor s stem Propagation Driven ifi which we have termed achi;e (PDM). the PROPAGATION DRIVEN MACHINES from its runtime system. CAM will require an extre;Eky hi,p$ throughput cannot tolerate much reason, we the basic machine cyc f eneral e. pattern matchin f,i; To e iminate the nee d" attern mate hing, ependency lattice we base our machine on a form of similar in structure described in [l], [2],[4] and [5]. The e%enE?% such a scheme is to re reient antecedent-consequent relations forward and b~ac~wa~,ra~~~sohhi~~ is useful in both The central data structure of the graph whose nodes PDM is a represent the current value of some parameter or proposition (e. ., value of sensor 23 is X" "the current condition in K chamber 19"ih'rghesl~~e~~ef,p:~~s~~ network nodes are the primitive sensor readers real time clocks whose PD8 spontaneous se uences all propagation in the net. Nodes ticking are connected by Above and Below in the links which reflect computational dependencies among noies. For example, form i if node Nl represents knowledge of the A and B and C) causes D then in the PDM net E wtid E-o-1, a TZlT?lrwiil be Above each of A, , . Some event in the model operator start-up, ',,"L, ;zFputation at a no E'HA'the net) identi-ffieFhz 6 modeled , R, that deals with some aspect environment, the causal e.g., the ru~es~~;t describe relationships Pressure S stem. Rules in ti?s set are then Cht%dr on r; 3 even;t;ll; Fe ;p,;gtzd azd from which they wi 1 all evaluated. An important PDM concept is structural augmentation of a node's of the network as a natural byproduct evaluation: evaluation of a s ueued newly nto rule causes the rule to be structurally knit the PDM network. any ex ressions During a node's evaluation, B not alread as new net nodes, P* present in the net are create expression's inked to the referencing node via Above links for expansion themselves. This and placed in "downward chaining" structurally process P introduces t:e rule into the PDM net. Once rule's instz$l.Fd by. downward chaining, the value evaluation via begin to receive constant PDM "upward chaining": changes in lower nodes' values in the net Will throu h this structure, ropagate upward reeva uated. f causing hig er nodes to be R Nodes subject to reevaluation by this mechanism are queued on Q. By using the same for both structural integration of rules in;~~~~~ current context and for the values, the PDM scheme P ropagation of a lows d namic i: a graczfil and mechanism. it 0 viates context-shifting the matching. need for most run-time pittern To illustrate, suppose some node's evaluation k$s called for zfrule set for monitoring the degree 0 enness x Relief-Valve-14 (RV-14) to be swappe and in. The PDM system will, say, then evaluate others) a R? ueue up for comparing the (among . Pipe-P"4sii$?14yf rule, RV-14 again;t the R ressure of This leads to nitting process: Rl looks for RV-14 and PP-lz determinin trigger a higher node, for communicating e.g., the a display ,~reirn;~;; results to R operators. for This dual-purpy;: use of the PDM queue, staking out net structure and nam;;; performing in net node co;~~:~~ion;ak~; pf-ygztion of changes ",,y~lz,ic system. R topologically upward sprouting, AE~ .&tsp;ft;heoBett;;e g;;pagating garbage collected). The system is e.sr:,tizi$g while others are atrophying sompi ed production rule system, but one in compilation" is values. as dynamic as the propagation of Rule sets can come and go as naturally as values can be changed. Semantically, the lowest nodes of the PDM network, and those that are present initially spontaneously ticking real-time clocks whose Lb are ove pointers gob;;iEensor-reading nodes. certain mon*toring Additionally, expressions are represented by nodes initially in the net. As the site's state changes, these 'caretaker" rules (net nodes) will make reference to other rules not ;kE;,turally part of the net references come to queuing them up.':,' bg evaluated. their evalua begin ----. tion draws to participa them into the PDM net, where they te in the monitoring process. 216 When the context shifts in a way that causes the ;a;Et&;r node to be uninterested in the paged-in , the caretaker node drops its reference to iktwe?!?tsE?f"and the top members of the rule set. effectively cutting the Below link Nodes with no Above pointers and which are not marked as "permanent" are subject to PDM collection. Garbage.collection is inve:;: oEu:E origmzl;ddownward effectivef;rbX chaining that knit in set, structurally removes the entire rule set by following Below pointers. SUMMARY A CAM generator system is a s ecial of knowledge acquisition R causality in large, do:z: l;,s sites. requires (1) ~ornpl~~"~~"hys~cZ~~~ models of knowled 3 e elicitation from the site experts, models of the conce ts causa T (2)alframe-li e common to physical sites and their topology, and (3) an efficient, yet flexible a!? en ineering solution to thEfproblems of controll+ng arge symbolic knowledge in rea?y~~~~. The production rule-like PDM model appears to be a realistic approach to real-time monitoring: it exhibits the breadth and thorou hness of classical production rule system, % ut without thz usual problems of pattern matching and database maintenance. The area of data driven real-time causal models of corn lex sites appears to be a very fertile an 8 I ph sical large y unexplored area of AI research. Research in this area will he1 brin following areas of AI closer together: E %a:: rame- systems, knowledge acquisition, causal modeling, and efficient implementation techniques for real-time symbol manipulation. The domain ' manageable because we are modeling s stems in whit; deep problem solving is not a % issue, and theoretically interesting becausEY of its breadth-wise complexity. REFERENCES ill [21 E31 [41 [51 [61 London, P. E., Dependency Networks as Representation for I?~?~~~~ Solvers,. Modelling in Genera? Department of Computer University Report ?R-698, 1978. of Maryland, Technical McDermott, D., Computer Pro AIM-402, 197 . 7 Flexibility and Efficiency hinTa ram for Designing Circuits, , Rieger, C. and Stanfill, C., A Causal Monitor Generator System, Forthcoming TR, Department of Computer Science, University of Maryland, 1980. Shortliffe, E. H., MYCIN: A Rule Based Computer Program for Advising Physicians Regardin Antimicrobial Therapy Selection, Memo ifi IM-251, Artificial Intelligence Laboratory, Stanford University, 1974. Sussman, G. J. and Knight, I. F. iZ*:tal Holloway J. Computer Aided Evolutionar K Desi n for 1999. Integrated Systems, IT, AHM-526, Rieger, C., System of Representation $orACo$??~;E~~~~~ and Grinberg, M. Design, Artificial Intelligence and Pattern puteran~:d~~7D~gn, J. co 217

1980

56

52

GENERATING RELEKWC FXPLANATICNS: NATURAL LANGUAGERESPCNSES 'I0QJESTIONSAl3OUl?DATAHASE STRUCIWB* Kathleen R. McKeown Department of Ccmputer and Information Science The Moore School University of Pennsylvania Philadelphia, Pa. 19104 ABSTRACTT The research described here is aimed at unresolved problems in both natural language generation and natural language interfaces to database systems. How relevant information is selected and then organized for the generation of responses to questions about database structure is examined. Due to limited space, this paper reports on only one method of explanation, called 'kanpare and contrast". In particular, it describes a specific constraint on relevancy and organization that can be used for this response type* I INTRODUCTICN Following Thanpson [141, the process of generating natural language may be divided into two interacting phases: (1) determining the content, force, and shape of what is to be said (the "strategic aznponent") and (2) transforming that message from an internal representation into English (the "tactical ccmponent"). The decisions made in the strategic component are the focal point of the current work. These decisions are being investigated through the development of a system for answering questions about database structure that require sane type of explanation or ' description. This mrk, therefore, has two goals: (1) providing a facility that is lacking in many natural language interfaces to database systems, and (2) exercising theories about the nature of natural language generation. The system has been designed and implementation is in its beginning stages [12]. The decisions that the strategic component of a natural language generator must make are of two different types: decisions of a semantic/pragmatic nature and decisions that are structural in nature, Given a guestion, the strategic component must select only that information relevant to its answer (semantic/pragmatic decisions). What is selected must then be organized appropriately (structural decisions). These two types of decisions are the issues this work addresses. Not covered in this paper are the syntactic issues and problems of lexical choice that a tactical ccmponent must address. * This work was partially supported by NSF Grant MCS 79-08401 and an IBM Fellowship. Structural issues are important since the generation of text and not simply the generation of single sentences is being considered. Anumber of organizational principles that can be used for structuring expository text have been identified WI. These are termed ccmpare and contrast, description, top-down 7 illustration through example, definition, bottcm-up description, and t3IELLogy. In this paper, discussion is limited to compare and contrast and its effect on the 7 organization and selection processes. II THE APPLICATION Current database systems, including those enhanced by a natural language interface (e.g. [61), are, in most cases, limited in their responses to providing lists or tables of objects in the database.* Thus, allowable questions are those which place restrictions upon a class of objects occurring in the database. To ask these kinds of questions, a user must already knaw what kind of information is stored in the database and must be aware of how that information is structured. The system whose design I am describing will answer questions about the structure and organization of the database (i.e. - meta-questions)**. The classes of meta-questions which will be accepted by the system include requests for definitions, requests for descriptions of information available in the database, questions about the differences between entity-classes, and questions about relations that hold between entities. Typical of such meta-questions are the following, taken from Malhotra [7]: What kind of data do you have? What do you know about unit cost? What is the difference between material cost and production cost? What is production cost? -------------------------------------------------- * Note that in sane systems, the list (especially in cases where it consists of only one object) may be embedded in a sentence, or a table may be introduced by a sentence which has been generated by the system ( e.g. - [43). ** I am not addressing the problem of deciding whether the question is about structure or contents. 306 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. III KNWLEDGIZ BEBBESENTATICJN In order for a system to answer meta-questions, it requires information beyond that normally encoded in a database schema. The knowledge base used in this system will be based on a logical database schema similar to that described by Mays 191. It will be augmented by definitional information, specifying restrictions on class membership, and contingent information, specifying attribute-values which hold for all members of a single class. A generalization hierarchy, with mutual exclusion and exhaustion on sub-classes, will be used to provide further organization for the information. For more detail on the knowledge representation to be used, see WI. Iv SAMf?LEQTJESTIONS Textual responses to meta-questions must be organized according to sane principle in order to convey information appropriately. The compare and contrast principle is effective in answering questions that ask explicitly about the difference between entity-classes occurring in the database. (It is also effective in augmenting definitions but this would require a paper in itself.) In this paper, the following two questions will be used to illustrate how the strategic ccmponent operates: (1) What is the difference between a part-time and a full-time student? (2) What is the difference between a raven and a writing desk? V SELECTION OF RELEVAW IEFOBMATION - Questions about the difference between entities require an assumption on the part of the speaker that there is sane similarity between the items in question. This similarity must be determined before the ways in which the entities differ can be pointed out. Entities can be contrasted along several different dimensions, all of which will not necessarily be required in a single response. These include: attributes super-classes subclasses relations related entities For sane entities, a comparison along the lines of one information type is more appropriate than along others. For example, ccmparing the attributes of part-time and full-time students (as in (A) belaw) can reasonably be part of an answer to question (1) r but a comparison of the attributes of raven and writing desk yields a ludicrous answer to question (2) (see (B) below). (A) A part-time student takes 2 or 3 courses/semester while a full-time student takes 3 or 4. (B) A writing desk has 4 legs while a raven has only 2. One factor influencing the type of information to be described is the "conceptual closeness" of the entities in question. The degree of closeness is indicated by the distance between the entity-classes in the knowledge base. Three features of the knowledge base are used in determining distance: the generalization hierarchy, database relationships, and definitional attributes. A test for closeness is made first via the generalization hierarchy and if that fails, then via relationships and definitional attributes. A successful generalization hierarchy test indicates the highest degree of closeness. Usually, this will apply to questions about two sub-types of a common class, as in: What is the difference between production cost and material cost? What is the difference between a part-time and a full-time student? In the generalization hierarchy, distance is determined by two factors: (1) the path between the entity-classes in question and the nearest ccnnmon super-class; and (2) the generality of the common super-class (path between the common super-class and the root node of the hierarchy). The path is measured by considering its depth and breadth in the generalization hierarchy, as well as the reasons for the branches taken (provided by the definitional attributes). Entities are considered close in concept if path (1) is sufficiently short and path (2) sufficiently long. If the test succeeds, a discussion of the similarity in the hierarchical class structure of the entities, as well ,as a comparison of their distinguishing attributes, is appropriate. Although the entities are not as close in concept if this test fails, sane similarities may nevertheless exist between them (e.g. - consider the difference between a graduate student and a teacher). A discussion of similarities may be based on relationships both participate in (e.g. - teaching) or entities both are related to (e.g. - courses). In other cases, similarities may be based on definitional attributes which hold for both entities. For both cases, a discussion of the similarities should be augmented by a description of the difference in hierarchical class structure. Entities that satisfy none of these tests are very different in concept, and a discussion of the class structure which separates them is informative. For example, for question (2) above, indicating that ravens belong to the class of animate objects , while writing desks are inanimate results in a better answer than a discussion of their attributes. 307 VI TEXT ORGATYIZATION There are several ways in which a text can be organized to achieve the canpare and contrast orientation. One approach is to describe similarities between the entities in question, followed by differences. Alternatively, the response can be organized around the entities themselves; a discussion of the characterizing attributes of one entity may be followed by a discussion of the second. Finally, although the question may ask about the difference between entities, it may be impossible to compare them on any basis and the ccanpare and contrast must be rejected. The determination of the specific text outline is made by the structural processor of the strategic oamponent. On the basis of the input question, the structural processor selects the organizing principle to be used (for the two sample questions, compare and contrast is selected). Then, on the basz of information available in the knowledge base, the decision is reevaluated and a ccxrnnitment made to one of the outlines described above. Because of this reliance on semantic information to resolve structural problems, a high degree of interaction must exist between the structural processor and the processor which addresses semantic and pragmatic issues. One type of semantic information which the structural processor uses in selecting an outline is, again, the distance between entity-classes in the knowledge base. For entities relatively close in concept, like the part-time and the full-time student, the text is organized by first presenting similarities and then differences. By first describing similarities, the response confirms the questioner's initial assumption that the entities are similar and provides the basis for contrasting them. Two entities which are very different in concept can be described by presenting first a discussion of one, followed by a discussion of the other. Entities which cannot be described using the ccmpare and contrast organization are those which have very little or no differences. For example, if one entity is a sub-concept of another, the two are essentially identical, and the compare and contrast organizing principle must be rejected and a new one selected. VII STRATEGICPRCCESSING Although dialogue facilities between the structural processor (SIR) of the strategic ccmponent and the semantic/pragmatic processor (S&P) have not yet been implemented, the following hypothetical dialogue gives an idea of the intended result. Question (1): What is the difference between a part-time and a full-time student? STR: notes form of query and selects COMPARE AND -T S&P: queries knowledge base: DISTANCE(part-time,full-time) --> very close (same immediate super-classes) SIR: retains COMPARE AND CONTRAST selects outline: SIMIL,ARITIES DIEFERENCFS: ATTRIBUTE-TYPE1 . . ATTRIBUI'E-TYPEn CONSEQUENCES* S&P: queries knowledge base and fills in outline: SIMILARITIES super-classes(part-time,full-time) --> graduate student attribute/value(part-time,full-time) --> degree-sought = MS or PhD DIFFERENCES attribute/value(part-time,full-time) -> courses-required = part-time: 1 or 2/semester full-time: 3 or 4/sanester -> source-of-inme = part-time: full-time job full-time: unknown CONSEQUENCES none STR: further organizational tasks, not described here, include determining paragraph breaks (see [12]). Here there is 1 paragraph. The tactical canponent, with additional information frcrn the strategic component, might translate this into: Both are graduate students going for a masters or Phd. A full-time student, however, takes 3 or 4 courses per semester, while a part-time student takes only 1 or 2 in addition to holding a full-time job. After engaging in similar dialogue for question (2) p the strategic mponent might produce outline (C) belaw, which the tactical ccanponent could translate as (D): (C) RAVEN FACTS: super-classes(raven) = raven E bird E animate object WRITING DESK FACTS: super-classes(writing desk)= writing desk E furniture E inanimate object CONS~S: bird and furniture inaznpatible 2 different objects (D) A raven is a bird and birds belong to the class of animate objects. A writing desk is ------------------------------------------------- * CDNS~S here involve OdY minimal inferences that can be made about the class structure. 308 a piece of furniture and furniture belongs to the class of inanimate objects. A bird can't be a piece of furniture and a piece of furniture can't be a bird since one is animate and the other isn't. A raven and a writing desk therefore, are 2 very different things. VIII RELATEDRESEARCH INGHNFPATICN - Those working on generation have concentrated on the syntactic and lexical choice problems that are associated with the tactical component (for example, DOI, [31, D-31, D11) - Research on planning and generation ([l], 121) comes closer to the problems I am addressing although it does not address the problem of relevancy and high-level text organization. Mann and Moore [81 deal with text organization for one particular domain in their generation system, but avoid the issue of relevancy. The selection of relevant information has been discussed by Hobbs and Robinson [5] who are interested in appropriate definitions. IX CGNCLus1oNs The effects of a specific metric, the "conceptual closeness" of the items being ccmpared, were shown on the organization and selection of relevant information for meta-question response generation. Other factors which influence the response, but were not discussed here include information about the user's knowledge and the preceding discourse. Further research will attempt to identify specific constraints from these two sources which shape the response. The research described here differs frcm previous work in generation in the following ways: 1. Previous work has concentrated on the problems in the tactical ccmponent of a generator. This work focusses on the strategic cunponent: selecting and organizing relevant information for appropriate explanation. 2. While previous work has dealt, for the most part, with the generation of single sentences, here the emphasis is on the generation of multi-sentence strings. When implemented, the application for generation will provide a facility for answering questions which the user of a database system has been shown to have about the structure of the database. In the process of describing or explaining structural properties of the database, theories about the nature of text structure and generation can be tested. ACKNB I would like to thank Dr. Aravind K. Joshi, Dr. Bonnie Webber, and Eric Mays for their detailed comments on the content and style of this paper. -S [l] Appelt, D. E. "Problem Solving Applied to Language Generation" In Proc. of the 18th Annual Meeting of the ACL. 1980, pp.-59x.- Philadel$icEz [2] Cohen, P. R. "On Knowing What to Say: Planning Speech Acts," Technical Report # 118, University of Toronto, Toronto, Canada, 1978. I31 Goldman,N. "Conceptual Generation" In R. C. Schank led.1 , Conceptual Information Processing. North-Holland Publishing Co. Amsterdam, 1975. [41 Grishman, R. "Response Generation ' Question-Answering Systems" In Proc. of the 17:: Annual Meeting of the ACL. --- La Jolla, Ca., August, --- 1979, pp. 99-101. [5] Hobbs, J. R. and J. J. Robinson "Why Ask?", Technical Note 169, SRI International, Menlo Park, Ca., October 1978. [63 Kaplan, S. J. "Cooperative Responses Fran a Portable Natural Language Data Base Query System", Ph.D. Dissertation, Ccmputer and Information Science Department, University of Pennsylvania, Pa., 1979. [71 Malhotra, A. "Design Criteria for a Knowledge-based English Language system for Management: an Experimental Analysis", MAC TR-146, MIT, Cambridge, Ma., 1975. [81 Mann, w. c. and J. A. Moore "Computer as Author - Results and Prospects", ISI/RR-79-82, ISI, Marina de1 Rey, Ca., 1980. [91 Mays, E. "Correcting Misconceptions About Database Structure" In Proc. of the Conference of the CSCSI. Victoria, May 1980, pp. 123-128. BritisrCaia, Canada, I101 McDonald, D. "Steps Toward a Psycholinguistic Model of Language Production", MIT AI Lab Working Paper 193, MIT, Cambridge, Ma., April 1979. [ll] McKeown, K. R. "Paraphrasing Using Given and New Information in a Question-Answer System", In Proc. of the 17th Annual Meeting of the ACL. La e--p --- Jolla, Ca., August, 1979, pp. 67-72. 1121 McKeown, K. R. "Generating Explanations and Descriptions: Applications to Questions about Database Structure," Technical Report # MS-CIS-80-9, University of Pennsylvania, Philadelphia, Pa., 1979. [131 Simns, R. and J. Slccum "Generating English Discourse from Semantic Networks." CACM. 15:lO (1972) pp. 891-905. [14] Thompson, H. "Strategy and Tactics: A Model for Language Production" In Papers fran the 13th Regional Meeting, Chicago Linguistic S&ieFlm 309

1980

57

53

THE SEMANTIC INTERPRETATION OF NOMINAL COMPOUNDS * Timothy Wilking Finin Coordinated Science Laboratory University of Illinois Urbana IL 61801 ABSTRACT This paper briefly introduces an approach to the problem of building semantic interpretations of nominal ComDounds, i.e. sequences of two or more nouns related through modification. Examples of the kinds of nominal compounds dealt with are: "engine repairs", "aircraft flight arrival", ~aluminum water pump", and "noun noun modification". I INTRODUCTION This paper briefly introduces an approach to the problem of building semantic COmDOundS, interpretations of nominal i.e. sequences of two or more nouns related through modification. The work presented in this paper is discussed in more detail in [3] and [4]. The semantics of ncminal compounds have been studied, either directly or indirectly, by linguists and AI researchers. impressive In an early study, Lees [8 3 developed an taxoncm of the forms. II81 and Downing f 21 More recently, Levi have attempted to capture the linguistic regularities evidenced by nominal ccmpounding. Rhyne explored the problem of generating canpounds from an underlying representation in [JO]. Bra&man Cl] used the problem of interpreting and representing nominal compounds as an example domain in the development of his SI-Net representational formalism in [l]. Gershman [6] and McDonald and Hayes-Roth [9] attempt to handle noun- noun modification in the context of more general semantic systems. In this work, the interpretation of nominal compounds is divided into three intertwined subproblems: lexical internretation (mapping words into concepts), modifier narsina (discovering the structure of compounds with more that two nominalsj and concert modification (assigning an interpretation to the modification of one concept by another), this paper. This last problem is the focus of The essential feature of this form of modification is that the underlying semantic relationship which exists between the two concepts is not explicit. Moreover, a large number of relationships might, in principal, exist between the two concepts. The selection of the most appropriate one can depend, in general, on a host of semantic, pragmatic and contextual factors. been As a part of this research, a computer program has written which builds an appropriate semantic interpretation when given a string of nouns. This program has been designed as one component of the natural language question answering system JETS 151, a successor to the PLANES query system [133. The interpretation is done by a set of semantic interpretation rules. Some of the rules are very specific, capturing the meaning of idioms and canned phrases. Other rules are very general, representing fundamental case-like relationships which * Ailthor's current address: Department of Ccmputer and Information Science, Philadelphia, PA 19104. University of Pennsylvania, This work was supported by the Office of Naval Research under Contract NOOOlll-75-C-0612. can hold between concepts. A strong attempt has been made to handle as much as possible with the more general, highly productive interpretation rules. The approach has been built around a frame-based representational system (derived from FRL [ill) which represents concepts and the relationships between them. The concepts are organized into an abstraction hierarchy which supports inheritance of attributes. The same representational system is used to encode the semantic interpretation rules. An important part of the system is the concert matcher which, given two concepts, determines whether the first describes the second and, if it does, how well. II THE PROBLEM Let's restrict our attention for a moment to the simplest of ccmpounds - those made up of just two nouns, both of which unambiguously refer to objects that we know and understand. What is the fundamental problem in interpreting the modification of the second noun by the first? The problem is to find the underlying relationship that the utterer intends to hold between the two concepts that the nouns denote. For example, in the compound "aircraft engine" the relationship is part of, in "meeting room" it is location, in "salt water" it is dissolved in. There are several aspects to this problem which make it difficult. First, the relationship is not always evident in the surface form of the compound. What is it about the compound GM cars which suggests the relationship made bv? The correct interpretation of this compound depends on our knowledge of several facts. We must know that a is the name of an organization that manufactures things, and in particular, automobiles. Another fact that helps to select this interpretation is that the identity of an artifact's manufacturer is a salient fact. It is even more important when the artifact is an automobile (as opposed to, say, a pencil). A second source of difficulty is the general lack of syntactic clues to guide the interpretation process. The interpretation of clauses involves discovering and making explicit the relationships between the verb and its "arguments", e.g. the subject, direct object, tense marker, aspect, etc. Clauses have well developed systems of syntactic clues and markers to guide interpretation. These include word order (e.g. the agent is usually expressed as the subject, which canes before an active verb), prepositions which suggest case roles, and morphemic markers. None of these clues exists in the case of nominal compounds. Third, even when the constituents are unambiguous, the result of compounding them may be multiply ambiguous. For example, a woman doctor may be a doctor who is a woman or a doctor whose patients are women. Similarly, Chicano fliQhtS may be those bound for Chicago, coming from Chicago or even those making a stop in Chicago. A fourth aspect is that compounds exhibit a variable degree of lexicalization and idiomaticity. In general, 310 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. the same compound form is used for lexical items (e.g. duck soup, hanger queen) and completely productive expression (e.g. engine maintenance, faculty meeting). Finally, I point out that it is possible for any two nouns to be ccmbined as a compound and be meaningful in some context. In fact, there can be arbitrarily many possible relationships between the two nouns, each relationship appropriate for a particular context. III m INTERPRETATION RULES The implemented system contains three components, one for each of the three sub-problems mentioned in the introduction. The lexical internreter maps the incoming surface words into one or more underlying concepts. The concert modifier takes a head concept and a potential modifying concept and produces a set of possible interpretations. Each interpretation has an associated score which rates its "likelihoodm. Finally, the modifier parser applies a parsing strategy which compares and combines the local decisions made by the other two components to produce a strong interpretation for the entire compound, without evaluating all of the possible structures (the number of which increases exponentially with the number of nouns in the ccmpound). The remainder of this paper discusses some of the interpretation rules that have been developed to drive the concept modifier. Three general classes of interpretation rules have been used for the interpretation of nominal compounds. The first class contains idiomatic rules - rules in which the relationship created is totally dependent on the identity of the rule's constituents. These rules will typically match surface lexical items directly. Often, the compounds will have an idiomatic or exocentric (R) meaning. As an example, consider the Navy's term for a plane with a very poor maintenance record - a "hanger queentl. The rule to interpret this phrase has a pattern which require an exact match to the words **hanger" and l*queenl*. The second class consists of productive rules These rules attempt to capture forms of modifian which are productive in the sense of defining a general pattern which can produce many instantiations. They are characterized by the semantic relationships they create between the modifying and modified concepts. That is, the nature of the relationship is a property of the rule and not the constituent concepts. The nature of the concepts only determines whether or not the rule applies and, perhaps, how strong the resulting interpretation is. For example, a rule for dissolved in could build interpretations of such compounds as l*salt water" and (tsugar water" and be triggered by compounds matching the description: (a NominalCompound with Modifier matching (a ChemicalCompound) Modified matching (a Liquid) preferably (a Water)) The third class contains the structural rules. These rules are characterized by the structural relationships they create between the modifying and modified concepts. The semantic nature of the relationship that a structural rule creates is a function of the concepts involved in the modification. Many of these rules are particularly useful for analyzing compounds which contain ncminalized verbs. IV STRUCTURAL RULES I have found this last class to be the most interesting and important, at least from a theoretical * An exocentric compound is one in which the modifier changes the basic semantic category of the head noun, as in hot dog and ladv finger. perspective. This class contains the most general semantic interpretation rules - precisely the ones which help to achieve a degree of closure with respect to semantic coverage [5]. Similar structural rules form the basis of the approaches of Bra&man Cl] and McDonald and Hayes-Roth C91. This section presents some of the structural rules I have catalogued. Each rule handles a compound with two constituents. RULE: RoleValue + Conceot. The first structural rule that I present is the most common. It interprets the modifying-concept as specifying or filling one of the roles of the modified concept. Some examples of compounds which can be successfully interpreted by this rule are: engine repair (a to-repair with object = (an engine)) January flight (a to-fly with time = (a January)) F4 flight (a to-fly with vehicle = (an F4)) engine housing (a housing with superpart = (an engine)) iron wheel (a wheel with raw-umterial = (a iron)) Note that when the compound fits the form **subject+verb" or "object+verbll this works very nicely. The applicability of this rule is not limited to such compounds, however, as the last two examples demonstrate. To apply this rule we must be able to answer two questions. First, which of the modified concept's roles can the modifier fill? Obviously some roles of the modified concept may be inappropriate. The concept for the to-repair event has many roles, such as an agent doing the repairing, an object being repaired, an instrument, a location, a time, etc. The concept representing an engine is clearly inappropriate as the filler for the agent and time roles, probably inappropriate as a filler for the location and instrument roles, and highly appropriate as the object's filler. Secondly, given that we have found a set of roles that the modifier may fill, how do we select the best one? Moreover, is there a way to measure how well the modifier fits a role? Raving such a figure of merit allows one to rate the overall interpretation. The process of determining which roles of a concept another may fill and assigning scores to the alternatives is called role fitting. This process returns a list of the roles that the modifier can fill and, for each, a measure of how "good" the fit is. Each possibility in this list represents one possible interpretation. Not all of the possibilities are worthy of becoming interpretations, however. A selection process is applied which takes into account the number of possible interpretations, their absolute scores and their scores relative to each other. Making a role fit into an interpretation involves making a new instantiation of the modified concept, and filling the appropriate role with modifier. Details of this process are presented in the next section. RULE: Conceot + RoleValue. This rule is similar to the first, except that the concepts change places. In interpretations produced by this rule, the modified concept is seen as filling a role in the modifier concept. Note that the object referred to by the compound is still an instance of the modified concept. Some examples where this rule yields the most appropriate interpretation are: drinking water (a water which is (an object of (a to-drink))) washing machine (a machine which is (an instrument of (a to-wash))) maintenance crew (a crew which is (an agent of (a to-maintain))) Again, the application of this rule is mediated by the role fitting process. 311 RULE' Concent + RoleNominal. This rule is app1i.G when the modified concept is in the class I call role nominals, nouns that refer to roles of other underlying concepts. &glish has but one productive system for naming role ncminals: the agent of an verb can commonly be referenced by adding the -er or -or suffix to the verb stem. This should not hide the possibility of interpreting many concepts as refer@ to a role in another related concept. Some examples are: a student is the recipient of a teaching, flowing, a pipe is the conduit of a a pump is the instrument of a pumping, and a book is the obiect of a reading. This rule tries to find an interpretation in which the modifier actually modifies the underlying concept to which the role nominal refers. For example, given *IF4 Pilot'*, the rule notes that ~pilotl* is a role nominal refering to the agent role of the to-fly event and attempts to find an interpretation in which l*F4" modifies that to-fly event. The result is something like "an F4 pilot is the agent of a to-fly event in which the vehicle is an F4". Some other examples are: cat food (an object of (a to-eat with agent = (a cat))) oil pump (an instrument of (a to-pump with object = (an oil))) dog house (a location of (a to-dwell with agent = (a dog))) Viewing a concept as a role nominal (e.g. food as the object of eating) serves to tie the concept to a characteristic activity in which it participates. It is very much like a relative clause except that the characteristic or habitual nature of the relationship is emphasized. RULE: RoleNominal + Concert. This rule is very similar to the previous one exceot that it anolies when the modifying concept is a role -nominal. Thgaction is to attempt an interpretation in which the modification is done, concept not by the first concept, but by the underlying to which it refers. For example, given the compound "pilot school", we can derive the concept for "an organization that teaches people to fly". This is done by noting that pilot refers to the agent of a to-fly event and then trying to modify "school" by this "to- fly". This, in turn, can be interpreted by the Conceot + RoleNominal rule if school is defined as *'an organization which is the agent of a to-teach". This leads to an attempt to interpret to-fly modifying to- teach. The RoleValue + Conceot rule interprets to-fly as filling the object (or discipline) role of to-teach. Some other examples of ccmpounds that benefit from this interpretation rule are newspaper glasses (glasses used to read a newspaper), driver education (teaching rfTple to drive), food bowl (a bowl used to eat food out . Other Structural Rules Other structural interpretation rules that-?-have identified include SDecific+Generic which applies when the modifier is a specialization of the modified concept (e.g. F4 planes, boy child), Generic+SDecific tiich applies when the the modifier is a generalization of the modified concept (e.g. Building NE43, the integer three), Eauivalence in which the resulting concept is descendant frcm both the modifier and modified concepts (e.g. woman doctor) and Attribute Transfer in which a salient attribute of the modifier is transferred to the modified concept (e.g. iron will, crescent wrench). V ROLE FITTING The process of role fitting is one in which we are given two concepts, a RoleValue-and a Host, and attempt to find appropriate roles in the Host concept in which the RoleValue concept can be placed. Briefly, the steps carried out by the program are: [ll Collect the local and inherited roles of the Host concept; [2] Filter out any inappropriate ones (e.g. structural ones); 133 For each remaining role, compute a score for accepting the RoleValue concept; [4] Select the most appropriate role(s). In the third step, the goodness-of-fit score is represented by a signed integer. Each role of a concept is divided into an arbitrary number of facets, each one representing a different aspect of the role. In computing the goodness of fit measure, each facet contributes to the overall score via a characteristic scoring function. The facets which currently participate include the following: Requirements descriptions candidate value &match. Preferences descriptions candidate value should match DefaultValue a default value. TypicalValues other very ccmmon values for this role. Modality one of Optional, Mandatory, Dependent or Prohibited. Multiplicity maximum and minimum number of values. Salience a measure of the role's importance with respect to the concept. For example, the scoring function for the reauirements facet yields a score increment of +I for each requirement that the candidate value matches and a negative infinity for any mismatch. For the preferences facet, we get a +4 for each matching preference description and a -1 for each mismatching description. The salience facet holds a value frcm a 5 point scale (i.e. Verylow, Low, Medium, High, VeryHigh). Its scoring function maps these into the integers -1, 0, 2, 4, 8. VI SUMMARY This paper is a brief introduction to an approach to the task of building semantic interpretations of nominal compounds. A nominal compound is a sequence of two or more nouns or nominal adjectives (i.e. non-predicating) related through modification. The concepts which the nouns (and the compound) denote are expressed in a frame- based representation system. The knowledge which drives the interpretation comes from the knowledge of the concepts themselves and from three classes of interpretation rules. Examples of the most general class of interpretation rules have been given. L-11 L-21 [31 141 151 C61 [71 181 r.91 II101 [ill [I21 El31 REFERENCES Tennant, H., **i%perience with the Evaluation of Natural Language aestion Answerers**, Proc. IJCAI- $Ltpkp Japan, 1979. *IAn English language Question Answering S steh fob a Large Relational Data Base", CACM, vol. 2Y , PP* 526-539, 1978. 312

1980

58

54

TOWARDS AN AI MODEL OF ARGUMENTATIGA. Lawrence Birnbaum, Margot Flowers, and Rod McGuire Yale University Department of Computer Science New Haven, Connecticut Abstract This paper describes a process model of human argumentation, and provides examples of its operation as implemented in a computer program. Our main concerns include such issues as the rules and structures underlying argumentation, how these relate to conversational rules, how reasoning is used in arguments, and how arguing and reasoning interact. implemented, and using them the program is capable of participating in the following argument fragment, concerning the question of who was responsible for the 1967 Arab-Israeli war. The program can assume either the Israeli or the Arab point of view. [ll Arab: Who started the 1967 War? [ 21 Israeli: The Arabs did, by blockading the Straits of Tiran. Introduction [31 Arab: But Israel attacked first. Engaging in an argument is a task of considerable complexity, involving the coordination of many different abilities and knowledge sources. Some questions that arise in trying to construct a process model of argumentation include: What sub-tasks comprise argumentation? What are the argumentation ? rules underlying What representations of argument structure are necessary to support these argument rules? What are the conversational rules required for dealing with arguments as a specific type of dialogue? How is the ability to reason about the world used in the argumentation process? How do reasoning interact? and argumentation To address these questions, we are in the process of building a system, called ABDUL/ILANA, that can adopt a point of view and engage in a political argument, supporting its beliefs with argument techniques using appropriate facts and reasoning. The program takes natural language input and produces natural language output. All of the rules and mechanisms described in this paper have been This work was supported in part by the Advanced Research Projects Agency of the Department of Defense, monitored by the Office of Naval Research under contract N00014-75-C-1111, and in part by the National Science Foundation under contract IST7918463. [41 Israeli: According to international law, blockades are acts of war< [51 Arab: Were we supposed to let you import American arms through the Straits? [6l Israeli: Israel was not importing arms through the Straits. The reason for the blockade was to keep Israel from importing oil from Iran. No matter which point of view the program adopts, it uses the same principles of argumentation, abstract knowledge, and historical facts. In this way, the argument is not over the facts but instead over the interpretation of the facts, slanted by the sympathies of each arguer. Arpument tasks What tasks must an arguer perform when faced with his opponent's responses? Before anything else, each input must be transformed into a meaning representation for use in further processing (see Birnbaum and Selfridge (1980)). But much more is necessary. An arguer must understand how new input relates to the structure of the argument as a whole, and this task usually requires making non-trivial inferences. For example, consider what the program (in the role of an Israeli) must do to relate the Arab's input [31, "But Israel attacked first", to the utterances which preceded it. The Israeli must realize that [31 constitutes evidence for the unstated claim that Israel started the 1967 War, which is contrary to his claim in [21. An arguer must also relate new input to what he knows about the domain, for two reasons. First, this allows the arguer to determine whether or not 313 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. he believes that some claim put to him is true. knowledge of how to look through this network for a Second, in doing this he can uncover relevant point which can be attacked or defended. As an information which may be useful in forming a example of the use of such rules, consider how the response. program (operating as the Arab) generates a Once the understanding sub-tasks sketched out response to input 141: above have been accomplished, an arguer must decide [41 Israeli: According to international how to respond. This involves several levels of law, blockades are acts of war. decision. First, he must make large-scale strategic choices, such as whether to attack his opponent's claims, defend his own, change the This point stands in a support relation to the previous Israeli claims of utterance [21: subject, and so forth. Second, he must determine which points should be taken up, and how evidence [2al The Arabs started the war. can be provided for or against those points. A A Third, he must use his knowledge and reasoning I I ability in order to actually produce such evidence. attack This final step ends with the generation of a response in natural language (see McGuire (1980)). I support I ---[3al The Israelis ArPument structures and rules -- I started the war. A There are basically two ways that propositions in an argument can relate to each other. One point [2bl The Arab blockade led to the war. 6 I support can be evidence for another point, in which case I the argument relationship is one of support, or it I [3bl The Israeli attack can challenge the other point, in which case the support led to the war. relationship is one of attack. Thus, for example, the ultimate analysis of the Israeli's statement: I 141 Blockades are acts of war. [ 21 Israeli: The Arabs did, by blockading the Straits of Tiran. Once the program has decided to follow an attack requires not only understanding strategy, the it needs to find a weak point in two separate the opponent's argument. In this case a search propositions: rule suggests traversing up the support links in the graph, starting with the most recent [2al The Arabs started the war. input. The first item considered is 141, the Israeli's claim that blockades are acts of war. However this [2bl The Arab blockade led to the war. proposition was already checked during the understanding phase and found to be one of. the but also realizing that the second proposition [2bI program's beliefs. stands in a relation to the first Thus, it is not a good support candidate for attack. Traversing the support link proposition 112al. The argument relations are leads to [2bl, the proposition that the Arab themselves complex structures, including such blockade led to the war. However, this too was f information as which inference rules enabled the establishment of verified during understanding. Following one more the relation, (See de Kleer et link leads to (1977) and Doyle (1979) for [2al, the claim that the Arabs al. a discussion of related techniques.) started the war, which the Arab does I& believe to be true. Hence, this is a good candidate for attack. One motivation for having these argument relations is that their local "geometry" can be Now the program considers the three tactics used by argument tactics in determining how to for attacking this proposition's respond to the input. simple support These are rules that relation. describe Tactic (a>, attacking the options as the main point, to how to go about attacking or defending a proposition based on the has already been used once, as can be determined by inspecting the argument graph. Tactic argument relations in which it (b), takes part. For attacking the evidence, can't be used in this case example, one such rule coordinates the three ways to attack a simple support relationship: since the evidence has already been rejected as a candidate for attack by the argument graph search rule. (a> Attack the main point directly; This leaves tactic cc>, attacking the claim that the evidence is adequate support for the main (b) Attack the supporting evidence; point. (c) Attack the claim that the evidence Having decided to attack this support relation between proposition [2al, that the Arabs started gives support for the main point. the war, and [2bl, that the Arab blockade led to the war, The representation of the entire history of the program now attempts to do that by inferring a justification for the blockade, using the argument, as a network of propositions its abstract knowledge of blockades and its factual connected by argument relations, forms an arpument knowledge of U.S. military aid to Israel. This graph. Argument graph search rules - - embody justification is ultimately generated as: 314 151 Arab: Were we supposed to let you import American arms through the Straits? An interesting point about question [51 is that it has the form of the standard argument gambit of asking one's opponent to support or justify a position. What makes the question rhetorical is the assumption that, in this case, there is no justification for demanding that arms importation be allowed, Reasoning & memory &I arguments We have been particularly concerned with investigating how reasoning and memory search interact with the argument process. Reasoning in an argument is not simply blind inference: requirements imposed by the structure of the argument constrain when and how inferences should be made. For example, consider how the program, when adopting the Israeli point of view, responds to question [ll: El1 Arab: Who started the 1967 War? [2al Israeli: The Arabs did, [2b] Israeli: . . . by blockading the Straits of Tiran. In this case, the generation of [2al does not require the use of argument rules or episodic memory retrieval. a "gut reaction" Instead it is derived by use of rule that always assigns blame for "bad" events to some other participant. However, as soon as such a claim is put forth in a serious argument, it must be supported. The support [2bl is then produced by a more complex use of inferential memory, activated only in response to the argument goal of providing support. Conversely, reasoning and memory guide the argument process by discovering information that affects subsequent argument choices. In particular, good rebuttals may often be found in the course of understanding the input. Consider how the program (in the role of the Arab) processes utterance [21: [21 Israeli: The Arabs did, by blockading the Straits of Tiran. To understand this input, the Arab must relate it to what he knows in order to verify its truth, and perhaps more importantly, uncover his relevant knowledge. How does the program verify the claim that the Arab blockade led to the war? Access to historical events in memory depends upon their organization into temporally ordered chains of related events (see Schank (1979)). These chains are searched by a process called causal-temporal (CT) search. CT search contains the system's knowledge about relations between causality and temporal ordering. Such knowledge includes, for example, the rule that any event up to and including the first event of an episodic sequence can be a cause of that sequence, but no subsequent event can. The program checks the plausibility of 121 above by employing CT search backwards in time from the start of the war, looking for the Arab blockade. However, in the course of this search it naturally discovers the initial event of the war episode, the Israeli attack, which is then noted as . a possible cause of the war. Search continues until the Arab blockade is verified. Then, when the time comes to respond to input [21, the Israeli attack has already been found as a possible rebuttal, so that no explicit search need be initiated. Thus a prime method for finding possible responses is simply to make use of facts that have been noticed by prior inferential memory processing. This mechanism also enables the system to deny false presuppositions without the necessity for a special rule. This is because in the course of relating input which contains a false presupposition to long-term memory, the program would discover the falsehood -- and often the reason why it's false. Having this in hand, it can then be used in a rebuttal. Conclusion Engaging in an argument requires several distinct kinds of knowledge: knowledge of the domain, knowledge of how to reason, and knowledge of how to argue. The proper coordination of these disparate knowledge sources can have a mutually beneficial effect in reducing undirected processing. This kind of result would be impossible in a simplistic model based on strictly separate processing stages. References Birnbaum, L., and Selfridge, M. 1980. Conceptual Analysis of Natural Language, in R. Schank and C. Riesbeck, eds., Inside Computer Understanding, Lawrence Erlbaum Associates, Hillsdale, N.J. de Kleer, J., Doyle, J., Steele, G., and Sussman, G. 1977. AMORD: Explicit Control of Reasoning, in Proc. on Artificial Intellipence and ACM SYUIP. Programming Languages,Rochester, N.Y. Doyle, J. 1979. A Truth Maintenance System, Artificial Intelligence, vol. 12, no. 3. McGuire, R. 1980. Political Primaries and Words of Pain, unpublished ms., Yale University, Dept. of Computer Science, New Haven, CT. Schank, R. 1979. Reminding and Memory Organization: An Introduction to MOPS, Research Report no. 170, Yale University, Dept. of. Computer Science, New Haven, CT. 315

1980

59

55

A PROGRAM MODEL AND KNOWLEDGE BASE FOR COMPUTER AIDED PROGRAM SYNTHESIS' Richard J. Wood Department of Computer Science University of Maryland College Park, Maryland 20742 1 Utroductaon Program synthesis is a complex task comprising many interacting subactivities and requiring access to a variety of knowledge sourcesr Recent investigations have discovered the inadequaciest;ef current synthesis techniques to keep pace with increasing difficulties of managing large intricate problem solutions* An alternative approach to software methodologies is the development of intelligent computer systems that mana e the vast amount of information assimilated an % accessed during this process= The system's ttintelli ence" is characterized not by an innate ability to P nvent solutions, but by the incorporation of an internal model of the problem domain and corresponding program solution. This project ex called the uent- domain expert (the programmer (the cons %?%%onsu tant Bf casting this investigation in the paradigm, *the. techniques and knowleci$e general to pro rammln can Isolated 5 2 1n y$~~ipy& exam nedo coo erative frameworkanof program s nthesis the fol!?owing four major categories of ac ivities have I been identified: domain from the client, Formulation: transformation, completion, % refinement of the problem requirements into terms recognized by the system= . tructu : the selection of known techniques for solving a task's subproblems and the combination of these solution fragments to form an overall task solution. . !&& P-ludaan : language the instantiationthztf construct schemata corres ond R to ste s of the solution and w ose execu e ion will achieve the overall program behavior, A system capable of supporting computer aided synthesis must have corn onents each of the above active ies, -& corresponding to Additionally, an automated consultant must access a rich knowledge base of programming information in constructing a model of the evolving the-e F This model must %%trnGf the activities syn;t$sls and must be ;;;;;;ed and consultant the he final program, The Interactive Pro ram Synthesizer (IPS) is a system designed to fulfl 1 - f the role of a consultant *The research described in this the Office N00014-76C-0471;: Naval re ortu;;e;unded Researc R b gran 8 Their support and encouragement are gratefully acknowledged+ oomplete process descriptions, but- at a cost of additional processing for referent identificationD This renort focuses on the nature of the program model and required for a Specificall successfu!? sy%h%?ed&~~~~ the pro rammi e' the architecture of the Interacti;; Program Syn hesizer, under current development, described, iS zk~ central datamE;gycture of the IPS system developin program which represents the 2 program during synthesis. The organizat on of the program model must accommodate operations that include the introduction of new terms from the user's problem description, the refinement and further definition of existing the detection &E%i! tion of inconsistencies in the and the efficient retraction of the incons !i stent! assertions* These activities occur in client-consultant programming and correspond to the initial problem description by the client, the explanations and clarifications required by the consultant and the rejection of unfruitful partial solutions* The program model is a record of all the assertions, inferences, and deductions made during the synthesis and the justifications for each assertion* The IPS program model is encoded as a semantic network., a data structure which facilitates the processing of synthesis activitiesi The nodes of describe the cE'f'esi;;jE;ce between an ofg;ratiioz and the set requisite achievement. When the sentence "The screen is thought of as a 40 b 86 byte arrayi" is processed by the system two 0 jects are 5 introduced into the model, 1) an object whose name is "screen" and 2) an object 77 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. which is an instantiation of the two dimensional frame ZZXnce (e,g with the information presented in the the a3 the extents of*each dimension and These two objects are t p~~gdf tSeiaarray entries) 1. *DEF link* that reflects the client's decision ?o consider the abstract object "screen', as a two dimensional array* Similarly, the processing of the sentence "To clear the screen store a blank in ever e position of the screenF,, introduces two objezn; corre;;;;;;; operation ,,clear,' a f to the d;;n;E -array operation and links them via a *REF link, The *REP and *RED links are used in a similar manner represent s z stem generated decisions (i.nferenfc!$ as opposed 0 user specificationss The division of link types parallels the distinction between the two domains of expertise of the client and the consultant. Clarifvina reauests to the user are identified by expressed SDEF and *REF yi%E " tZrmirio1;~~ ob.iects (e.g "screenI, and "clear,' in the above e&&nnle) ;E;',,e tk;E;ysi;; automatically infers information array store), *RED ObJeCtS (e@g., 2-D array and two state 9 When an inimica~t~~;~;action between in the model is unravels Rieger&London, 1977 B the system the cu&ent solution and selects alternative strategies for achievin the *RED state before causing a retraction of tha f state* If the offender is a must appeal to the user for a *REF state the system restatement of the goal* The system cannot jud e the soundness of a user-supplied decom osition e an % must turn to the user for an alterna e decompositionr Other link types exist in the program model $Z$%dency) ~ZerZnce, feature-description, report. beyond the scope of this The reader is directed to other projects that investigate the foundation networks (e*g,, 1979115 [Bra&man, 19779 semantic a:df [Fahlman, The program model is instantiating constructed by schematic programming knowledge with problem specific data presented by the user. The programming knowledge base consists of facts and program construction techniques considered primi;;;; to programming and employed during 9r'; descriptions This collection includes: combination to of data types and rules for their form new (2) criteria abstract required t pes; (3) techniques by .a. type descrip ion; I problem for dz;;mposltlon. solving recognition of st$?s ,;zy interactions: and (4) methods for construction of expressions,* statements, conditionals, A fundamental input and out ut characteristic of e knowledge base is that the facts are applicable ",E many programming domains, _ _ The IPS kn;;;;zge base is organized in a hierarchical system, efficient or anization ac ivities: 2 of knowledge for iE0 s nthesis recogn;kcn and inference. % eatures presented during user's behavioral ~~crlpt=on *task sugg;;; potential programming objects represent abstract domain Identification of a particular programmin ob Jet ts. supplies information normally associated object w th !3 the object., but not stated in the user's discourse, These inferences provide a basis for queries to the user requesting additional selection information, candidatess of a particular object from a set % The programming frames contain information describing defining characteristics and potential roles of an object in a program, the sentence While processing "The screen is thought of as a 40 by 86 byte array*", for example, the prototypical t..~ dimensional array frame is retrieved *The IPS is designed to communicate with the user in English. translated Curre;;;xr the sentences into corresponding are rn:daZ! manipulation functions. This transformation will ultimately utilize a kevword narser built around a dictionary of programming terminology, instantiated with the data presented * the sentence. ",;g$ult Addit;onall , ci in;;;;;kon .a out the array frame imovides some characteristics common?;' used f indicies of the dimensions an d erminology for components of referencing the array (ergs, ,'rowl and "column for the dimensions), It also suggests queries to the user about requisite information (e&g,, is the size fixed? is this a specific individual?) type description or a kz%ted at?) ., are the values cdnt%ied selectional queries in the uses of the'airay Features describing the ~%%ti% (e-.g., the ability to define subregions of an array> are included in the frame but not processed immediately. If later assertions refer to these roles they can be retrieved from the protytpical frame and instantiated, The knowledge base contains frames for both programming objects and operations, Object frames contain defining characteristics and potential roles of an ob'ect, while o by the set of s ates requisi e i t erations are described for their correct execution and the post conditions and side-effects of the action+ This note describes the nature of two corn onents ? of a computer aided program synthesis the internal program %?w?fdge base model and the structures are of programming information. 19801 part of a larger project C%Z" directed towards the development of I theoretical modekf of program synthesis and an implementation incor orates e this Eode P. rogra%Y:ng system ""2 inves igating project simple assembly microprocessor language programming oth,a activities particular and knowledge ?sed by a co&ultant during the construction of a software package that manages a video dis lay program synthesis in t e concrete and R buffer, By examinin uncluttere % realm of-assembl P language prog ams (as contrasted to abstract high- eve1 languages f progress towards a successful corn uter aided programming system can advance in much t e same manner that advances to R general Investigat!~%??%o problem solvi resulted from the blocks-wor d domain. T Trig Thanks to Chuck Rieger, Steve Small, and Randy for reading drafts of this paper and making help ul B commentsI [Br=$mhm;a'n197~lJ Represent!in A Structural Paradigm for Inc,, Repor & 'KAdwledge, Bolt Beranek and Newman, 3605, Many 1978. [Fahlman, 19791 . Fahlman,*S.Ei ,NETL,: A n fer B ?%!bri!!&'!%a%?$@'%! -, M1T YPress' Results in Knowledge Based J?roc& a-I-79, Tokoyo, [Rieger&London, 19771 Rleger, C, & London, Pa, Subgoal Protection and Unravellin . m, f ambridge, Mass., during Plan S~;;~~;~7). pr0c.a [Wood, 19801 ;,oyf 9 R,J, t Computer Aided Pro ram Synthesis, . of Maryland, TR-861, Jan, $980. 78

1980

6

56

Knowledge Representation for Syntactic/Semantic Processing Robert J. Bobrow Bolt Beranek and Newman Inc. 50 Moulton St. Cambridge, Mass. 02238 Bonnie L. Webber Department of Computer Science The Moore School of Elelectrical Engineering D2 University of Pennsylvania Philadelphia, Pa. 19104 1. Introduction This paper discusses some theoretical implications of recasting parsing and semantic interpretation as a type of inference process which we call incrementalhescription refinement.* It draws upon our recent experience with RUS. a framework for natural language processing developed at BBN and in use in several different natural language systems across the country (for details see [II, C21 and C71>. RUS is a very practical system that is as efficient as a semantic grammar like the SOPHIE parser 161 and as flexible and extensible processor lik? LUiAR modular syntactic/semantic CIOI. It achieves this ccmbination of efficiency and flexibility by cascading Cl21 syntactic and semantic processing- producing the semantic interpretation of an input utterance incremental1 y during the parsing process, and using it to guide the operation of the parser. *The research reported in this paw was supported in part by the .Advanced Research Proj ects Agency, and was monitored by ONR under Contract No. NOOO14-77-C-0378. Because RUS provides a very clean interface between syntactic and semantic processing, it has been possible to experiment with a variety of knowledge representations in the different implementations noted above. The most recent such implementation uses the KL-ONE formalism c31, [41, to represent the knowledge needed for incremental processing. (This implementation has been dubbed PSI-KLONE, for "Earsing and Semantic InterpretatioGsmL-ONE1l.) KL-ONE is a uniform Ebject-centered representational scheme based on the idea of structured inheritance in a lattice-structured taxonomy of generic knowledge. As we shall discuss iater, - PSI-KLONE takes advantage of KL-ONE's taxonomic lattice [ill which ccmbines the properties of an inheritance network with those of-a discrimination net. The next section of this paper describes the syntactic/semantic cascade in general terms, and then gives a short example of its operation in the PSI-KLONE implementation. We then define the concent of an incremental describtion refinement (IDR)' process to use as a* paradigm for usrstanding the operation of the semantic component of-the cascade. section of the paw, requirements for a general This introduces the last which discusses the frame-like knowledge representation if it is to be capable of supporting such an IDR process. 2. The Syntactic/Semantic Cascade Within the RUS framework, the interaction between the. parser and the semantic interpreter (the interpreter) takes place incrementally as the parser scans the input string from left to right, one word at a time. The semantic interpretation of each syntactic constituent is produced in parallel with the determination of its syntactic structure. Knowledge developed in the course of producing the interpretation is used to control further action by the parser. Parsing supports the processes of semantic interpretation and discourse inference (not discussed in this paper) by finding the constituents of each phrase, determining their syntactic stS;ucture, and labelling their functional relationship to the phrase as a whole (the * We use an extended notion of functional relation here that includes surface syntactic relations, logical syntactic (or shallow case structure) relations, and relations useful for determining discourse structures such as primary focus. 316 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. matrix). These labels are proposed purely on the basisof syntactimrmation, but are intended to reflect a constituent's functional role in the matrix, apd not simply its internal syntactic structure. We will refer to these labels as functional or syntactic labels for constituents. The parser and interpreter engage in a dialogue consisting of transmissions from the parser and responses from the interpreter. A transmission is a proposal by syntax that some snecific functional relation holds between a previously parsed and interpreted constituent and the matrix phrase whose parsing and interpretation is in progress. The proposal takes the form of a matrix/label/constituent triple. The interpreter either rejects the proposal or accepts it and returns a pointer to a KL-ONE data-re which represents-its knowledge of the resulting phrase. (This pointer is not analyzed by the parser, but is rather used in the description of the matrix that syntax includes in its next proposal (transmission) to extend the matrix.) The parser is implemented as an ATN 191, and transmissions occur as actions on the arcs of the ATN grammar. The failure of an arc because of a semantic rejection of a transmission is treated exactly like the failure of an arc because of a syntactic mismatch; alternative arcs on the source state are attempted, and if none are successful, a back-up occurs. 2.1. The role of the semantic interpreter in a cascaded system The PSI-KLONE interpreter must perform two related tasks: 1. provide. feedback to the parser by checking the semantic plausibility of syntactic labels for proposed constituents of a phrase, and 2. build semantic individual phrases interpretations for The mechanism for performing both these tasks is based on the idea of mapping between the (syntactic) functional labels provided by the parser and a set of extended case-frame or semantic relations (defined by the inters) that can hold between a constituent and its matrix phrase. The mapping of functional labels to semantic relations is clearly one to many. For example, the logical subject of a clause whose main verb is "hiV1 might be the agent of the act (e.g. "'Ihe boy hit . ..I'> or the instrument (e.g. "The brick hit 11 . . . 1. A semantic relation (or semantic role), on the other hand, must completely specify the role played by the interpretation of- the constituent the interpretation of the matrix phrase. in example, a noun phrase (NP) can serve various functions in a clause, including logical subject (LSUBJ), logical object (LOBJ), surface subject (SSUBJ), and first NP (FIRSTNP). The task of the interpreter is to determine which, if any, semantic relation could hold between a matrix phrase and a parsed and interpreted constituent, given a functional label proposed by the parser. This task is accomplished with the aid of a- set of pattern-action relation mapping rules (RMRULES) that specify how a given fun- label cmapped into a semantic relation. An RMRULE has a pattern (a matrix/label/constituent triple) that specifies thext=in which it applies, in terms of: 0 the syntactic shape of the matrix (e.g. "It is a transitive clause whose main verb is 'run'."), and the interpretation and semantic role assigned to other constituents (e.8. 'IThe logical subject must be a person and be the Agent of the clause"), o the proposed functional label, and o the interpretation of the be added. constituent to The action of the RMRULE is to map the given functionallabel onto a semantic relation. A proposed syntactic label is semantically plausible if its proposal triple matches the pattern triple(s) of some RMRULE(s). KL-ONE is a good language for describing structured objects such as phrases built up out of constituents, and for representing classes of objects such as the matrix/label/constituent triples that satisfy the constraints given by RMRULE patterns. In PSI-KLONE, each RMRULE pattern is represented as a KL-ONE structure called a Generic Concept (see section 2.2). These Concep55??5- am in a taxonomy that is used as a discrimination net to determine the set of patterns which match each triple. We refer to this as the taxonomy of syntactic/semantic shapes; note that it is generally a lattice and not simply a tree structure. Associated with each semantic relation is a rule (an IRULE) that specifies how the interpretation of the constituent is to be used in building the interpretation of the matrix phrase. When all the constituents of a matrix have been assigned appropriate semantic relations, the interpretation of a phrase is built up by executing all of the IRULEs that apply to the phrase. The separation of RMRULEs from IRULEs allows PSI-KLONE to take full advantage of the properties of the syntactic/semantic cascade. As each new constituent is proposed by the parser, the interpreter uses the RMRULEs to determine which IRULEs apply to that constituent; but it does not actually apply them until the parser indicates that all constituents have been found. This buys *That is, the constituent interpretable as a person. labelled LSUBJ must be 317 efficiency by rejecting constituent label assigrments which have no hope of semantic of the PHRASE, and may have 0 or more other SYNTACTIC-CONSTITUENTS which are Modifiers. The interpretation, while deferring the construction of double arrow or SuperC Cable between PHRASE and an interpretation until the syntactic well-formedness of the entire phrase is verified. SYNTACTIC-CONSTITUi?~indicates that every instance of PHRASE is thereby a SYNTACTIC-CONSTITUENT. I 2.2. An example of the cascade The simnlified taxonomv* for our example is As a simplified example of the parser-interpreter interaction, and the use of the KL-ONE taxonomy of syntactic/semantic shapes in this interaction, we will briefly describe the process of parsing the clause "John ran the drill press." The simplified ATN grammar we use for this example is shown in Fig. 2-1. given in Fig: 2-3. This ind*icates that any CLAUSE whose Head is the verb "run" Figure 2-l: A simplified ATN For readers unfamiliar with KL-ONE, we will explain three of its major constructs as we note the information represented in the simple taxonomy shown in Fig. 2-2. In E-ONE, Generic Concepts (ovals in the diagram, boldfaceinhm represent Figure 2-2: A simple EL-ONE network description templates, from which individual descriptions or Individual Concepts (shaded ovals, also boldface in text) are formed. In Fig. 2-2, the most general description * SYNTACTIC-CONSTITUENT, which is specialized by tiz two descriptions, PHRASE and WORD. All KL-ONE descriptions are structured objects. The only structuring device of concern here is the Role. A Role (drawn as a small square, and underEd in text) represents a type of relationship between two objects, such as the relation between a whole and one of its parts. Every Role on a Generic Concept indicates what type of object can fill the Role, and how many distinct instances of the relation represented by the Role can occur. The restriction on fillers of a Role is given by a pointer to a Generic Concept, and the ntanber of possible instances of the Role is shown by a number facet (indicated in the form "M < # < NH in thegs In our diagram we indicate that every PHRASE has a WORD associated with it which fills the Head Role Figure 2-3: A simple EL-ONE Syntactic Taxonomy (independent of tense and person/nunber agreement) is an example of a RunCLAUSE. There are two classes of RunCLAUSEs represented in the taxonmy - those whose LSUBJ is a person (the Per sonRunCI.AUSEs), and those whose LSUBJ is a machine (the MachineRunCLAUSEs). The class of PersonRunCLAUSEs is again sub-divided, and its subclasses are RunMachineCLAUSE (in which the LOBJ must be a machine), RunRaceCLAUSE (in which the LOBJ is a race), and SimpleRunCLAUSE (which has no LOBJ). If we get an active sentence like "John ran the drill press", the first stage in the parsing is to PUSH for an NP from the CLAUSE network. For simplicity we assume that the result of this is to parse the noun phrase "John" and produce a pointer to NPl, an Individual Concept which is an instance of the Generic pattern PersonNP. This is the result of interaction of the parser and interpreter *To reduce clutter, several superC cables to the Concept NP have been left out. 318 at a lower level of the ATN. Since it is not yet clear what role NPl plays in the clause (i.e. because the clause may be active or passive), the parser must hold onto NPl until it has analyzed the verb. Thus the first transmission from the parser to the interpreter at this level is the proposal that rrrunV (the root of "ran") is the Head of a CLAUSE. The interpreter accepts this and returns a pointer to a new Individual Concept Ckl which it places as an instance of RunCLAUSE. Since the parser has by now determined that the clause is a simple active clause, it can now transmit the proposal that NPl is the LSUBJ of CLl. Because NPl is an instance of a PersonNP, the interpreter can tell that it satisfies the restrictions on the LSUBJ of one of the specializations of RunCLAUSE, and thus it is a semantically plausible assignment. The interpreter fills in the LSUBJ Role of CL1 with NPl, and connects CL1 to PersonRunCLAUSE, since that is the only subConcept of RunCLAUSE which can have a PersonNP as its LSUBJ. Finally, the parser PUSHes for an NP, resulting in a pointer to NP2, an instance of MachineNP. This is transmitted to the interpreter as the LOBJ of CLI. Since CL1 is a PersonRunCLAUSE, the taxonomy indicates that it can be either an instance of a RunRaceCLAUSE or a RunMBCLAUSE, or a SimpleRunCLAUSE. Since-P2 has been classifieras an instance of MachineNP, it is not compatible with being the LOBJ of a RunRaceCLAUSE (whose LOBJ must be interpretable as a race). On the other'hand NP2 is compatible with the restriction on the filler of the LOBJ Role of RunMachineCLAUSE. We assume that the taxonomy indicates all the acceptable subcategories of PersonRunCLAUSE~Thus it is only semantically plausible for NP2 to fill the LOBJ Role of CL1 if CL1 is an instance of RunMachineCLAUSE. This being the case, the interpreter can join CL1 to RunMachineCLAUSE and fill its LOBJ Role with NP2, creating a new version of CL1 whmit returns to the parser. At this point, since there are no more words in the string, the parser transmits a special message to the interpreter, indicating that there are no more constituents to be added to CLI. The interpreter responds by finding the IRULEs inherited by CL1 from RunMachineCLAUSE, PersonRunCLAUSE, etc. and using the actions on those IRULEs to create the interpretation of CLl. It associates that interpretation with CL1 and returns a pointer to CLl, now a fully parsed and interpreted clause, to the parser. *Actually, the interpreter creates a Generic subConcept of RunCLAUSE, in order to facilitate sharing of information between alternative paths in the parse, but we will ignore this detail in the remainder of the example. 3. Incremental Description We view the cascaded Refinement analysis and semantic interaction of svntactic interpretation as implementing a recognition paradigm we refer to as intiremental description refinement. In this paradigm we assume we are initially given a domain of structured objects, a space of descriptions, and rules that determine whic$ descriptions apply to each object in the domain. As an example, consider the domain to be strings of words, the structured descriptions to be the parse trees of some grammar, and say that a parse tree applies to a string of words if the leaves in the tree correspond to the sequence of words in the string. In general we assume each description is structured, not only describing the object as a whole, but having ccmponents that describe the parts of the and their relationship to the whole as well. object We consider a situation that corresponds to left-to-right parsing. A machine is presented with facts about an object or its parts in some specified order, such as learning the words in a string one by one in a left-to-right order. As it learns more properties of the object the machine must determine which descriptions are compatible with its current knowledge of the properties object and its parts. the process of: Incremental description refinement o determining the set of the (IDR) is of descriptions ccmpatible with an object known to have a given set of properties, and o refining the set of descriptions as more properties are learned. More precisely, for every set of properties P = {pl ,...,pn) of some object 0 or its parts, there is an associated set of descriptions C(P), the descriptive cover of P. The descriptive cover of P consists of- those descriptions which might possibly be applicable to 0, given that 0 has the properties PI ,-**I b; that is, the set of descriptions which apply to at least one object which has all the properties in P. As one learns more about some object, the set of descriptions consistent with that knowledge shrinks. Hence, the basic step of any IDR process is to take (1) a set of properties P, and (2) its cover C(P), and (3) some extension of P into a set P', and to produce C(P') by removing inapplicable elements from C(P). The difficulty is that it is usually impractical, if not impossible, to represent C(P) extensionally: in many cases C(P) will be infinite. (For example, until the number of words in a string is learned, the nLPnber of parse trees in C(P) remains infinite no matter how many words in the string are known.) Thus, the *We assume that at least one to each object in the domain. description applies 319 covering set must be represented intensionally, with the consequence that "removing elements" becomes an inference process which determines the intensional representation of C(P1) given the intensional representation of C(P). Note that just as any element of C(P), represented extensionally, may be structured, so may the intensional representation of C(P) be structured as well. The trick in designing an efficient and effective IDR process is to choose a synergistic inference process/intensional representation pair. One example is the use of a discrimination tree. In such a tree each terminal node represents an individual description, and each non-terminal node represents the set of descriptions corresponding to the terminals below it. Every branch indicates a test or discrimination based on some property (or properties) of the object to be described. Each newly learned property of an object allows the IDR process to take a single step down the tree, as long as the properties are learned in an order ccmpatible with the tree's structure. Each step thus reduces the set of descriptions subsumed. Another IDR process is the operation of a constraint propagation system. In such a system an object is described by a set of nodes, each of which bears a label chosen from some fixed set. The nodes are linkedinto a network, and there is a constraint relation that specifies which pairs of labels can occur on adjoining (i.e. linked) nodes. The facts learned about an object are either links between previously known nodes, or label sets which specify the possible labels at a single node. A descriptive cover is simply the cross-product of some collection of node label sets. The refinement operation consists of (I) extending the analysis to a new node, (2) removing all incompatible labels from adjacent nodes and (3) propagating the effects. Unlike the use of a discrimination net, constraint propagation does not require that information about nodes be considered in some a priori fixed order. - As mentioned earlier, in the RUS framework we are attempting to refine the semantic description of an utterance in parallel with determining its syntactic structure. The relevant properties for this IDR process include the descriptions of various constituents and their functional relations to their matrix (cf. Section 2). Unfortunately, surface variations such as passive forms and dative movement make it difficult to assume any particular order of discovery of properties as the parser considers words in a left to right order. However, the taxonomic lattice of KL-ONE can be used as a generalization of a discrimination tree which is order independent. The actual operation used in PSI-KLONE involves an extended notion of constraint propagation operating on nodes in the taxonomic lattice, and thus the resulting system has interesting analogies to both simpler forms of IDR processes. The complete algorithm for the IDR process in PSI-KLONE is too ccmplex to cover in this paper, and will be described in more detail in a forthcoming report. However, the reader is urged to return to the example in Sec. 2.2 and reconsider 320 it as an IDR process as we have described above. Briefly, we can view the KL-ONE taxonomy of syntactic/semantic shapes as a set of discrimination trees, each with a root labelled by some syntactic phrase type. At each level in a tree the branches make discriminations based on the properties of some single labelled constituent (El, such as the LSUBJ of a CLAUSE. The parser first proposes a phrase type such as CLAUSE, and the IDR process determines which tree has a root with that label. That root beccmes the current active node in the IDR process. All further refim isone within the subtree dominated by an active node. As the parser proposes and transmits new LC's to the IDR, the IDR may respond in one of two ways: 1. it may reject the LC because it is not compatiblewith any branch below the currently active node(s), or 2. it may accept the LC, and replace the current -58376 node(s) with the (set of) node(s) which can be reached by branches whose discriminations are compatible with the LC. 4. The IDR Process and Knowledge Representation We-have- identified four critical characteristics of any general representation scheme that can support an IDR process in which descriptions are structured and covering descriptions are represented intensionally. In such a scheme it must be possible to efficiently infer from the representation: 1. what properties of a structured object provide sufficient information to guarantee the applicability of a description to (some portion of) that object - i.e., criteriality conditions, 2. what mappings are possible between classes of relations - e.g. how functional relationships between syntactic constituents map onto semantic relationships 3. which pairs of descriptions are mutually incompatible - i.e., cannot both apply to a single individual 4. which sub-categorizations of descriptions are exhaustive - i.e., at least one of the sub-categories applies to anything to which the more general description applies. Wr analysis of the assumptions implicit in the current implementation of PSI-KLONE has led us to an understanding of the importance of these four points in a IDR. By making these four points explicit in the next implementation we expect to be able to deal with a larger class of phenomena than the current system handles. In the following sections we illustrate these four points in terms of the behavior of the current version of PSI-KLONE and the improvements we expect to be able to make with more information explicit of types of 4.1. Criteriality Conditions The point here is an obvious one, but bears repeating. If a taxonomy is to be used for recognition, then there must be some way, based on partial evidence, to get into it at the right place for the recognition (IDR) process to begin. That is, for any ultimately recognizable phrase there must be at least one criteria1 condition, i.e. a collection of facts which is sufficient to ensure the abnlicabilitv of some particular descris In the' syntactic"/semantic taxonomy, the criteria1 condition is often, for a phrase, the properties of belonging to a particular syntactic category (e.g., noun phrase, clause, etc.) and having a particular lexical item as head. Recalling the example given in Section 2.2, the evidence that the input had the shape of a CLAUSE and had the verb llrunV as its head constituted sufficient conditions to enter the taxonomy at the node RunCLAUSE - i.e., a specialization of CLAUSE whose head is filled by the verb "run". Without the notion of criteria1 properties, we cannot ensure the applicability of any description and therefore have no way of continuing the recognition process. 4.2. Mapping Syntactic to Semantic Relations In RB, the parser intermittently sends messages to the interpreter asking whether it is semantically plausible for a constituent to fill a specified functional role. The interpreter's ability to answer this question ccmes from its RMRULEs and their organization. This is based on the assumption that a potential constituent can fill some functional role in the matrix phrase if and only if it also fills a semantic role compatible with: o that functional role o the interpretation of that constituent o the head of that matrix phrase o the roles filled of that phrase by the other constituents o other syntactic/semantic propert that phrase and its constituents. ies of With respect to the first of these points, one effective way of representing the compatibility restrictions between syntactic and semantic relations derives from the fact that each purely syntactic relation can be viewed as an abstraction of the syntactic properties shared by some class of semantic relations (i.e., that they have syntactically identical argunents). If 1. a general frame-like system is used to represent the system's syntactic/semantic knowledge, 2. possible syntactic and semantic relations are represented therein as "slots" in a frame, and 3. there is an abstraction hierarchy among slots (the Role hierarchy in KL-ONE), as well as the more common IS-A hierarchy among frames (the SUPERC link between concepts in KL-ONE), then the interpreter can make use of this abstraction hierarchy in answering questions from the parser. As an example, consider a question from the parser, loosely translatable as "Can the PP 'on Sunday' be a PP-modifier of the NP being built whose head is 'party'?". Fig. 4-1jo Figure 4-l: A Simple NP Taxonomy We assume the NP headed by "party" has already been classified by the interpreter as a CelebrationNP. As indicated in Fig. 4-l this concept inherits from EventNP two specializations of the general PP-modifier relation applicable to NP - location-PP-modifier and time-PP-modifier. Thus "on SundaT can be one of% p-modifiers iff it can be either its location-PFmodifier or its time-PP-modifier. The next sEtion will discuss howt& decision can be made. The point here is that there must be some indication of which syntactic relations can map onto which semantic ones, and under what circumstances. An abstraction hierarchy among Roles provides one method of doing so. 4.3. Explicit compatibility/incompatibility annotations As noted above, the semantic interpreter must be able to decide if the interpretation assigned to the already parsed constituent is compatible with the type restrictions on the argLPnents of a semantic relation. For example, the PP "on Sunday" can be a PP-modifier of an NP whose Head is "party" if it isccmpatible with being either a time-PP, and hence capable of instantiating the relation time-PP-modifier, or a location-PP and hencs instaziating the relation location-PP-modifier. - izing There are two plausible strateg the somewhat informal notion of ies for formal compatibility: *In this, as in Fig. 4-1, we assume for simplicity that only these two semantic relations are consistent with the syntactic relation PP-modifier for an NP whose head is "party". - 321 1. a constituent is judged ccmpatible with a restriction if its syntactic/semantic shape (and hence interpretation) guarantees consistency with the type restrictions, or 2. it is 3-&W compatible if its interpretation does not guarantee inconsistency. - Consider the problem of rejecting "on Sunday" as a location-PP-modifier. Conceivably one could reject it on thegrounds that "Sunday" doesn't have a syntactic/semantic shape that guarantees that it is a location-NP. This is essentially the strategy followed by the current version of PSI-KLONE. More specifically, the PSI-KLONE system searches along the superC cables of a constituent to find just those semantic relations which are guaranteed to be canpatible with the interpretation of the constituent and matrix. However, "birthday that strategy would have to reject present" as being compatible with apparel-NP (thereby rejecting "Mary wore her birthday present to New York"), vehicle-NP (thereby rejecting "Mary drove her Boston to birthday present from Philadelphia"), rejecting animate-NP (thereby "Mary fed her birthday present some Little Friskies"), etc. Thus, we believe that future systems should incorporate the second strategy, at least as a fall-back when no interpretation is found using only the first strategy. This strategy also makes it easier for the system to handle pronouns and other semantically empty NPs (e.g. etc.) whose "thing" "stuff" syntactic/semantic shapes' guarantei almost nothing, but which are compatible with many semantic interpretations. The imp1 ication here for both processing language and knowledge repr mesentation is that: 1. incompatibility must be marked in the representation, and explicitly 2. the most useful strategy for determining compatibility involves not being able to show explicit inccmpatibility. One caveat and one further observ ation: this strategy is not by itself effective in certain cases of metonymy, which Webster's defines as "the use of the name of one thing for that of another associated with or suggested-by it." semantics would reject For example, "the hamburger" as the subject of,a clause like "the hamburger is getting impatient" which might occur in a conversation between a waiter and a short-order cook. However, the taxonomy would be able to provide information *If we assume something like "hamburger" being an instance of definite-food-NP, which is marked as incompatible with animate-NP, the restriction on the subject of "impatient". needed to resolve the metonymy, since it would indicate that "the hamburger" is possibly being used metonymously to refer to some discourse entity which is both describable by an animate-NP and associated with some (unique) hamburger. The observation concerns the way in which semantic interpretation was done in LUNAR [lOI, which was to judge semantic compatibility solely on the basis of positive syntactic/semantic evidence. A semantic interpretation rule could only be applied to a fragment of a parse tree if the rule's left-hand side - a syntactic/semantic template - could be matched against the fragment. The only kinds of semantic constraints actually used in the LUNAR templates were predicates on the head of some tree constituent -- e.g. that the head of the NP object of a PP constituent were of class element, etc. rock, Given this restriction, LUNAR w-t be able to handle an utterance like "give me analyses of alLaninun in NASA's gift to the Royal Academy", where clearly "gift to the Royal Academy" is not incompatible with rock. 4.4. Explicit marking of exhaustive sub-categorization in the taxonomy The algorithm we have developed for incremental description refinement requires that the IDR process be able to distinguish exhaustive from non-exhaustive sub-categorization in the taxonomy of syntactic/semantic shapes. Exhaustiveness marking plays a role similar to that played by inclusive or in a logical framework. That is, it justifies the application of case-analysis techniques to the problem of determining if a proposed constituent is ccmpatible with a given syntactic role. The interpreter is justified in rejecting a proposed label for a constituent only if it has considered all possible ways in which it can correspond to a semantic relation. Exhaustiveness marking also make it possible to infer positive information from negative information as was done in the example in section 2.2. There, the interpreter inferred that the clause was a RunMachineCLAUSE, because it was known to be a PersonRunCLAUSE and the proposed LOBJ was incompatible with it being a RunRaceCLAUSE. Such reasoning is justified only if the subcategories RunMachineCLAUSE, SimpleRunCLAUSE and RunRaceCLAUSE exhaust the possibilities under PersonRunCLAUSE. These types of inference do not always ccme up in systems that are primarily used to reason about inherited properties and defaults. For example, as long as one knows that DOG and CAT are both specializations of PET, one knows what properties they inherit from PET. It is irrelevant to an inheritance process whether there are any other kinds of PET such as MONKEY, BOA-CONSTRICTOR or TARANTULA. Many formalisms, including KL-ONE, do not require the sub-categorization of a node to be exhaustive. So there are two options vis-a-vis the way exhaustiveness can be indicated. A recognition algorithm can act as if every node were exhaustively sub-categorized - a type of Closed World Assumption L-81 - this is essentially the way 322 the current PSI-KLONE system operates. Unfortunately, there are other uses of KL-ONE in the natural language system in which concepts are subcategorized but it is clear that an exhaustive subcategorization has not been made. If the meaning of the links in the representation scheme is to be well-defined, it must be possible to distinguish exhaustive from non-exhaustive sub-categorization. The implication for both knowledge representation and inference is that some clear stand must be taken vis-a-vis the representation of exhaustive sub-categorizations. 5. Conclusion The approach we have taken in RUS is midway between completely decoupled syntactic and semantic processing and the totally merged processing that is characteristic of semantic grammars. RUS has already proven the robustness of this approach in several different systems, each using different knowledge representation techniques for the semantic ccmponent. The RUS grammar is a substantial and general grammar for English, more extensive than the grammar in the LUNAR system ClOl. Although the grammar is represented as an ATN, we have been able to greatly reduce the backtracking that normally occurs in the operation of an ATN parser, allowing RUS to approach the performance of a "deterministic" parser [21. With the aid of a "grammar compiler" [51 this makes it possible to achieve parsing times on the order of .X CPU seconds, on a DEC KLIO, for twenty word sentences. In this paper we have- focused on the latest embodiment of the RUS framework in the PSI-KLONE system -- in particular on the nature of its cascaded syntactic/semantic interactions and the incremental description refinement process they support. We believe that what we are learning about such cascaded structures and IDR processes in building PSI-KLONE is of value for the design of both natural language systems and knowledge representation systems. ACKNOWLEDGEMENTS Cur work on the PSI-KLONE system has not been done in vacua - our colleagues in this research effort includeEd Barton, Madeleine Bates, Ron Bra&-man, Phil Cohen, David Israel, Hector Levesque, Candy Sidner and Bill Woods. We hope this paper reflects well on our joint effort. The authors wish to thank Madeleine Bates, Danny Bobrow, Ron Bra&man, David Israel, Candy Sidner, Brian Smith, and Dave Waltz for their helpful comments on earlier versions of this paper. Our special thanks go to Susan Chase, whose gourmet feasts and general support made the preparation of this paper much more enjoyable than it might otherwise have been. [II PI L-31 [41 [51 Ii61 c71 [81 191 Cl01 Cl11 r121 Woods, W. A. Cascaded ATN Grammars. Amer. J. Computational Linguistics 6(l), Jan:-Mar., 1980. REFERENCES Bobrow, R. J. The RUS System. -- BBN Repom8, Bolt Beranek and Newnan Inc,, 1978. . Bobrow, R. J. & Webber, B. L. PSI-KLONE - Parsing and Semantic Interpretation in the BBN Natural Language Understanding System. In CSCSI/CSEIO Annual Conference. 7z?xrim,*1980. Brachnan, R. J. On the Epistemological Status of Semantic Networks. In Findler, Nicholas V. (editor), Associative Networks - The Representation and Use of Knowledge inomputers, . AcaGi'Gs, New York, 379. Bract-man, R. J. An Introduction to KL-ONE. In Bra&man, R.J., et al. (editors), Research in Natural Language Understanding, Annual Rep- 31 Aug. 79), pages -- -- 13;116, Bolt Berzek and Nekanan Inc, Cambridge, MA, 1980. Burton, R. & Woods, W. A. A Compiling System for Augmented Transition Networks. In COLING 76. Sixth International Conference ?%7Ziputational Linguistics, Ottawa, Canada, June, 1976. Burton, R.. Seelv Br0wn.J. Semantic Grammar: A Technique for ConstrwNa?%ral LanguazInterfaces to Instructional Svstems. BBNTeport 3587, E!ol-?&%%k And Newnan Inc., May, 1977. Mark, W. S. & Barton, G. E. The RUSGrammar Parsing System. m 3243, Genermrmarch Laboratories, 1980. Reiter, R. Closed World Data Bases. In Gallaire, H. & Minker, J. (editor), Logic and Data Bases, . -- Plenum Press, 197T Woods, W. A. Transition Network Grammars for Natural Language Analysis. CACM 13(10), October, 1970. Woods, W. A., Kaplan, R. M. & Nash-Webber. B. The Lunar Sciences Natural Language ' -- Information Systmnal Report. BBN Report 2378, Bolt Beranek andwnan Inc., June, 1972. Woods, W. A. Taxonomic Lattice Structures for Situation Recognition. In Theoretical Issues in Natural Language Processing -r- 1978. - ACL amSIGART, July, 323

1980

60

57

LANGUAGE AND MEMORY: GENERALIZATION AS A PART OF UNDERSTANDING Michael Lebowitz Department of Computer Science Yale University, P.O. Box 2158 New Haven, Connecticut 06520 ABSTRACT process. Here I will present the program's ability to remember Integrated Partial events and make generalizations about This paper presents the the stories it reads. Parser (IPP), a computer model that combines text understanding and memory of events. An extended example of the program's ability to understand MEMORY IN UNDERSTANDING newspaper stories and make generalizations that are useful for memory organization is presented. The addition of information to long-term memory is an integral part of IPP's operation. actively INTRODUCTION Furthermore, the memory update process changes the structure of memory by noticing similarities new memory Memory of specific events has not been a among events, creating structures serious part of previous Artificial Intelligence based upon generalizations about these language understanding systems. Most understanding events, and using these new structures in storing events. systems have been designed to simply parse natural language into an internal representation, use the I will representation for output tasks, and then discard In order to illustrate IPP's memory, present three stories taken This has been true even for systems primarily directly from it. concerned with creating high-level and show how IPP incorporates them into story newspapers, The stories all describe kidnappings that representations in terms of structures such as memory. scripts 1121 [31 [6l, plans and goals [6l [71, took place in Italy. frames [II and other such structures, as well as more syntactically oriented systems. (Sl) Boston Globe, 5 February 79, Italy Three gunmen kidnapped a 67 year-old retired The Integrated Partial Parser (IPP) is an industrialist yesterday outside his house near understanding program that addresses the problems this north Italian town, police said. of integrating parsing and memory updating. By making use of memory, IPP is able to achieve a high (S2) New York Times, 15 April 79, Italy A building contractor kidnapped here on Jan. level of performance when understanding texts that 17 was released last night after payment of an it has not been specially prepared for. IPP has been designed to read and remember news stories undisclosed ransom, the police said. about international terrorism taken directly from newspapers and the UPI news wire. (S3) New York Times, 25 June 79, Italy Kidnappers released an Italian shoe manufacturer here today after payment of an IPP performs six different understanding undisclosed ransom, the police said. tasks, in a single, integrated process. These tasks include the addition of new events to After a person has read these three stories he long-term memory, being reminded of previous would have undoubtedly drawn some conclusions about stories, noticing the interesting aspects of kidnapping in Italy. The similar nature of the stories, making generalizations to improve the victims - all businessmen of one sort or another - quality of its world knowledge, predicting likely In what follows I will future events, and, of course, parsing stories from is immediately apparent. show the natural language into an internal representation. actual generalizations IPP made upon reading this sequence of stories. The output that follows In this paper, I will mention IPP's parsing is from IPP's operation on the three stories above. abilities only in passing. The interested reader is refered to [41 for more details of the parsing Story: Sl (2 5 79) ITALY -------------------- (THREE GUNMEN KIDNAPPED A 67 YEAR-OLD RETIRED INDUSTRIALIST YESTERDAY OUTSIDE HIS HOUSE NEAR THIS This work was supported in part by the Advanced NORTH ITALIAN TOWN POLICE SAID) Research Projects Agency of the Department of Defense and monitored under the Office of Naval *** Parsing and incremental memory access *** Research under contract N00014-75-C-1111. 324 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. >>> Beginning final memory incorporation . . . Story analysis: EVl (S-MOP = S-EXTORT) HOSTAGES AGE OLD OCCUPATION-TYPE RETIRED ROLE BUSINESSMAN POL-POS ESTAB GENDER MALE ACTOR NUMBER 3 METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Indexing EVl as variant of S-EXTORT >>> Memory incorporation complete Stories are represented in IPP by MOPS (Memory Organization Packets) [51. MOPS were designed to represent events in memory, and are thus particularly well suited for maintaining a permanent the purpose of episodic memory. Sl is represented by S-EXTORT, a simple MOP (S-MOP) that is intended to capture the common aspects of extortion and SKIDNAP, a script that represents the low-level events of a kidnapping. IPP's long term memory is organized around S-MOPS, structures that represent generalities among scripts, and the ways in which different stories vary from the S-MOPS. The first step in adding a new story to memory is to extract the various features of each instantiated S-MOP. These features consist of the scripts that are instantiated subordinate to the S-MOP (methods and results, typically), and features from each of the role fillers (such as the actor and victim). The output above shows the features that IPP accumulated about the S-EXTORT S-MOP as it read Sl. Since there were no similar events in memory when IPP read this story, its addition to memory consisted of indexing it under S-EXTORT using simply each of the features as an index. Now consider they way IPP reads S2. The collection of the features of the S-EXTORT occurs during parsing of the story. Story: S2 (4 15 79) ITALY (A BUILDING CONTRACTOR KIDNAPPED HERE ON JAN 17 WAS RELEASED LAST NIGHT AFTER PAYMENT OF AN UNDISCLOSED RANSOM THE POLICE SAID) *** Parsing and incremental memory access *** >>> Beginning final memory incorporation . . . Story analysis: EV3 (S-MOP = S-EXTORT) RESULTS SCRIPT SS-GET-RANSOM SCRIPT SS-RELEASE-HOSTAGES HOSTAGES ROLE BUSINESSMAN POL-POS ESTAB GENDER MALE METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Creating more specific S-EXTORT (SpMl) from events EV3 EVl with features: HOSTAGES ROLE BUSINESSMAN POL-POS ESTAB GENDER MALE METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Reminded of: EVl (during spec-MOP [EVl is from Sll >>> Memory incorporation complete creation) While reading S2, IPP adds it to long-term memory. However, as it uses the indexing procedure described above, it naturally finds Sl, which is already indexed under S-EXTORT in many of the same ways as S2 would be. Thus IPP is reminded of Sl, and what is more, since there are such a large number of features in common, IPP makes a tentative generalization that these features normally occur together. Informally speaking, IPP has concluded that the targets of kidnappings in Italy are frequently businessmen. The new generalization causes the creation of a new node in memory, known as a spec-MOP, to be used to remember events that are examples of the generalization. A spec-MOP is simply a new MOP, equivalent in kind to the S-MOPS that IPP starts with, that is used to embody the generalizations that have been made. The events that went into the making of the generalization are then indexed off of the spec-MOP by any additional features they may have. Story: S3 (6 25 79) ITALY (KIDNAPPERS RELEASED AN ITALIAN SHOE MANUFACTURER HERE TODAY AFTER PAYMENT OF AN UNDISCLOSED RANSOM THE POLICE SAID) Processing: KIDNAPPERS : Interesting token - KIDNAPPERS Instantiated SKIDNAP -- S-EXTORT 325 >>> Beginning memory update . . . New features: EV5 (S-EXTORT) METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Best existing S-MOP(s) -- SpMl [the spec-MOP just created] Predicted features (from SpMl) HOSTAGES GENDER MALE POL-POS ESTAB ROLE BUSINESSMAN >>> Memory update complete . . [rest of the parsing process] >>> Beginning final memory incorporation . . . Story analysis: EV5 (S-MOP = S-EXTORT) RESULTS SCRIPT SS-GET-RANSOM HOSTAGES NATIONALITY ITALY ROLE BUSINESSMAN POL-POS ESTAB GENDER MALE RESULTS SCRIPT SS-RELEASE-HOSTAGES METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Creating more specific S-EXTORT (SpM2) than SpMl from events EV5 EV3 with features: RESULTS SCRIPT SS-GET-RANSOM SCRIPT SS-RELEASE-HOSTAGES HOSTAGES ROLE BUSINESSMAN POL-POS ESTAB GENDER MALE METHODS SCRIPT SKIDNAP NATION LOCATION ITALY Reminded of: EV3 (during spec-MOP creation) [EV3 is from S21 >>> Memory incorporation complete As it finishes reading S3, IPP completes its addition of the new event to memory. The processing is basically the same as that we saw for s2, but instead of beginning the updating process with the basic S-EXTORT S-MOP, IPP has already decided that the new story should be considered as a variant of the first spec-MOP created. IPP is able to create another new spec-MOP from S3, this one including all the features of the first spec-MOP, plus the RELEASE-HOSTAGES and GET-RANSOM results of S-EXTORT. IPP has noticed that these are frequent results of kidnappings of businessmen in Italy. So after having read these three stories, IPP has begun to create a model of kidnapping in Italy. This is the sort of behavior displayed by interested human readers. It is also the kind of behavior that cannot be captured by an understanding system that does not involve long-term memory. CONCLUSION The inclusion of long-term memory as a part of IPP has been a key factor in allowing the program to be a powerful, robust understanding system. (To date it has a vocabulary of about 3200 words and has processed over 500 stories from newspapers and the UP1 news wire, producing an accurate representation for 60 - 70%.) Furthermore, the addition of memory makes possible the inclusion of many familiar language-related phenomena, such as reminding and generalization, into a model of the understanding. As shown in the examples in this paper, memory updating can be implemented as a natural part of language processing. In fact without it, the ability of a program to fully understand a story is doubtful, since it cannot improve its knowledge of the world, and adapt its processing to fit the knowledge it has obtained - important measures of the human understanding ability. A program without episodic memory can never conclude that Italian kidnappings are often against businessmen, or any of a myriad of other generalizations that people make all the time. And that is a crucial part of understanding. [II [21 [31 [41 I51 t61 [71 REFERENCES Charniak, E. On the use of framed knowledge in language comprehension. Research Report #137, Department of Computer Science, Yale University, 1978. Cullingford, R. Script application: computer understanding of newspaper stories. Research Report #116, Department of Computer Science, Yale University, 1978 DeJong, G. F. (1979) Skimming stories in real time: An experiment in integrated understanding. Research Report #158, Department of Computer Science, Yale University. Lebowitz, M. Reading with a purpose. In Proceedings of 17th Annual Meeting; of the Association of Computational Linguistics, g Diego, CA, 1979. Schank, R. C. Reminding and memory organization: An introduction to MOPS. Research Report #l70, Department of Computer Science, Yale University, 1979. Schank, R. C and Abelson, R. P. Scripts, Plans, Goals and Understanding. Lawrence Erlbaum Associates, Hillsdale, New Jersey, 1977. Wilensky, R. Understanding Goal-Based Stories. Research Report #l40, Department of Computer Science, Yale University, 1978. 326

1980

61

58

FAILURES IN NATURAL LANGUACE SYSTEIG : APPLICATIONS ?o DATA BASE WERY SYSTEMS Eric Mays Department of Canputer and Information Science University of Pennsylvania Philadelphia, PA 19104 ABSTRACT A significant class of failures in interactions with data base query systems are attributable to misconceptions or incunplete knowledge regarding the danain of discourse on the part of the user. This paper describes several types of user failures, namely, intensional failures of presumptions. These failures are distinguished fran extensional failures of presumptions since they are dependent on the structure rather than the contents of the data base e A knowledge representation has been developed for the recognition of intensional failures that are due to the assumption of non-existent relationships between entities. Several other intensional failures which depend on more sophisticated knowledge representations are also discussed. Appropriate forms of corrective behavior are outlined which would enable the user to formulate queries directed to the solution of his/her particular task and compatible with the knowledge organization. I INTRCDUcrICN An important aspect of natural language interaction with intelligent systems is the ability to deal constructively with failure. Failures can be viewed as being of two types. One can be ascribed to a lack of syntactic, semantic, or pragmatic coverage by the system. This will be termed system failure , and manifests itself in the inability of the system to assign an interpretation to the user's input. Recent work has been done in responding to these types of failures, see for example, Weischedel and Black [aI, and Kwasny and Sondheimer E31 - A second class of failures may be termed user failures. A user failure arises when his/hermiefs about the domain of system. ** discourse diverge from those of the *This work is partially supported by a grant frcm the National Science Foundation, NSF-W 79-08401. **Sane user beliefs regarding the domain of discourse are implicitly encoded in questions posedtothe system. The beliefs held by the system are explicit in its knowledge representation, either procedurally or declaratively. To avoid confusion, a clear distinction should be made between failures and errors. An error occurs when the system's response to an input is incorrect. Errors generally manifest themselves as incorrect resolution of ambiguities in word sense or modifier placement. These errors would usually be detected by the user when presented with a paraphrase that differs in a meaningful way frcan the original input [61. More serious errors result fran incorrect coding of domain knowledge, and are often undetectable by the user. This paper concerns itself with the recognition and correction of user failures in natural language data base query systems -- in particular, failures that arise fran the user@s beliefs about the structure, rather than the cOntent, of the data base. The data base model that has been implemented for the recognition and correction of simple user failures about the data base structure is presented. Several other failures which depend on more sophisticated knowledge representation are also discussed. II PRESUPFDSITICNANDPRESUMPTICN The linguistic notion of presupposition provides a formal basis for the inference of a significant class of user beliefs. There is a less restrictive notion, presumption, which allows the inference of larger class of user beliefs, namely, that knowledge which the user must assume when posing a question. A presupposition is a proposition that is entailed by all the direct answers of a question.*** A presumption is either a presupposition or it is a proposition that is entailed by all but one of the direct answers of a question [21. Hence, presupposition is a stronger version of presumption , and a presupposition is a preswtption by definition. For example, question (la) has several direct answers such as, "John", "Sue" , etc., and, of course, "no one". Proposition (lb) is entailed by all the direct answers to (la) except the last one, i.e., "no -------m-e ***The complete definition of includes the condition that the question, direct answer pair presupposition. presupposition negation of a entails the 327 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. one". Therefore, (lb) is a presumption of (la). Proposition (Id) is a presupposition of (lc), since it is entailed by all of the question's direct answers. (la) Which faculty members teach CSEllO? (lb) Faculty members teach CSEllO. (lc) When does John take CSEllO? (ld) John takes CSEllO. Presumptions can be classified on the basis of what is asserted -- i.e., an "intensional" statement about the structure of the data base or an "extensional" statement about its contents. Thus an extensional failure of a presumption occurs based on the current contents of the data base, while an intensional failure occurs based on the structure or organization. For example, question (2a) presumes propositions (2b), (2c), and ( a) l Presumption (2b) is subject to intensional failure if the data base does not allow for the relation "teach" to hold between "faculty" and "course" entities. An extensional failure of presumption (2b) would occur if the data base did not contain any "faculty member" that "teaches" a "course". Also note that the truth of (2b) is a pre-condition for the truth of (2c) l (2a) Which faculty members teach CSEllO? (2b) Faculty members teach courses. (2~) Faculty members teach CSEllO. (Xi) CSEllO is a course. Although a presumption which fails intensionally will of necessity fail extensionally, it is important to differentiate between them, since an intensional failure that occurs will occur consistently for a given data base structure, whereas extensional failure is a transitory function of the current contents of the data base. This is not meant to imply that a data base structure is not subject to change. However, such a change usually represents a fundamental modification of the organization of the enterprise that is modelled. One can observe that structural modifications occur over long periods of time (many months to years, for example), while the data base contents are subject to change over relatively shorter periods of time (hourly, daily, or monthly, for example). Kaplan [2] has investigated the -tation and correction of extensional failures of preslrmptions. The approach taken there involves accessing the contents of the data base to determine if a presumption has a non-empty extension. The remainder of this paper discusses several ways a presumption might be subject to intensional failure. These inferences are made from the structural information of the data base. III DATA F3ASE M3DEL --- In order to recognize failures of presumptions concerning the structure of the data base, it is necessary to use a robust data model. The discussion here will assume a data base model similar to that proposed by Lee and Gerritsen [4], which incorporates the generalization dimension developed by Smith and Smith [7] into the entity-relationship model [U * Basically, entities participate in relationships along two orthogonal dimensions, aggregation bxmg dissimilar entities) and generalization (among similar entities), as well as having attributes that assume values. As an example of this type of structure consider the data base model fragment for a typical university in figure 1. Entity sets are designated by ovals, aggregation relationships by diamonds, and generalization relationships by edges from the super-entity set to the sub-entity set. The parallel arcs denote mutual exclusion. 328 Mutual exclusion is used to infer the difference between "men that are also faculty" (a possibly non-empty set) and "men that are also women" (an empty set by definition), for exmple given figure 1. This distinction can be made by prohibiting the traversal of a path in the data model that includes two entity sets which are mutually exclusive. Furthermore, the path in the generalization dimension is restricted to "upward" traversals followed by "downward" traversals. An upward (downward) traversal is from a sub-entity (super-entity) set to a super-entity (sub-entity) set. This restriction is made to prevent over-specialization of an entity set when traversing downward edges. The set of inferences that can be made in the presence of this restriction is not overly constrained, since any two entity sets that have a ccmmon intersection (sub-entity set) will also have a common union (super-entity set).* IV INTENSICNAL FAILURES A. Non-existent Relationships The most basic intensional failure that can occur is the presumption of a non-existent relationship between entity sets. In the university data base model fragment given above, such a failure occurs in the question "Which faculty take courses?". This question presumes that a "take" relationship could exist between "faculty" and "courses" entities. Since no such relationship can be established, that presumption has failed intensionally. Recognizing the failure is only part of the problem -- it is also useful to provide the user with related intensional knowledge. Given a relation R, entities X and Y, and a failed presumption (R X Y), salient intensional knowledge can be found by abstracting on R, X, or Y to create a new relation. For example, consider the following exchange: Q: Which faculty take courses?" A: "I don't believe that faculty can take courses. Faculty teach courses. Students take courses." A similar failure occurs in the presumption of a non-existent attribute of an entity set. For example, What is the cost of all courses taught by teaching assistants?", incorrectly presumes that in this data base, "cost" is an attribute of "courses". B. Inapplicable Functions Intensional failures may also occur when attempting to apply a function on a domain. The question, What is the average grade in CSEllO?", will cause no processing problems provided grades are assigned over the real numbers. But if grades ranged fron A to F, then the system should inform the user that averages can not be performed on character data. (Note that the clever system designer might trap this case and make numerical assignments to the letter grades.) A more significant aspect is the notion of a function to be meaningful over a particular dcmain. That is, certain operations, even though they might be applicable, may not be meaningful. An example would be "average social security number". The user who requested such a amputation does not really understand what the data is supposed to represent. In such cases a short explanation regarding the function of the data would be appropriate. To achieve this type of behavior, of course, the data base model must be augmented to include type and functional information. C. Higher Order Failures The mutual exclusion operator allows the detection of a failure when the question specifies a restriction of an entity set by any two of its mutually exclusive sub-entity sets. For examplep Which teachers that advise students take courses?" presumes that there could be same "teachers" that are both "faculty" and "students". Since this situation could never arise, given the structure in figure 1 , it should be cammunicated to the user as an intensional failure. If an exhaustiveness operator is incorporated as well, unnecessary restrictions of an entity set by disjunction of all of its exhaustive sub-entity sets can be detected. Although this would not constitute a failure, it does indicate that there is scane misconception regarding the structure of the data base on the part of the user. If the sub-entity sets were known to be exhaustive by the user, there would be no reason to make the restriction. As an example, the addition of the fact that "grads" and "undergrads" were exhaustive sub-entity sets of "students" would cause this misconception to arise in the question "Which students are either grads or undergrads?". The following behavior would be desired in these cases: Q: "Which teachers that advise students take courses?" A: "Faculty advise students. Students take courses. I don't believe that a teacher can be both a faculty member and a student." *See [5] for a more detailed description. 329 D. Data Currency Same failures depend on the currency of the data. One such example occurs in a naval data base about ships, subs, and aircraft. The question "What is the position of the Kitty Hawk?" presumes that timely data is maintained. Actually, positions of friendly vessels are current, while those of enemy ships might be hopelessly out of date. In this case, the failures would be extensional since the last update of the attribute must be checked for currency. It may be the case that sane data is current while other data is not. However, the update processing time lag from actual event occurence to capture in the data base might be sufficiently long that such presumptions might be subject to intensional failure. Thus the user could be made aware that current data was never available. v coNcLus1oN In this paper we have discussed several.kinds of failures of presumptions that depend on knowledge about the structure or organization of the data base. It is important to distinguish between structure and content, since there is a significant difference in the rate at which they change. When responding to intensional failures of presumptions, simply pointing out the failure is in most cases inadequate. The user must also be informed with regard to related knowledge about the structure of the data base in order to formulate queries directed at solving his/her particular problem. The technique described for recognizing intensional failures that are due to the presumption of non-existent relationships between entities and attributes of entities has been implemented. Further work is aimed at developing knowledge representations for temporal and functional information. We hope to eventually develop a general account of user failures in natural language query systems. I would like to thank Peter Buneman, Aravind Joshi, Kathy McKeown, and Bonnie Webber for their valuable camments on an earlier draft of this paper l [2] Kaplan, S.J., Cooperative Responses Fran a Portable Natural Language Data Base Query -em; Ph.D. Dision, Ccmpz and Information Science Department, University of Pennsylvania, Philadelphia, PA., 1979, [31 Kwasny, S.C.r and Sondheimer, N.K., "Ungrammaticality and Extra-Grannnaticality in Natural Language Understanding sys terns" , Proceedings of the Conference of the Association for Cqutational Linguistics, La Jolla, CA., August 1979. [4] Lee, R.M. and Gerritsen, R., "A Hybrid Representation for Database Semantics", Working Paper 78-01-01, Decision Sciences Department, University of Pennsylvania, 1978. [51 Maysr E., "Correcting Misconceptions About Data Base Structure", Proceedings of the Conference of the Canadian Society for Computational Studies of Intelligence, Victoria, British Columbia, Canada, May 1980. [63 McKeown, K., "Paraphrasing Using Given and New Information in a Question-Answer system" , Proceedings of the Conference of the Association for Canputational Linguistics, La Jolla, CA., August 1979. [7] Smith, J.M. and Smith, D.C.P., "Database Abstractions: Aggregation and Generalization", ACM Transactions on Database Systems, Vol. 2, No. 2, June 1977. 183 Weischedel, R.M., and Black, J., "Responding Intelligently to Unparseable Sentences", American Journal of Canputational Linguistics, Vol. 6, No. 2, April-June 1980. [l] Chen, P.P.S., "The Entity-Relationship Model -- Tawards a Unified View of Data", ACM Transactions on Database Systems, Vol. 1, No. 1, March 1976. 330

1980

62

59

WHEN EXPECTATION FAILS: -- Towards a self-correcting inference system Richard H. Granger, Jr. Artificial Intelligence Project Department of Information and Computer Science University of California Irvine, California 92717 ABSTRACT Contextual understanding depends on a reader's ability to correctly infer a context within which to interpret the events in a story. This "context-selection problem" has traditionally been expressed in terms of heuristics for making the correct initial selection of a story context. This paper presents a view of context selection as an ongoing process spread throuqhout the understanding process. This view requires that the understander be capable of recognizing and correcting erroneous initial context inferences. A computer program called ARTHUR is described, which selects the correct context for a story by dynamically re-evaluating its own initial inferences in light of subsequent information in a story. INTRODUCTION Consider the following simple story: [l] Geoffrey Huggins walked into the Roger Sherman movie theater. He went up to the balcony, where Willy North was waiting with a gram of cocaine. Geoff paid Willy in large bills and left quickly. Why did Geoff go into the movie theater? Most people infer that he did so in order to buy some coke, since that was the outcome of the story. The alternative possibility, that Geoff went to the theater to see a movie and then coincidentally ran into Willy and decided to buy some aoke from him, seems to go virtually unnoticed by most readers in informal experiments. On the basis of pure logic, either of these inferences is equally plausible. However, people overwhelmingly choose the first inference to explain this story, maintaining that Geoff did not go into the theater to see a movie. The problem is that the most plausible initial inference from the story's first sentence is that Geoff did go inside to see a movie. Hence, selection of the correct inference about Geoff's goal in this story requires rejection of this initial inference. This paper describes a program called ARTHUR (A Reader THat understands Reflectively) which understands stories like [l] by generating tentative initial context inferences and then re-evaluating its own inferences in light of subsequent information in the story. By this process ARTHUR understands misleading and surprising stories, and expresses its surprise in English. For example, from the above story, ARTHUR answers the following question about Geoff's intentions: Q) Why did Geoff go into the movie theater? A) AT FIRST I THOUGHT IT WAS BECAUSE HE WANTED TO WATCH A MOVIE, BUT ACTUALLY IT'S BECAUSE HE l.&NTED To BUY COCAINE. (For a much more complete description of ARTHUR, see Granger [19801). We call the problem of finding the correct inference in a story the "context-selection problem" (after the "script-selection problem" in Cullingford El9781 and Dejong [1979], which is a special case (see Section 4.2)). All the "contexts" (or "context inferences") referred to in this paper are goals, plans or scripts, as presented by Schank and Abelson [1977]. Other theories of contextual understanding (Charniak C19721, Schank [19731, Wilks [1975], Schank and Abelson [1977], Cullingford [19781, [1978]) involve Wilensky the selection of a context which is then used to interpret subsequent events in the story, but these theories fail to understand stories such as [l], in which the initially selected context turns out to be incorrect. ARThllR operates by maintaining an "inference-fate g rayh", containing the tentative inferences generated during story processing, alony with information about the current status of each inference. BACKGROUND: SCRIPTS, PLANS, GOALS & UNDERSTANDING 2.1 Contextual understanding ARTHUR's representational scheme is adopted from Schank and Abelson's [1977] framework for representing human intentions (goals) and methods of achieving those goals (plans and scripts). The problem ARTHUR addresses is the process by which a given story representation is generated from a story. It will be seen that this process of mapping a story onto a representation is not straightforward, and may involve the generation of a number of intermediate representations which are discarded by the time the final story representation is complete. Recall the first sentence of story r11: "Geoffrey Huggins walked into the Roger Sherman movie theater." ARlHUR's attempt to infer a context for this event is based on knowledge of typical functions associated with objects and locations. In this instance, a movie theater is a known location with an associated activity: "scripty" viewing a movie. Hence, whenever a story character goes to such a location, one of the plausible inferences from this action is that the character may intend to perform this activity. Seeing a movie also has a default goal associated with it: being entertained. Thus, ARTHUR infers that Geoff plans to see a movie to entertain himself. 301 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. When the next sentence is read, "He went up to the balcony, where Willy North was waiting with a gram of cocaine", ARTHUR again performs this bottom-up inference process, resulting in the inference that Geoff may have been planning to take part in a drug sale. Now ARTHUR attempts to connect this inference with the previously inferred goal of watching a movie for entertainment. Now, however, ARTHUR fails to find a connection between the goal of wanting to see a movie and the action of meeting a uocaine dealer. Understanding the story requires ARTHUR to resolve this connection failure. 2.2 Correcting an erroneous inference - Having failed to specify a connecting inferential path between the initial goal and the new action, ARTHUR generates an alternative goal inference from the action. In this case, the new inference is that Geoff wanted to entertain himself by intoxicating himself with cocaine. (Note that this inference too is only a tentative inference, and could be supplanted if it failed to account for the other events in the story.) ARTHUR now has a disconnected representation for the story so far: it has generated two separate goal inferences to explain Geoff's two actions. ARTHUR thinks that Geoff went to the theater in order to see a movie, but that he then met up with Willy in order to buy some coke. This is not an adequate representation for the story at this point. The correct representation would indicate that Geoff performed both of his actions in service of a single goal of getting coke, and that he never intended to see a movie there at all; the theater was just a meeting place. Hence, ARTHUR instead infers that Geoff's action of going into the theater was in service of the newly inferred goal, and discards the initial inference (wanting to see a movie) which previously explained this action. We call this process supplanting an inference: ARTHUR supplants its initial "see-movie" inference by the new "yet-coke" inference, as the explanation for Geoff's two actions. ARTHUR's representation of the story now consists of a single inference about Geoff's intentions (he wanted to acquire some coke) and two plans performed in service of that goal (getting to the movie theater and getting to Willy), each of which was carried out by a physical action (PTRANSing to the theater and PTRANSing to Willy). At this point, the initial goal inference (that Geoff wanted to see a movie) has been supplanted: it is no longer considered to be a valid inference about Geoff's intentions in light of the events in the story. 3.1 OPERATICN OF THE ARTHUR PROGRAM Annotated run-time output The following represents actual annotated run-time output of the ARTHUR program. The input to the program is the following deceptively simple story: [2] Mary picked up a magazine. She swatted a fly. The first sentence causes ARTHUR to generate the plausible inference that Mary plans to read the magazine for entertainment, since that is stored in ARTHUR's memory as the default use for a magazine. ARTHUR's internal representation of this situation consists of an "explanation triple": a goal (being entertained), an event (picking up the magazine), and an inferential path connecting the event and goal (reading the magazine). The following ARTHUR output is generated from the processing of the second sentence. (ARTHUR'S output has been shortened and simplified here for pedagogical and financial reasons.) :CURRENT EXPLANATION-GRAPH: GOAL: (E-ENTERTAIN (PLANNER MARY) (OBJECT MAG)) EW: (GRASP (ACTOR MARY) (OE+JECT M&)) PATHO: (READ (PLANNER MARY) (OBJECT MAG)) ARTHUR's explanation of the first sentence'has a goal (being ENTERTAINed), an act (GRASPiq a magazine) and an inferential path connecting the action and goal (READing the magazine). :NEXTSENTENCE CD: (PROPEL (ACTOR MARY) (OBJECT NIL) (TO FLY)) The Conceptual Dependency for Mary's acti she struck a fly with an unknown object. On: :FAILURE TO CONNECT TO EXISTING GOAL CONTEXT: ARTHUR's initial goal inference (Mary planned to entertain herself by reading the magazine) fails to explain her action of swatting a fly. :SUPPLANTING WITH NEW PLAUSIBLE GOAL CONTEXT: (PHYS-STATE (PLANNER MARY) (OEUECT MAG) (VAL -10)) ARTHUR now generates an alternative goal on the basis of Mary's new action: she may want to destroy the fly, i.e., want its physical state to be -10. This new goal also serves to explain her previous action (getting a magazine) as a precondition to the action of swatting the fly, once AKI'HUR infers that the magazine was the INSTRument in Mary's plan to damage the fly. :FINAL EXPLANATION-TRIPLE: GOALl: (PHYS-STATE (PLANNER MARY) (OBJECT FLY) (VAL -10)) EVl: (GRASP (ACtOR MARY) (OBJECT MAG)) PATHl: (DELTA-CONTROL (PLANNER MARY) (OBJECT Mpc,) EV2: (PROPEL (ACTOR MARY) (OBJECT MAG) (TO FLY)) PAm2: (CHANGE-~Ys-STATE (PLANNER MARY) (ow~cr FLY) (DIRECTION NM;) (INSTR MAG)) 302 This representation says that Mary wanted to destroy a fly (GOALl), so she planned to damage it (PATH2). Her first step in doing so was to get an instrumental object (PATHl). These two plans were realized (Events 1,2) by her picking up a magazine and hitting the fly with it. :READY FOR QUESTIONS: >Why did Mary pick up a magazine? AT FIRST I THOUGHT IT c91S BECAUSE SHE WANTED To READ IT, BUT ACTUALLY IT'S BECAUSE SHE WTED TO USE IT TO GET RID OF A FLY. The question asks for the inferred goal under- lying Mary's action of GRASPing the magazine. This answer is generated according to ARTHUR's suw$aed inference about the action (READ) and the active inference about the action (CHANGE-PhYS-STATE) . The English generation mechanism used is described in Granger [1980]. 3.2 The parsimony principle AKIWUR's answer as given above is not the only possible interpretation of the story; it is only one of the following three alternatives, all of which are valid on the basis of what the story says: (2a) Mary picked up a magazine to read it. She then was annoyed by a fly, and she swatted it with the magazine she was holding. (2b) Mary picked up a magazine to read it. she then was annoyed by a fly, and she swatted it with a flyswatter that was handy. (2~) Mary picked up a magazine to swat a fly with it. The last interpretation (2~) reflects a story representation which consists of a single goal, getting rid of a fly, which both of Mary's actions were performed in service of. The other interpretations both consist of two separate goals, each of which explains one of Mary's actions. In [21, as in [U, the interpretation generated by the reader is the most parsimonious of the possible interpretations. That is, the preferred interpretation is the one in which the fewest number of inferred goals of a story character account for the maximum number of his actions. We ammarize this observation in the, following principle: The Parsimony Principle The best context inference is the one which accents for the most actions of a story character. - 303 In other words, the decisicn to replace a previous inference by a new one is not based on the explicit contradiction of that inference by subsequent information in the story. Example [2], for instance, has three possible interpretations, none of which can be ruled out on strictly logical grounds. Rather, the reader prefers the most parsimnious story representation over less parsimonious ones, that is, the representation which includes the fewest goal inferences to account for the actions in the story. This is true even when it requires the reader to do the extra work of replacing one of its own previous inferences, as in example [2]. CATEGORIES OF ERRONEOUS INEERENCES 4.1 Goals ARTHUR is capable of recognizing and correcting erroneous context inferences in order to maintain a parsimonious explanation of a story. The examples given so far have dealt only with erroneous goal inferences, but other conceptual categories of inferences can be generated erroneously as well. In this section, examples of other classes of erroneous inferences will be given, and it will be shown why each different class presents its own unique difficulties to ARTHUR's correction processes. 4.2 Plans and scripts -- Consider the following simple story: [3] Carl was bored. He picked up the newspaper. He reached under it to get the tennis racket that the newspaper had been covering. This is an example in which ARTHUR correctly infers the goal of the story character, but erroneously infers the plan that he is going to perform in service of his goal. ARTHUR first infers that Carl will read the newspaper to alleviate his boredom, but this inference fails to explain why Carl then gets his tennis racket. ARTHUR at this point attempts to supplant the initial goal inference, but in this case ARTHUR knows that that goal was correctly inferred, because it was implicitly stated in the first sentence of the story (that Carl was bored). Hence ARTHUR infers instead that it erroneously inferred the plan by which Carl intended to satisfy his goal (reading the newspaper). Rather, Carl planned to alleviate his boredom by playing tennis. The problem now is to connect Carl's action of picking up the newspaper with his plan of playing tennis. Instead of using the newspaper as a functional object (in this case, reading material), Carl has treated it as an instrumental object that must be moved as a precondition to the implementation of his intended plan. (Preconditions are discussed in Schank and Abelson [1977]) e ARTHUR recognizes that an object can be used either functionally or instrumentally. Furthermore, when an action is performed as a precondition to a plan, typically the objects used in the action are used instrumentally, as in [31 . ARTHUR's initial inference about Carl's plan was based on the functionality of a newspaper. It is able to supplant this inference by an inference that Carl instead used the newspaper instrumentally, as a precondition to getting to his tennis racket, which in turn was a presumed precondition to using the racket to play tennis with. Hence, correcting this erroneous plan inference required ARTHUR to re-evaluate its inference about the intended use of a functional object. 4.3 Causal state changes -- Consider the following example: [4] Kathy and Chris were playing golf. Kathy hit a shot deep into the rough. We assume that Kathy did not intend to hit her ball into the rough, since she's playing golf, which implies that she probably has a goal of winning or at least playing well. Her action, therefore, is probably not goal-oriented behavior, but is accidental: that is, it is an action which causally results in some state which may have an effect on her goal. This situation differs from stories like [l], 121 and [31, in that ARTHUR does not change its mind about its inference of Kathy's goal. Rather than assuming that .the goal inference was erroneous, ARTHUR infers that the causal result state hinders the achievement of Kathy’s goal. Any causal state which affects a character's goal, either positively or negatively, appears in ARTHUR's story representation in one of the following four relationships to an existing goal: l- the state helps the achievement of the goal; 2 - the state hinders achievement of the goal; 3 - the state aZG%Z the goal entirely; or 4- the state thwarts the goal entirely. If ARTHUR did assume that Kathy's shot was intentional, then the concomitant inference is that she didn't really want to win the game at all; or, in other words, that the initial inference was erroneous. This is the case in the following example: [S] Kathy and Chris were playing golf. Kathy hit a shot deep into the rough. She wanted to let her good friend Chris win the game. Understanding this story requires ARTHUR first to infer that Kathy intends to win the game; then to notice that her action has hindered her goal, and finally to recognize that the initial goal inference was erroneous, and to supplant it by the inference that Kathy actually intended to lose the game, not win it. 4.4 Travelling down the garden path . . . -- If the correct context inference for a story remains unknown until some significant fraction of the story has been read, the story can be thought of as a "garden path" story. This term is borrowed from so-called garden path sentences, in which the correct representation of the sentence is not resolved until relatively late in the sentence. We will call a garden path story any story which causes the reader to generte an initial inference which turns out to be erroneous on the basis of subsequent story events. Obvious examples of garden path stories are those in which we experience a surprise ending, e.g., mystery stories, jokes, fables. Since ARTHUR operates by generating tentative initial inferences and then re-evaluating those inferences in light of subsequent information, ARTHUR understands simple garden path stories. Not all garden path stories cause us to experience surprise. For example , many readers of story [2] do not notice that Mary might have been planning to read the magazine, unless that intermediate inference is pinted out to them. Hence we hypothesize that the processes involved in understanding stories with surprise endings must differ from the processes of understanding other garden path stories. Hence, ARTHUR's understanding mechanism is not entirely psychologically plausible in that it does not differentiate between stories with surprise endings and other garden path stories. A more sophisticated version of ARTHUR (call it “Macro-MTHUR”) might differentiate between "strong" default inferences and "weak" tentative inferences when generating an initial context inference. If a strong initial inference is generated, then MacARTHUR would consciously "notice" this inference being supplanted, thereby experiencing surprise that the inference was incorrect. Conversely, if the initial inference is weak, MacAKTHUR may not commit itself to that inference, but rather may choose to keep around other possible alternatives. In this case MacARTHUR would onlv exmrience further specification of the initial- tentative set of inferences, rather than supplanting a sinqle strong inference. The question of-when readers processes consciously versus unconsciously is still an open question in psychology. Future psychological studies of the cognitive phenomena underlying human story understanding (such as in Thorndyke [19761, [1977], Norman and Bobrow [19751 , and Hayes-Roth [1977], to name a few) may be able to provide data which will shed further light on this issue. 304 a)NCLUSIONS: WHERE WE'VE BEEN/ WHERE WE'RE HEADING REFERENCES 5.1 Process and representation in understanding This paper has presented a process for [ll Cullingford, R. (1978). Application: Script Computer Understanding of Newspaper Stories. Ph.D. Thesis, Yale University, New Haven, building story representations which contain inferences not explicitly stated in the story. The representations themselves are not new; they are based on those presented by Schank and Abelson [19771. What is new here is the process of corm m [2l Granger, R. (19813). Adaptive Understanding: Correcting Erroneous Inferences. Ph.D. Thesis, Yale University, New Haven, Conn. arriving at a story representation. Most contextual understanders (e.g. Charniak E19751, Cullingford [1978], Wilensky [1978]) would fail to arrive at the correct story representations for any of the examples in this paper, because initial statements in the examples trigger inferences which prove to be erroneous in light of subsequent story statements. ARTHUR's processing of these examples shows that arriving at a given story representation may require the reader to generate a nunber of intermediate inferences which get discarded along the way, and which therefore play no role in the final representation of the story. [3] Hayes-Roth, B. (1977). Implications of human pattern processing for the desiyn of artificial knowledge systems. In Pattern-directed inference systems (Waterman and Hayes-Roth, eds.). Academic Press, N.Y. [4] Kintsch, h. (1977) e Memory @ Cognition. Wiley, New York. [S] Norman, D. and Bobrow, D (1975). On the role of active memory processes in perception and cognition. In The structure of human memory -- (Coffer, C., ed.). Freeman, San Francisco. Thus a final story representation may not completely wecify the process by which it was generated, since there may have been intermediate inferences which are not contained in the final [6] Schank, R. C. and Abelson, R. P. (1977). Scripts, Plans, Goals and Understanding. Erlbaum Press, Hill=, E. representation. Yet we know that when people have understood one of these examples, they can express these intermediate inferences with phrases like "At first I thought X, but actually it's Y." ARTHUR keeps track of its intermediate inferences while understanding a story, and maintains an "inference-fate graph" containing all inferences generated during story processing, whether they end up in the final story representation or not. [71 Thorndyke, P. (1977). Pattern-directed processing of knowledge from texts. In Pattern-directed inference systems (Waterman and Hayes-Roth, eds.). Academic Press, N.Y. [8] Wilensky, R. (1978). Understanding Goal-based Stories. Ph.D. Thesis, Yale University, New Haven, Conn. The point here is that the relationship between a given story representation on the one hand, and the process of arriving at that representation on the other, may be far from straightforward. The path to a final story representation may involve sidetracks and spurious inferences which must be recognized and corrected. Therefore, specifying the representations corresponding to natural language inputs is not enough for a theory of natural language processing; such a theory must also include descriptions of the processes by which a final representation is constructed. ARTHUR has demonstrated one area in which specification of process and representation diverge: the area of correcting erroneous inferences during understanding. Further work will be directed towards specifying other conditions under which process and representation are not straightforwardly related in natural language tasks. [9] Wilks, Y. (1975). Seven Theses on Artificial Intelligence and Natural Language, Research Report No. 17, Instituto per gli Studi Semantici e Cognitivi, Castagnola, Switzerland. 305

1980

63

60

ORGANIZING MEMORY AND KEEPING IT ORGANIZED Janet L. Kolodner Dept. of Computer Science Yale University, P. 0. BOX 2158 New Haven, CT 06520 ABSTRACT Maintaining good memory organization is important in large memory systems. This paper presents a scheme for automatically reorganizing event information in memory. The processes are implemented in a computer program called CYRUS. INTRODUCTION People are quite good at retrieving episodes from their long term memories. In fact, they are much better at information retrieval than any current computer system. Psychologists have described human memory as a reconstructive process (e.g., 131). When people attempt to remember events and episodes from their lives, they often must go through a complicated reasoning and search process ([6l and [51). These processes are dependent on good memory Drganization. In order to keep memory well organized as new data is added, memory organization must support the creation of new memory categories and the building up of generalized knowledge. If a memory held only 10 events that could be described as meetings, a "meetings" category would be useful. But, unless new meeting sub-categories were created as additional meetings were added to the memory, retrieval of meetings would become very inefficient. Thus, a memory system needs the ability to create new categories automatically from old ones. CYRUS is a computer program which implements a theory of human memory organization and retrieval. The program is designed to store and retrieve episodic information about important people, and is based on a theory of the way people organize and remember information about themselves. Right now, it,holds information about Secretaries of State Cyrus Vance and Edmund Muskie. CYRUS answers questions posed to it in English, using search strategies [51 to search its memory and a set of constructive strategies 121 to construct search keys. CYRUS is connected to the FRUMP program [ll, which produces summaries of stories off the UP1 *This work was supported in part by the Advanced Research Projects Agency of the Department of Defense and monitored under the Office of Naval Research under contract NOOOl4-75-C-1111. wire. When FRUMP sends CYRUS new information about Vance or Muskie, CYRUS automatically updates its memory. As it does that updating, it reorganizes its memory and builds up generalized knowledge about the information it stores. CYRUS' memory is organized around episodes using Memory Organization Packets (MOPS) [41. Because episodes include references to persons who participated in them, their locations, other episodes they are related to, etc., they are good global organizers of otherwise disjoint information. For example, Camp David and Menachim Begin have in common that Camp David was the location of an important summit conference that Begin attended. There are a number of problems that must be addressed in maintaining a self-updating memory. 1. what constitutes a good category in memory, i.e., what attributes does a good category have? 2. what kind of knowledge must be stored about each category to enable retrieval and new category creation? 3. how do categories relate to each other? 4. when is it appropriate to reorganize a category into smaller pieces? 5. how can generalized knowledge be added to new categories? The remainder of this paper will address some of these problems. RECOGNIZING SIMILARITIES BETWEEN EPISODES People notice common aspects of episodes and make generalizations in the normal course of understanding. Reorganization of memory requires noticing similarities and making generalizations based on those similarities. Generalized knowledge is needed to predict future occurrences, to elaborate on a context being understood, to help direct memory update, and as an aid in directing search during retrieval. Like people, CYRUS notices similarities between episodes and makes generalizations from them. Similar episodes in CYRUS are stored in the same MOP, along with the generalized knowledge built UP from them. MOPS act as event categories in memory holding episodes and knowledge about those episodes. The generalized information a MOP 331 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. holds resides in its "content frame" 141 and includes such things as typical preconditions and enablement conditions for its episodes, their typical sequence of events, larger episodes they are usually part of, their usual results, typical location, duration, participants, etc. The structure of individual episodes provides a framework for deciding whether two episodes are similar to each other. If, on two different trips, Vance is welcomed at the airport and then has a meeting with the president of the country he is in, then the episodic structure of the two trips will look alike, and we can say that the two episodes are similar. While on the second trip, he might be reminded [41 of the first one because of their similarities. In the same way, a person hearing about the two trips might be reminded of the first when hearing about the second. If the result of the first trip had been a signed accord, then he may predict, or at least hope, that an accord would be signed at the end of this trip also. If an accord is signed, he will generalize that when the first event of a diplomatic trip is a meeting with the head of state, then an accord will be reached. Later, he will be able to use that knowledge in understanding and retrieval. REORGANIZING EVENT CATEGORIES In order for such reminding and subsequent generalization to occur in CYRUS, its MOPS must be highly structured internally. Episodes are indexed in MOPS according to their important aspects. Indexing in a MOP is by content-frame components and includes sequence of events, location, participants, duration, etc. When an event is indexed similarly to an event already in memory, reminding occurs and generalizations are made based on their similarities. As a result, a sub-MOP of the larger MOP is formed to hold those episodes and their generalizations. With the addition of more episodes, new sub-MOPS are further divided. In creating new MOPS and building up generalized information, new knowledge, which can be used for later understanding and retrieval, is added to the data base. The actual processing when an event is added to a MOP depends on its relationship to events already in the MOP. One of the following four things is true about each component of an event description: 1. It is unique to this event in the MOP 2. It is semi-unique to this event in the MOP (it has happened once or a small number of times before) 3. It is often true of events in the MOP 4. It is typical of events in the MOP In case 1, when the descriptive property is unique, the event is indexed under that aspect in the MOP. For instance, one of the discriminations CYRUS makes on meetings is the topic of the contract being discussed. A meeting about the Camp David Accords is indexed under "contract topic = peace", and a meeting about military aid- to Pakistan is indexed under "contract topic = military aid". The first time CYRUS hears about a meeting in which Vance discusses mili tary aid, it will index that meeting uniquely in the "diplomatic meetings" MOP under the property "contract topic = military aid". If it were also the first meeting he had with a defense minister, then it would also be indexed uniquely under the occupation of its participants (because meetings are occupational). If a property being indexed has occured once before in a MOP (case 21, then reminding occurs, the two events are compared to see which other common aspects they have, and generalizations are made. When Vance meets again about military aid, CYRUS is reminded of the prior meeting because both have the same topic. It checks the descriptions of both to see what other similarities they have. If both, for example, are with defense ministers, it will conclude that meetings about military aid are usually with defense ministers. It also begins indexing within the new MOP: *Jr*********** Adding SMEET actor (Vance) others (defense minister of Israel) topic (Military aid to Israel) place (Jerusalem) to memory . . . Reminded of SMEET actor (Vance) others (defense minister of Pakistan) topic (Military aid to Pakistan) place (Washington) because both are "diplomatic meetings' both have contract topic "military aid" creating new MOP: meetings about military aid generalizing that when Vance meets about military aid, often he meets with a defense minister *****Jr******* Later, if CYRUS hears about a third meeting whose topic is military aid, it will assume that the meeting is with the defense minister of the country requesting aid (unless it is given contrary information). If asked for the participants of that event' i&w.'.11 be able to answer "probably the defense minister". If, on the other hand, a number of meetings about military aid with participants something other than defense ministers are added to memory, CYRUS will remove that generalization and attempt a better one instead. On entering the next meetings about military aid to memory, CYRUS will index them along with other events already indexed there. A new meeting about military aid will be entered into the "meetings about military aid" sub-MOP of "diplomatic meetings" and will be indexed within that MOP (case 3). In this way, reminding, generalization, and new MOP creation will occur within newly created MOPS. If a new meeting about military aid to Pakistan is added to memory, CYRUS will be reminded of the first because both will be indexed under "contract sides = Pakistan" in the "meetings about military aid" MOP. 332 No discrimination is done on properties that are typical (case 4) of events in a MOP (i.e., almost all events in the MOP fit that description). In that way, generalization can control the expansion of MOPS in memory. If memory has generalized that meetings are called to discuss contracts, then the fact that the topic of a later meeting is a contract will never be indexed. Appropriate aspects of the contract, however, will be indexed. Thus, if a new event with no unique aspects is added to memory, reminding of individual events does not occur, but generalizations already made are confirmed or disconfirmed. Disconfinned generalizations are removed. When CYRUS hears about yet another meeting in the Mid East about the Camp David Accords, it will not be reminded of any specific episodes, but will put the new meeting into the MOPS it fits into. *Jr*********** Adding SMEET actor (Vance) others (Begin) topic (Camp David Accords) place (Jerusalem) to memory . . . Putting it into MOP: meetings with Begin in Israel confirming generalizations Putting it into MOP: meetings about the Camp David Accords with Israeli participants confirming generalizations Putting it into MOP: meetings in Israel confirming generalizations a.. ***Jr********* IMPLICATIONS IN RETRIEVAL What are the implications of this indexing scheme in retrieval? The retrievability of an event depends on how distinct its description is, or how many of its features turn out to be significant. As events with similar properties are added to memory, their common aspects lose significance as good retrieval cues and category specifiers (case 4). An event with no unique or semi-unique descriptors will become lost in memory or "forgotten". Since events are indexed by their differences, they can be retrieved whenever an appropriate set of those differences is specified, but specification of only common aspects of events will not allow distinct episodes to be retrieved. Generalized knowledge can be used during retrieval in a number of ways. One important use is in guiding search strategy application (see 151). Generalized knowledge can also be used for search key elaboration. There is not always enough information given in a question to direct search to a relevant MOP or to a unique episode within a MOP. Generalizations and a MOP's indexing scheme can be used to direct the filling in of missing details. Only those aspects of a MOP that are indexed need be elaborated. Generalized information can be used to answer questions when more specific information can't be found. If CYRUS has made a generalization that Gromyko is usually a participant in SALT negotiations, it will be able to give the answer "probably Gromyko" to "Last time Vance negotiated SALT, who did he meet with?", even if it could not retrieve the last negotiating episode. In the same way, generalizations can be used for making predictions during understanding. CONCLUSIONS Good memory organization is crucial in a large information system. Some important processes memory organization must support include dynamic reorganization of memory, creation of new memory categories, and generalization. In this paper, we've tried to show how a long-term memory for episodic information can keep itself usefully organized. This requires a good initial organization plus rules for reorganization. Reorganization of memory happens through reminding, or noticing similarities between episodes, and generalization. The generalizations produced are useful both for controlling later memory reorganization and for retrieval. Some related problems which have not been addressed in this paper, but which are important, are updating generalized knowledge (particularly recovery from bad generalizations), judging the usefulness of new categories, and guiding indexing properly so that new MOPS and generalizations are useful and relevant. Those topics are currently being addressed. REFERENCES [ll DeJong, G. F. (1979). Skimming stories in real time: An experiment in integrated understanding. Research Report #158. Department of Computer Science, Yale University. [21 Kolodner, J. L. (1980). Organization and Retrieval from Long Term Episodic Memory. Ph.D. Thesis (forthcoming). Department of Computer Science, Yale University. [31 Norman, D. A. and Bobrow, D. G. (1977). Descriptions: a basis for memory acquisition and retrieval. Report #7703. Center for Human Information Processing, La Jolla, California. [41 Schank, R. C. (1979). Reminding and memory organization: An introduction to MOPS. Research Report #170. Department of Computer Science, Yale University. 151 Schank, R. C. and Kolodner, J. L. (1979). Retrieving Information from an Episodic Memory Research Report #159. Department of Computer Science, Yale University. also in IJCAI-6. [61 Williams, M. D. (1978). The process of retrieval from very long term memory. Center for Human Information Processing Memo CHIP-75, La Jolla, California. 333

1980

64

61

Meta-planning Robert Wilensky Computer Science Division De artment of EECS University o F California, Berkeley Berkeley, California 94720 1.0 INTRODUCTION This paper is concerned with the problems of planning related and understanding. because These problems are natural understander must apply knEwledge about people s language goals and necessary f lans in order to make the inferences o in a sto explain the behavior of a character story u~er~~~~;;;ky, 19,;3Fa). a planner, it mus? Thus while embody a theory of plilning knowledge. I have construction story dt?;e+;get rl * understanding PIZYhApp?ieZh$Z anliZm> F": concerned not with t R rogram. This paper is itself, but e understanding mechanism which is that part of its inde lanning knowledge is used to 21 * endent of whe her that knowledge if someone's behavior or to generate a p% "f,'", one's own use. One part of this theory of knowledge is essentially world knowled includes a plan;;i;fg e. classification of in entional f them, plans are used to achieve goals, etc.) and an actual body of knowledge about particula; elements (e getting some hing from someone 'f' 2 asking for som 7 thing is a way of . When one attempts to use this world knowledge to understand the intentions of a story's characters, a number of problems soon become apparent. particular, what ' difficult in understan%ng a person s behavii: iz iE':perating under, but the fact so much understanding the goal and that t#$% are usually numerous a situation. It is f oals and plans present in he interactions between these intentional elements that cause much of complexity in both understanding and planning. For example, consider the stories: following (1) John was in a hurr Vegas, I* to get to Las but he no iced that there were a lot of cops around so he stuck to the speed limit. (2) John was noticed eating dinner when he to break in to his house. that a thief was ti$;;f he finished his dessert, John called the police. Likewise, (2) strikes most strange since John should have rea%", 'Fo t:z z intruder more strongly. The unusualness of this ;;; ry is due not to kno;r;dge ;,";ut the plans goals involved, scheduling of these plans. ap arent unproductive 1 more intelligent planner would have dealt with t? e threat immediately, and then perhaps returned to ;+s meal when that situation had been disposed . Thus to understand the behavior of character, or to enerate an intelligent plana it is necessary f0 take into account th;! interacti ns between goals. programs. e.g., Sussman 1975, 9 y;;;py;+f;g deal interactions bv providi&tkpec??-!E pr~&a~~%emechanisms to deal with particular situations. Sussman s HACKER has a celebrated FZL tZ""~iEt knows about goals clobbering "brother oalslt, and detects this bug in plans suggested $Y the plan synthesizer. The difficulty with this ty e is that burying this knowledge a t of solution out how to plan in a rocedure assures that such knowled not %e shared by a program that wishe 2 et;o;zi this knowledge to understand someone else's behavior in a complicated situation. addition, there is a lot i"f structure as I hope to show to this knowle6ge that is missed in this fashion, to the tasks and which is extreme~l.l~seful both of planning as as plan understanding. 2.0 META-PLANNING One solution to this problem is to create a second body of planning knowledge that is called meta-planning. By this I mean that knowledge about now to plan should itself be expressed in P ;;ys,;f a set of g alsaf,;r thesg;an;pg '5 E recess meta-goals , achieve th meta-plans e~re~m~~Z1~n~~.atheMe~~~~oa~~~~~~~~ mechanism (or plan understa der) that is u produce a plan of action if or explanation 7 ed to from ordinary plans. For consider the following situation, example either Prom the point of view of plan understanding or plan generation: (7) John's wife called him and told him they were all out of milk. He decided to pick some up on his way home from work. Most intelligent planners would come up with John's -plan, assumi by a grocery store on T they knew that they pass he ;;ute home. In order to produce this plan, is necessary to go through the following processes: 1. 334 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. 2. A&j;sting one's plans according1 case, the plan is modifie 3' so Ia: to Produce a route that takes the planner near the grocery store. 2. The "go home" plan is suspended at the point at which the grocery store is reached. 3. 4. The "get milk" plan is executed. The "go home" plan is resumed. 2.1 Kinds Of Meta-goals The following is a brief description of the more important meta-goals along with the situations in so far encountered, which the and some standard plans applicable I arise o them. This list is not meant to be corn lete. merely reflects the current sta e ? of 0:: analysis. META-GOALS, SITUATIONS, AND META-PLANS 1 . Don't Waste Resources Situations to Detect 1 . Goal Overlap Associated meta-plans: 1 . Schedule Common Subgoals First 2. Plan Integration In terms.of meta-planning, this situation has the ;f;;o;;; structure: There ' an imnortant - d "Don't Waste Resou*Ees". Plan Piggybacking (find a new 'a g~~nflalls k",;k goa~;~ultaneously meta-pian fulfills the 'Don't Waste Resources" meta-goal. The advantage of the meta- is that the problem of how to Ei lanning approach eal with complex goal interactions can be stated as a problem to be solved ,,by applies to the same planning mechanism one ordinary" goals. For example, one may first try out a number of canned sol;;;ons, then some standard planning procedures, lf all else fails, try to construct a novel solution. Note that there are at least three important differences between meta-plannin and planning using constraints + Constraints and plan generators Ze asGike:ZE in that constraints reject plans, but they don't themselves propose new ones. In constrast meta-goals not violations, but su gest only pick up f new plans to fix the problem. Me a-goals are declarative structures, and thus may be used in the ex lanation process as well as in planning. In 3 are domain a dition, independent, meta-goals encoding about planning in general. only knowledge McDermott's notion of a *AotP,'n 0; secondary task comes closest to meta-planning I propose here. A policy is essentially constraint. ex licitly Theli;;rnary x* ifferences r;epeEezz;ted polic 3 and a - oal are inclu e goals tha 'i that meta-goal: are not constraints necessarily facts about p ,"~ni~~'as E meta-goals refer CXI;T,;~ policies their domain, include domain information; may specific of policies often entail the creation psuedo-tasks, whereas meta-goals have meta-plans that deviate less from the structure of normal plans. Hayes-Roth and Hayes-Roth (1978) uses the term meta-nlannina to refer to decisions about the planning process. While my use of the term similar to theirs the cyanning decisions under t K include all types of is meta-planning is name, and their not formulated in terms of explicit meta- 5 oals term to and meta-plans. I use the re er of this knowledge t and only to a subset convenien ly only when that knowledge is expressable in meta-goals and meta-plans. terms of explicit 2. Multiple Planning Options (more l%tin g:l) plan is applicable to a Associated meta-plans: 1 . Select Less Costly Plan 3. Plan Non-integrability situations in which the execution o two plahs will adversely affect one another e.5, one undoes B subgoal established by the other Associated meta-plans: 1. Schedule Sequentially 4. Recurring Goals (A goal repeatedly) arises Associated meta-plans: 1. 9 ubsume Recurring Goal Establish state that fulfills a r re:ondition for a plan for ndures he goal and which 7 over a P%~ of time see Wilensky, -5. Recursive subgoals (a subgoal is identical to a higher level goa , causing a potential infinite loop $ Associated meta-plans: 1. Select Alternate Plan 2. Achieve As Many Goals As Possible Situations to detect 1. Goal Conflict Associated meta-plans: various conflict resolution plans (see Wilensky 1978a) 2. Ass mme;;$;otGoal 3[ Conflict (Both goa s be accomplished‘if A kgr ormed F erformed beff;;; B, but B being difficulty) poses no 335 Associated meta-plans: Schedule Innocuous Plan First 2. Plan Splicing (If one plan has already been started, suspend it 6 divert to the other plan, ie"w p%?%s been executed) original plan when 3. Goal Competition (Goal interference with the goal of another planner) Associated meta-plans: 1. various anti-plans (plans to deal specificly with opposition) 2. various plans for resolving the competition 3. Maximize the Value of the Achieved Goals Situations to detect 1. Unresolvable Goal Conflict Associated meta-plans: 1. Abandon Less Important Goal 4. Don't Violate Desireable States Situations to detect 1. Danger Associated meta-plans: 1. Create a preservation goal 2. Maintenance Time Associated meta-plans: 1. Perform Maintenance 3. Anticipated Preservation Goal performance of another plan $?? cause the plann r e to preservation goal have a Associated meta-plans: 1. Select Alternate Plan 2. Protective Modification (Modify original plan so as n t to provoke preservation goal P Ef a stored lan for this % rotective clot ing, it R- wou d f oal is to wear be scheduled efore the initial plan. If not, then we could establish a subgoal of getting a raincoat. This mia:; spawn a plan that involves going would violate outside, the Recursive Subgoals condition. another x: The ~;ta--~la;ai~e;; * lan. t?nd to choose "Achieve one, the s Many Subgoals As Possible" meta-goal is activated, as a goal conflict is now seen to exist. resol- ution The meta-plans for goal conflict are attempted. If they fail, then an unresolvable goal conflict situation exists, and Maximize the Goals Achieved is activated. The meta-plan here is to abandon the less goal. The nlanner selects whichever important goal he values more and then abandons the other. 3 .O APPLICATIONS We are meta-planning understanding interactions :Ytple goal length el currently attemptin F use in two programs. AM,? sto ?$StE:derstand uses kno;;EtfEsabout Y w involving S. As PAM has been discussed sewhere, we will forego a discuss- ion of rts use of meta-planning here. Meta- P lanning is also being used in the developmen of a planning program called PANDORA (Plan ANalyzer with D namic Revision and Application . ai Organization, PANDORA is given a d,,esrcrit ;io;ozfsa $tuation and. creates. a plan R PANDORA . may have in that situation. developmen$z, dynamically told about new and changes it plans accordingly. References 11 Hayes-Roth, B. and Hayes-Roth, F. (1978). Cognitive Processes in Planning. RAND Report R-2366-ONR. 21 McDermott, Drew (1?78>. PLanning and $cting. In Cognitive Science vol. 2, no. . 31 Sacerdoti, E. D. (1977). A Structure for Plans and Behavior. ElZevier North-H?iIlanu, Amsterdam. 41 Schank, R. C. and Abelson, R. P. (1977). %%%%in~?@~'L%%&e%$lbaum Press, Billsdale, N.J. 51 Sussman G. J. (1975). A Corn uter Model of $%:l~o~;quisition. Aiiie?&&ETs~,- 61 Wilensky R. (1978). Understanding goal-based stories. Yale University Research Report #140. 336

1980

65

62

NARRATIVE TEXT SUMMARIZATION Wendy Lehnert Yale University Department of Computer Science New Haven, Connecticut ABSTRACT In order to summarize a story it is necessary to access a high level analysis that highlights the story's central concepts. A technique of memory representation based on affect units appears to provide the necessary foundation for such an analysis. Affect units are conceptual structures that overlap with each other when a narrative is cohesive. When overlapping intersections are interpreted as arcs in a graph of affect units, the resulting graph encodes the plot of the story. Structural features of the graph then reveal which concepts are central to the story. Affect unit analysis is currently being investigated as a processing strategy for narrative summarization. When a reader summarizes a story, vast amounts of information in memory are selectively ignored in order to produce a distilled version of the original narrative. This process of simplification relies on a global structuring of memory that allows search procedures to concentrate on central elements of the story while ignoring peripheral details. It is apparent that some hierarchical structure is holding memory together, but the precise formulation of this structure is much more elusive. How is the hierarchical ordering of a memory representation constructed at the time of understanding? Exactly what elements of the memory representation are critical in building this structure? What search processes examine memory during summarization? How are summaries produced after memory has been accessed? In this paper we will outline a strategy for narrative summarization that addresses each of these issues. This proposed representation for high level narrative analysis relies on affect units. An affect unit is an abstract structure composed of three affect states and four affect links. AFFECT STATES AFFECT LINES Positive Events (+> Motivation cm> Negative Events (-1 Actualization (a> Mental States (M) Termination (t) Equivalence (e> ------- This work was supported in part by ARPA contract N00014-75-C-1111 and in part by NSF contract IST7918463. For example, if John wants to buy a house, his desire is a mental state (Ml. If John subsequently buys the house, his desire is actualized by a positive event (+>. But if someone else buys it instead, John will experience that transaction as a negative event (-1 signalling actualization failure. These particular affect states are derived by recognizing an initiated goal (Ml, an achieved goal (+>, and a thwarted goal (-1. The status of a goal is just one way that an affect state can be recognized. A more complete account of affect state recognition is presented in [31. All affect states are relative to a particular character. If another buyer (Mary) takes the house, we have a negative event for John, and a positive event for Mary. We use a diagonal cross-character link to identify their two affect states as reactions to the same event: John Mary wants to buy (, M MJ" wants to buy a house is sold -/+ buys house The above configuration of four states and three links is the affect unit for "competition." Two actors have a goal, and success for one means failure for the other. "Success" and "failure" are primitive affect units contained within the competition unit. Success is recognized whenever a mental state is actualized by a positive event. Failure is the non-actualization of a mental state through a negative event. Now suppose John decides to get even by setting the house on fire. And suppose further that it takes two tries to get it going. John Mary wants to buy L M U ,:.' a wants to buy buys house house is sold desires fire 4- can't set fire e e desires fire "GM 5 gets fire going +-+-k house burns down The sale of the house to Mary motivates John to set the house on fire (Ml. This mental state fails to be actualized (-1 the first time he tries to commit 337 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. arson. But his desire persists in an equivalent mental state (M) and is then successfully actualized (+I by John setting the fire. This fire is a positive event (+I for John, but a negative event (-1 for Mary who suffers a loss. We first derive a baseline summary from the pivotal unit by accessing a "generational frame" associated with the pivotal unit. For example, a generational frame for retaliation is: "When Y caused a negative event for X, X caused a negative event for Y.” "Loss" is an affect unit that occurs whenever a negative event terminates a positive event in the sense of removing whatever satisfaction was derived from that positive event. When a loss wipes out a previous success, we get the affect unit for "fleeting success." When a smaller unit is embedded in a larger unit (e.g. "loss" is embedded in "fleeting success"), we recognize the structure of the larger unit as a "top level" affect unit. Using this convention, our story about John and Mary contains 4 top level affect units. (11 (21 (31 This is a conceptually abstract description of retaliation. To produce a reasonable summary, we must (1) instantiate the generational frame, and (2) augment it with information from units adjacent to the pivotal unit. We will try to convey what's involved by showing how a baseline summary evolves into a reasonable summary with the addition of information from adjacent units. (This sequence is not intended to reflect actual processing stages). Sl = Retaliation (the baseline summary) "When Mary prevented John from gett wanted, John set her house on fire. ing 11 something he s2 = Sl + Competition (11 represents "competition", (21 "fleeting success", and (3) "perseverance after failure." A fourth affect unit is recognized by merging the two equivalent mental states of John: "When Mary bought something that John set her house on fire." wanted, John s3 = S2 + Fleeting Success "retaliation" 'When Mary bought a hous set the house on fire." e that John wanted, John ,(x> mcA- aG; A, s4 = S3 + Perseverance After Failure "When Mary bought a house that John wanted, set the house on fire after two tries." John The unspecified (X) in the retaliation unit can be any affect state. In our story, John's negative event happened to be a positive event for Mary. If the information from the perseverance unit seems less important than the other contributions, there is a good reason for this. "Perseverence after failure" resides between two equivalent mental states that are merged within the retaliation unit. It is often desirable to ignore units that are lost when equivalent mental states are merged. Top level affect units for a narrative can be used as the basis for a graph structure that describes narrative cohesion. The nodes of the graph represent top level affect units, and an arc exists between two nodes whenever the corresponding affect units share at least one common affect state. The affect unit graph structure for our simple story looks like: Suppose for comparison, that John gave up on his intended arson after the first unsuccessful attempt. Then our affect analysis for the story would be a truncated version of the original: wants to buy M MiP wants to buy 4 /------+ buys house house is sold desires fire mS;; can't set fire a 4, Where C = "competition", F = "fleeting success", R = "retaliation", and P = "perseverance after failure." In general, the affect unit graph for a cohesive narrative will be connected. And in many cases, the graph will have a unique node whose degree (number of incident arcs) is maximal over all nodes in the graph. In our example, the retaliation unit has a uniquely maximal degree of 3. We will call any node of maximal degree a "pivotal unit." If a story has a unique pivotal unit, then that unit encodes the "gist" of the story. A good summary for the story will be based on the pivotal unit and its adjacent units. We st ill have a competition unit, but the other level units are now "motivation" and "fail ure": top "motivation" "failure" The a ffect unit graph now contains three connected units , with motivation acting as the pivotal unit: 338 The baseline summary is therefore built from a generational frame associated with motivation: “When a negative event happened to X, X wanted Z.” Augmenting this baseline summary with information from the competition and failure units, we derive a reasonable summary: Sl = Motivation (the baseline summary) “When Mary prevented John from getting something he wanted, John wanted to set her house on fire.” s2 = Sl + Competition “When Mary bought a house that wanted to set it on fire.” s3 = S2 + Failure “When Mary bought a house that John wanted, unsuccessfully tried to set it on fire.” John wanted, John These two examples illustrate how pivotal units and their adjacent units can be used to drive processes of narrative summarization. While many simple stories will succumb to an algorithm that uses a pivotal unit for the baseline summary, other stories yield affect unit graphs that do not have unique pivotal units. For example, consider “The Gift of the Magi” by 0. Henry. In this story a young couple want to buy each other Christmas presents. They are very poor but Della has long beautiful hair, and Jim has a prized pocket watch. To get money for the presents, Della sells her hair and Jim sells his pocket watch. Then she buys him a gold watch chain, and he buys her expensive ornaments for her hair. When they realize what they’ve done, they feel consoled by the love behind each other’s sacrifices. The affect unit analysis is perfectly symmetrical across the two characters. Both characters have affect units for nested subgoals, a regrettable mistake, two distinct losses, and a hidden blessing. The affect unit graph for this story is connected, but there is no unique pivotal unit: Both “HM” and “WM” are pivotal units. These units correspond to their regrettable mistakes. Let the family of a node N be the set of nodes adjacent to N. Then this graph can be partitioned into the families of “HM” and “WM.” “HN”, “WN”, “HLl”, and “WLl” are boundary units in the sense that each of their families cross this partition. It is not easy to come up with a one sentence summary of “The Gift of the Magi ,I’ but it can be done by concentrating on the boundary units of maximal degree (“HN” and “WN”). These are the units for their nested subgoals: “Della sold her long locks of hair to buy her husband a watch chain, and he sold his watch to buy her ornaments for her hair.” This example shows how the summarization algorithm must be sensitive to structural features of affect unit graphs. In this case the connected graph can be partitioned into two families of two pivotal units, and the simplest summary originates from the boundary units of maximal degree. The process of narrative text summarization relies on (1) a high level of conceptual representation that readily encodes coherence within the narrative, and (2) a process of language generation that can easily be driven by that high level memory representation. In this paper we have attempted to show how affect units and their resulting graph structures are well-suited to these requirements. We have necessarily omitted important explanations concerning techniques of recognition for affect units and the processes of generation that express target summaries in English. The representational system itself requires further explication concerning which affect unit configurations are legitimate (there are 15 legal configurations of the form “state” - “link” _ “state” rather than 36). the combinatorially possible Using these 15 primitive configurations, we can represent speech acts, voluntary compliance, coerced compliance, the notion of a double-cross, and similar abstractions of complexity [31 e equivalent conceptual The use of affect units in narrative summarization is currently being explored by psychological experiments on text comprehension and within a computer implementation for the BORIS system 121. While related work on text summarization has been conducted using story grammars , there are serious flaws in that approach due to the top-down nature of story grammars Ill. These difficulties will not arise with affect unit approach because affect units are constructed by bottom-up processing at the time of understanding. The resulting affect unit graphs are consequently far more flexible in their content and structure than the rigid hierarchies of fixed story grammars. This flexibility is the key to recognizing a diverse range of plot structures without recourse to an a priori taxonomy of all possible plot types. REFERENCES 111 Black, J. B., and Wilensky, R. (1979) “An Evaluation of Story Grammars .” Cognitive Science. Vol. 3, No. 3, pp. 213-230. [21 Dyer, M. and Lehnert, W, (1980) “Memory Organization and Search Processes for Narratives . ” Department of Computer Science TR i/175. Yale University. New Haven, Conn. [31 Lehnert, W. (1980) “Affect Units and Narrative Summarization” Department of Computer Science TR #179. Yale University. New Haven, Conn. 339

1980

66

63

HEARSAY-III: A Domain-Independent Framework for Ex Robert Balzer Lee Ermzrn Philip London Chuck Williams USC/Information Sciences Institute* Marina del Rey, CA 90291 Abstract Hearsay-Ill is a conceptually simple extenslon of the basic ideas in the Hearsay-II speech-understanding system [a]. That domain- dependent expert system was, in turn, a product of a tradition of increasingly sophisticated production-rule-based expert systems. The use of production systems to encapsulate expert knowledge in manageable and relatively independent chunks has been a strong recurrent theme in Al. These systems have steadily grown more sophisticated in their pattern-match and action languages, and in their conflict-resolution mechanisms [la]. In this paper, we describe the Hearsay-Ill framework, concentrating on its departures from Hearsay-II. 1. The Heritage From Hearsay-II Hearsay-II provided two major advances -- the structuring of the workspace, called the bhckboard in Hearsay, and the structuring of the search, via scheduling mechanisms. The blackboard provided a two-dimensional structure for incrementally building hierarchical interpretations of the utterance: - levels which contained different representations (and levels of abstraction) of the domain (phones, syllables, words, phrases, etc.). - a location dimension (the time within the spoken utterance) which positioned each partial interpretation within its level. Knowledge sources (KSs), relatively large production rules, were agents which reacted to blackboard changes produced by other KSs and in turn produced new changes. The expertise was thus organized around the activity of building higher-level, more encompassing partial interpretations from several nearby lower- level partial interpretations (e.g., aggregating three contiguous syllables into a word) and producing lower-level ones from higher-level ones (e.g., predicting an edjacent word on the basis of an exisiting phrase interpretation). Within this aggregation-based interpretation-building paradigm, Hearsay-II also provided a method for exploring alternative interpretations, i.e., handling search. Interpretations conflicted if they occupied the same or overlapping locations of a level; conflicting interpretations competed as alternatives. Thus, in addition to organizing activity around the interpretation-building process, Hearsay-II also had to allocate resources among competing interpretations. This required expertise in the form of critics and evaluators, and necessitated a more complex ‘This research wss supported by Defense Advrnced Research Projects Agency contrect DAK15 72 C 0308 View.s and conclusions contained in this document ere those of the authors and rho&l not be interpreted em representing the official opinion or policy of DARPA, the U.S. Government, or eny other person or egency connected with thorn. scheduler, which at each point one previously-matched KS.** chose for execution the action of 2. The Directions for Hearsay-III To this heritage, we bring two notions that motivate most of our changes: - Through simple generalization, the can be made domain independent. Hearsay approach - Scheduling is itself so complex a task that the Hearsay blackboard-oriented knowledge-based approach is needed to build adequate schedulers.*** Our generalizations activities: consist of systematizing the main blackboard - aggregating several interpretations at one a composite interpretation at a higher level, level into - manipulating alternative interpretations (by creating a placeholder for an unmade decision, indicating the alternatives of that decision, and ultimately replacing the placeholder by a selected alternative), and - criticizing proposed interpretations. The complexity of scheduling is handled by introducing a separate, scheduling blackboard whose base data is the dynamically created activation records of KSs. These include both the domain-dependent KSs, which react to the regular, domain blackboard, and scheduling KSs, which react to changes on the scheduling blackboard as well. The organization of these activations (with agendas, priorities, etc.) is left to the application writer; Hearsay-Ill provides only the basic mechanisms for building expert systems. Thus domain KSs can be viewed as the legal move generators (competence knowledge) with the scheduling KSs controlling search (performance knowledge). 3. Blackboard Structure In Hearsay-II, nodes on the blackboard, which represented partial interpretations, were called hypotheses. In Hearsay-Ill, we adopt the more neutral term unit. Hearsay-Ill provides primitives for creating units and aggregating them, i.e., associating them hierarchically. The blackboard is implemented in a general- purpose, typed, relational database system (built on top of INTERLJSP), called A83. AP3 has a pattern-matching language; this “A good discussion of scheduhng WI Hearsay-II csn be found in (51 “‘This notlon, in one form owemple, [6], (21. and [III another, is common to 8 number of others, for 108 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. is used for retrieval from the blackboard. AP3 also has demons; the triggering pattern which the application writer supplies as part of the definition of a KS is turned into an AP3 demon. The blackboard levels of Hearsay-II have been generalized somewhat into a tree-structure of classes. Each unit is created permanently as an instance of some class. The unit is, by inheritance, also an instance of all superclasses of that class. The apex of the class tree is the general class Unit. The immediate subclasses of Unit are DomainUnit and SchedulingUnit; these classes serve to define the domain and scheduling blackboards. All other subclasses are declared by the application writer, appropriate to his domain. For example, in the SAFE application, which is a system for building formal specifications of programs from informal specifications [l], one of the subclasses of DomainUnit is ControlFragment, and it has subclasses Sequence, Parallel, Loop, Conditional, Demon, etc. The semantics of the unit classes other than Unit, DomainUnit, and SchedulingUnit are left to the application writer. Any unit may serve to denote competing alternative interpretations. Such a unit, called a Choice Set, represents a choice point in the problem-solving. The Choice Set is a place- holder for the interpretation it represents; it can be dealt with as any other unit, including its incorporation as a component into higher-level units. Associated with a Choice Set unit are the alternatives of the choice. These may be explicit existing units or they may be implicit in a generator function associated with the Choice Set. When appropriate, a KS may execute a Select operation on a Choice Set, replacing it with the selected alternative. The Selection can be done in a destructive, irrevocable manner, or it can be done in a new contezt, retaining the ability to Select another alternative. Contexts are described more in Section 5. Hearsay-II’s location dimension (e.g., time-within-utterance in the speech-understanding domain) is not imposed on the Hearsay-III blackboard. The application writer may create such a dimension, either inherently in the interconnection structure of units or explicitly as values associated with the units. The flexibility of the underlying relational database system allows such constructs to have first-class status, for example, to be used in KS triggering patterns. 4. Scheduling Hearsay-Ill retains Hearsay-II’s basic sequencing of KS execution: When the triggering pattern of a KS is matched by a configuration of data on the blackboard, an activation record is created containing the information needed to execute the KS in the environment of the match. At some later time, the activation record may be selected and saecuted, i.e., the KS’s action, which is arbitrary code, is run. The executing KS has available to it the blackboard data that triggered it, which usually serves as the initial focus for the activity of the execution. Each KS execution is indivisible; it runs to completion and is not interrupted for the execution of any other KS activation. The effect of a KS execution is an updated blackboard. lndpendent activations of the same KS can pursue the same exploration by retrieving (potentially private) state information from the blackboard. The scheduling problem is: given the current state of the system, select the appropriate activation record to execute next. The separation of KS execution from triggering allows for complex schedultng schemes (i.e., a large collection of activations may be available from which to select). To allow the application writer to use the Hearsay problem-solving features for building such schemes, several mechanisms were added in Hearsay-Ill: - Each activation record is a unit on the scheduling blackboard. The application writer supplies, as part of the definition of each KS, code to be executed when the triggering pattern is matched; this code computes a scheduling-blackboard class (level) in which the activation record will be created. - When executed, scheduling KSs are expected to make changes to the scheduling blackboard to facilitate organizing the selection of activation records. In addition to triggering on changes to the domain blackboard, scheduling KSs can trigger on changes to the scheduling blackboard, including the creation of activation records. The actions a scheduling KS may take include associating information with activation records (e.g., assigning priorities) and creating new units that represent meta-information about the domain blackboard (e.g., pointers to the current highest-rated units on the domain blackboard). The scheduling blackboard is the database for the scheduling problem. - The application writer provides a base scheduler procedure that actually calls the primitive Eaecute operation for executing KS activations. We intend the base scheduler to be very simple; most of the knowledge about scheduling should be in the scheduling KSs. For example, if the scheduling KSs organize the activation records into a queue, the base scheduler need consist simply of a loop that removes the first element from the queue and calls for its execution. If the queue is ever empty, the base scheduler simply terminates, marking the end of system execution. 5. Context Mechanism While Choice Sets provide a means for representing an unmade decision about alternative interpretations, we still need a method of investigating those alternatives independently. For that, Hearsay-Ill supports a context mechanism similar to those found in Al programming languages such as QA4 [lo] and CONNIVER [9]. The method by which KS triggering interacts with the context mechanism allows controlled pursuit of alternative lines of reasoning. A KS triggers in the most general context (highest in the tree) in which its pattern matches. ixecution of that KS occurs in the same context and, unless it explicitly switches contexts, its changes are made in that context and are inherited down toward the leaves. Contexts are sometimes denoted as unsuitable for executing KSs -- a condition called poisoned. Poisoned contexts arise from the violation of a Hearsay constraint (e.g., attempting to aggregate conflicting units). In addition, a KS can explicitly poison a context if, for example, the KS discovers a violated domain constraint. A KS activation whose execution context is poisoned is placed in a wait state until the context is unpoisoned. Special KSs, called poison handlers, are allowed to run in poisoned contexts, and specifically serve to diagnose and correct the problems that gave rise to the poisoning. A common application for the context mechanism arises when alternative interpretations lack good “locality”. First consider the 109 exampfe of SAFE% Pfanning Phase, which uses Choice Sets to represent alternative interpretations for control fragments. In the case of the input sentence “Send an acknowledgment to the imp and pass the message on to the host.” a Choice Set served well. The possible interpretations for this sentence include being put in parallel or in sequence with an existing structure; since all alternatives would be positioned identically in the existing aggregate structure, the Choice Set unit can be placed where the chosen interpretation eventually will go. In some cases, however, locality is lacking. An example is the input sentence, “After receiving the message, the imp passes it to the host.” The possible interpretations for this include a demon (“The occurrance of r triggers y”) and a sequence to be embedded in an existing procedure (“After r do y”). Since the demon interpretation resides at the same structural level as the procedure into which the sequence would be embedded, there is no convenient place to put the Choice Set representing these alternatives. Instead, the KSs producing these alternative interpretations put them in brother contexts, so that each can be pursued independently. 6. Relational Database As mentioned earlier, the blackboard and all publicly accessible Hearsay-Ill data structures are represented in the AP3 relational database. In addition, any domain information which is to cause KS firing must also be represented in the database. This is because KSs are AP3 demons, and their triggering is controlled by activity in the database. The AP3 database is similar to those available in languages such as PLANNER [7], but also includes strong typing for each of the relational arguments in both assertion and retrieval. These typed relational capabilities are available for modeling directly the application domain. 7. Implementation and Current Status The Hearsay-Ill system is implemented in AP3, which in turn is implemented in INTERLISP [12]. AP3 was chosen as an implementation language because it already contained the mechanisms needed to support Hearsay-Ill (e.g., contexts, demons and constraints, and strong typing). In fact, the design of Hearsay-Ill’s initial implementation was almost trivial, being largely a set of AP3 usage conventions. However, efficiency considerations have forced a substantial implementation effort. Hearsay-Ill has been tested on two small applications: a cryptarithmatic problem and a cryptogram decoding problem. Three major implementation efforts are currently underway. The first of these, as described above, is the reimplementation of the SAFE system [l]. Second, Hearsay is being used as the basis for a system for producing natural language descriptions of expert= system data structures [8]. Finally, the system is being used as the basis for a “jitterer” which automatically transforms a program so tHat a transformation chosen by a user is applicable 143. The Hearsay-Ill architecture seems to be a helpful one. The separation of competence knowledge from p.erformance knowledge helps in rapidly formulating the expert knowledge required for a solution. Pretiminary experience with the larger applications now under development seem to bear this out, and seem to indicate that performance (scheduling) is a difficult issue. The flexibility that the Hearsay-III architecture gives toward developing scheduling algorithms will undoubtably go a long way toward simplifying this aspect of the overall problem-solving process. Acknowledgments We wish to thank Jeff Barnett, Mark Fox, and Bill Mann for their contributions to the Hearsay-Ill design. Neil Goldman has provided excellent and responsive support of the AP3 relational database system. Steve Fickas, Neil Goldman, Bill Mann, Jim Moore, and Dave Wile have served as helpful and patient initial users of the Hearsay-Ill system. References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. Balzer, R., N. Goldman, and D. Wile, “Informality in Program Specifications,” IEEE Trans. Software Enp. SE-4, (21, March 1978. Davis, R., Meta-Rules: Reasoning About Control, MIT Al Laboratory, Al Memo 576, March 1980. Erman, L. D., F. Hayes-Roth, V. R. Lesser, and D. R. Reddy, “The Hearsay-II Speech-Understanding System: Integrating Knowledge to Resolve Uncertainty,” Computing Surveys 12, (2), June 1980. (To appear) Fickas, S., “Automatic Goal-Directed Program Transformation,” in 1st NationaL Artificial Intelligence Conf., Palo Alto, CA, August 1980. (submitted) Hayes-Roth, F., and V. R. Lesser, “Focus of Attention in the Hearsay-II System,” in Proc. 5th International Joint Conference on Artificial Intefligence, pp. 27-35, Cambridge, MA, 1977. Hayes-Roth, B., and F. Hayes-Roth, Cognitive Prooesses in Phning, The Rand Corporation, Technical Report R-2366-ONR, 1979. Hewitt, C. E., Description and Theoretical Analysis (Using Schemata) of PLANNER: A Language for Proving Theorems and Manipulating Models in a Robot, MIT Al Laboratory, Technical Report TR-258, 1972. Mann, W. C., and J. A. Moore, Computer as Author -- Results and Prospects, USC/Information Sciences Institute, Technical Report RR-79-82, 1979. McDermott, D., and G. J. Sussman, The CONNIVER Reference Manual, MIT Al Laboratory, Memo 259a, 1974. Rulifson, J. F., R. J. Waldinger, and J. A, Derksen, “A Language for Writing Problem-Solving Programs,” in ZFIP 71, pp. 201-205, North-Holland, Amsterdam, 1972. Stefik, M., Planning with Constraints, Ph.D. thesis, Stanford University, Computer Science Department, January 1980. Teitelman, W., lnterlisp Reference ManuaL, Xerox Palo Alto Research Center, 1978. Waterman, D. A., and F. Hayes-Roth, Pattern-Directed fnference Systems, Academic Press, New York, 1978. 110

1980

67

64

QUANTIFYING AND SIMULATING THE BEHAVIOR OF KNOWLEDGE-BASED INTERPRETATION SYSTEMS* V.R. Lesser, S. Reed and J. Pavlin Computer and Information Science Department University of Massachusetts Amherst, Mass. 01003 ABSTRACT The beginnings of a methodology for quantifying the performance of knowledge-sources (KSs) and schedulers in a knowledge-based interpretation sys tern are presented. As part of this methodology, measures for the “reliability” of an intermediate state of system processing and the effectiveness of KSs and schedulers are developed. Based on the measures, techniques for simulating KSs and schedulers of arbitrary effectiveness are described. I INTRODUCTION The development and performance-tuning of a knowledge-based interpretation sys tern like the Hearsay-II speech understanding system [I] is still an art. There currently does not exist sufficient formal methodology for relating the performance characteristics of such a system to the performance characteristics of its components; i.e., knowledge-sources (KSs) and schedulers**. For that matter, there does not even exist an adequate framework for quantifying the performance of the system and its components in an uniform and integrated way. Thus, when the initial operational configuration, Cl, of the Hearsay-II speech understanding system had poor performance, there existed no methodology for detailing in a quantifiable way what types of performance improvements in specific components would be needed National -0412 and Projects of Naval The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or imp1 ied, of the National Science Foundation, the Defense Advanced Research Projects Agency, or the US Government. made some efforts in this we feel that his model of these abstract to capture adequately important issues. (A detailed be found in [41.) to improve significantly the overal 1 system performance . Therefore, the development of the completely reorganized C2 configuration, which turned out to have much superior performance, was based on “seat of the pants” intuitions. These intuit ions were testable only when the new set of KSs were integrated together into a working system. In the following sections we present the beginnings of a methodology for quantifying the performance of KSs and schedulers. We then show how this methodology can be used to simulate the performance of an upgraded component in a working system, so that more accurate estimates can be made of the overall performance improvement that would be realized if the component were actually upgraded. II A MODEL FOR A HEARSAY-LIKE ---v-p KNOWLEDGE-BASED SYSTEM -- 111 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. into more abstract and encompassing higher-level hypotheses (partial interpretations). The lower-level hypotheses are said to support the higher-level hypotheses. This aggregation process, accomplished by synthesis KSs , involves the detection of local consistency (or inconsistency) relationships among hypotheses*. In KS processing, the belief-values of supporting hypotheses are not changed as a result of detection of local consistency, as would be the case in a relaxation process. The construction of a higher-level hypothesis and its associated belief-value is an explicit encoding of the nature and degree of consistency found among its supporting hypotheses. The hope is that this incremental aggregation process resolves the uncertainty among competing interpretations and, simultaneously, distinguishes between correct and incorrect interpretations. We have not discussed KSs as solely reducing the uncertainty in the system, because uncertainty is a measure of the distribution of belief-values and does not reflect the accuracy of hypotheses. We feel that KS processing causes changes in both certainty and accuracy in a system’s database and we have developed a measure, called ‘lreliability”, that combines the two. A good KS will produce higher-level hypotheses which are more reliable than the lower-level hypotheses supporting them. III MEASURING THE SYSTEM STATE Basic to our view of processing in knowledge-based systems is the concept of system state. The system state at any point in time is -- the current set of hypotheses and their relationships to the input data. It is through measures of the system state that we can talk in a uniform manner about the performance of KSs and schedulers. For purposes of measuring reliability, we associate a hidden attribute with each hypothesis which we call its truth-value. This attribute measures the closeness of the hypothesized event to the correct event. For task domains in which a solution is either totally correct or incorrect, we quantify truth-value as either 1 (true) or 0 (false) , while in domains in which there are solutions of varying degrees of acceptability, truth-values range between 1 and 0. One way of evaluating an intermediate state of processing is by measuring the reliability of the set of all possible complete interpretations (final answers> that are supported (at least in part) by hypotheses in the current state. We feel that a direct measure of this sort is not feasible because it is very difficult in general to relate the set of partial interpretations to the very large set of complete interpretations they can support. We take an alternative approach, called + To simplify this presentation, we focus here on synthesis KSs only, though prediction, verification, and extrapolation KSs also have a place in our model. reflecting-back, which is based on two premises. First, the creation of a high-level hypothesis is a result of detecting the constitency among its supporting hypotheses. This creation process iS an alternative to actually changing the belief-values of the supporting hypotheses, as occurs in the relaxation paradigm. Thus, the creation of a high-level hypothesis implicitly changes the reliability of its supporting hypo the se s . This change can be traced down to the input hypotheses whose reliability is implicitly improved to the extent that they are aggregated into reliable high-level hypo the se s . Second, we assume that processing which implicitly improves the reliability of the input hypotheses al so improves the reliability of the complete interpretations supported by these hypotheses. In the reflecting-back approach, we associate with each input hypothesis the highest belief hypothesis it supports. The truth- and belief-values of this highest belief hypothesis are reflected-back to the input hypothesis . The process is illustrated in Figure 1. It should be stressed that the hypotheses’ truth-values and reflected-back values are used only for measuring the system state; they are not available to KSs during processing. Our measure for the reliability of an intermediate system state is based on a measure of the reliability of input competitor sets computed from the reflected-back belief- and truth-values. Intuitively, a measure of reliability for a competitor set should have the following properties, based on both the be1 ief- and truth-values of its hypotheses: 1. With respect to accuracy, reliability should be high if a true hypothesis has high belief-value or if a false hypothesis has a low belief-value, while reliability should be low if a true hypothesis has low belief-value or a false hypothesis has high belief-value; 2. With respect to uncertainty, reliability should be high if one hypothesis in a competitor set has a high belief-value and the rest have low belief-values, while reliability should be low if all the hypotheses have similar belief-values. A measure for the reliability, RC(S), of a competitor set, S, that captures and adequately combines both of these properties is: RC(S) = 1 - avg iTV(h)-BV(h) I h in S which is equivalent, in the case of binary truth-values, to the correllation of truth- and belief-values: RC( s) = avg CBV(h)*TV(h) + (l-BV(h) )*(I-TV(h))1 h in S where BV(h) is the belief-value of hypothesis h and 112 TV(h) is the truth-value of hypothesis h. Other measures may also be applicable, but of those we considered this one best captures our intuitive notion of reliability. Based on this measure of competitor set reliability, we can construe t, for instance, a measure of processing effectiveness associated with the current intermediate sys tern state. This measure is the average over the input competitor sets of the difference between the initial reliability and the reliability of the intermediate state. The initial reliability is the average reliability of the competitor sets formed from input hypotheses, using the initial belief- and truth-values. The reliability of the intermediate -- state is the average reliability of the competitor sets formed from the input hypotheses, where the reflected-back truth- and belief-values of hypotheses are used in place of the original ones. In Figure 1, if the three competitor sets were the only ones on the input level and each one contained just the two hypotheses shown, then the initial rel iab il i ty , RI, is .517, the reliability of the intermediate state, RS, is .575, and processing effectiveness, PE, is .058: RI = l/3 (.7+.6)/2 + (.5+.4)/2 + (.7+.2)/21 = .517 RS q 1/3[(.6+.5)/2 + (.7+.5)/2 + (.45+.7)/21 = .575 PE q RS - RI = .058 A positive PE value, as in this example, indicates that some uncertainty and error has been resolved. The larger the value, the more resolution, the more effective is the processing. IV MEASURING THE RESOLVING POWER OF A KS ___-- - We define the instantaneous resolving power of a KS as a change in reliability due to a single KS - execution. This change is measured on competitor sets construe ted from the KS input hypotheses. Thus, instead of calculating the reflected-back reliability of the entire system state, the procedure is localized only to the subset of the state directly affected by the KS execution. We Figure 1: an example of the reflecting back process. INPUT LEVEL: COMPETITOR SETS The system state contains three competitor sets (represented by boxes) on the input level, and a number of higher level hypotheses. The hypotheses are represented as circles with belief-values on the left and truth-values CT=1 and F=O) on the right. The initial values are in the bottom of the circle, the reflected-back values are in the top half. The thick 1 ines show the support 1 ink to the highest belief-value hypothesis supported by each input hypothesis and indicate the source of the reflected-back values. Hypotheses which intervene between the inputs and their highest-belief supported hypotheses are not shown. 113 measure the change in reliablilty as a result of KS execution by measuring before KS processing the reliability of the KS input competitor sets, then measuring after KS processing the reliability of these sets based on values reflected-back from the KS output hypotheses, and finally taking the difference between the results obtained. The resolving power of a KS can now be defined ---- as its average instantaneous resolving power over a series of system executions. This view of KS resolving power does not take into account the global impact of KS execution on the entire system state and on future processing. Rather, it is a local, instantaneous measure of the effects of KS execution. The global effects occur through KS interactions, which we believe should be separated from our measure of the resolving power of a single KS. V SIMULATING A KS -- Given a formal measure of KS resolving power, we can simulate KSs of any desired power. This is accomplished by introducing an lloracle" which knows how to judge the closeness of a hypothesized interpretation to the correct interpretation (this is the source of the truth-values). Our reliability measures can thus be calculated during processing rather than in retrospect, after the system has completed processing. Therefore, a system does not hav% to complete execution in order to be evaluated. A KS is simulated in stages: candidate We believe that our approach to simulating KSs of different resolving power, which makes heavy use of an oracle, will prove useful in designing and debugging knowledge-based systems**. However, there are some limitations: ------------------ I In many cases, it is relatively easy to design a KS which provides moderate accuracy. Most of the effort in knowledge engineering is spent in increasing this accuracy to gain superior performance. Our simulation of KS resolving power is based on a combination of simple knowledge about local consistency and reference to an oracle, while real KSs infer truth from local consistency alone (and falsehood from local inconsistency). The behavior of different simulated KSs sharing similar errors in knowledge will not be correlated due to our statistical approach to KS simulation. Given these limitations, we do not expect a simulated KS to behave exactly the same as a real KS. We hope, however, the essential behavior of a KS has been captured so that system phenomena are adequately modelled. In order to validate our models of KS power, we plan to analyze the behavior of KSs in some existing knowledge-based systems. A measures of KS power will be taken for an existing KS and then the KS will be replaced by a simulated KS of the same power, and the overall system behavior compared in the two cases. The results of these experiments should give us some understanding of the extent to which data derived from our simulation studies can be used to predict the behavior of real systems. VI SIMULATION OF - ACCURACY IN THE -- SCHEDULER Reliability measures can also be used in the simulation of a scheduler of a specific accuracy. The task of a scheduler is choosing a KS instantiation for execution. A KS instantiation is a KS-stimulus pair, where the stimulus is the set of hypotheses which caused the KS to be considered for scheduling. The scheduler evaluates alternative instantiations according to its knowledge of the characteristics of the KSs, the stimuli, and the current state of processing. The effects of future processing are not factored into this model of scheduling; we take an instantaneous view of scheduling decisions. Because of this, we are unable to model scheduling algorithms such as the "shortfall density scoring method" 171 which use information about future processing. We hope to develop a formulation that includes this type of information. A good scheduler chooses for execution the KS instantiation that will most improve the reliability of the current system state. The accuracy of a single scheduling decision is defined --- relative to the performance of an optimum scheduler, which uses accurate information about the resolving power of the KSs and the reliability of the KS stimuli and system state. The accuracy of a scheduler is the average of the accuracy of -- many scheduling decisions. We steps: view the optimum scheduling process in ------------------ ** The work of Paxton C5l comes the closest our approach, but was much more limited. two to 114 1. For each KS instantiation on the scheduling queue, make accurate predictions concerning its instantaneous resolving power. These predictions involve determining the truth-value of the stimulus hypotheses (using the oracle) and knowledge of the resolving power of the KS. 2. Make accurate predictions as to the global system state which would result from scheduling each instantiation given the predictions of step 1. These predictions will determine optimum ratings for the instantiations and result in an optimum schedule. Our approach to modelling the scheduler is to obtain statistically accurate ratings for the instantiations, based on the optimum schedule, and then choose for execution an instantiation from within the ordering which results. The position in the ordering of the chosen instantiation depends on the desired accuracy of the scheduler being modelled; the closer to the top of the order, the more accurate the scheduler. We feel it would be an error to model scheduling only as a function of the truth-value of stimulus hypotheses. Real schedulers do not have access to the truth-values of hypotheses, but only infer truth from belief-values and processing history. The point is that two instantiations of the same KS, whose stimulus hypotheses have equivalent characteristics (same belief-value, level of abstraction, database region, processing history, etc.) except for their truth-values would be rated the same by even the best scheduler. Additionally, in order to determine the rating of a KS instantiation, real schedulers [ 31 consider other factors, besides the characteristics of the stimulus hypotheses. For example, schedulers take into account such factors as the balance between depth-first vs. breadth-first processing or between executing KSs that work in areas with rich processing history vs. executing KSs that work where little processing has been done. These additional considerations are, in fact, heuristics which attempt to capture the concept of improvement in the reliability of the system state. Thus, in our view, a scheduler should be characterized in terms of its ability to estimate the improvement in system state reliability, rather than its ability to detect the truthfulness of the instantiation's stimulus hypotheses. We could have modelled the scheduler just as we modelled KSs, with a candidate evaluator and a scheduling resolver. The candidate evaluator would take the generated KS instantiations and give them ratings based on simple scheduling knowledge. The scheduling resolver would minimally alter these ratings (with statistical perturbation) to produce an ordering for the instantiations which corresponds to a desired scheduler accuracy. For several reasons, too complicated to discuss in this short paper, we have not used such an approach for modelling schedulers. Further details of this issue and a more detailed formulation of scheduling measures ar e disc this paper c41. in an extended version of VII SUMMARY This work represents the beginnings of a methodology for understanding in quantitative terms the relationship between performance of a knowledge-based system and the characteristics of its ccmponents. This quantification may also allow us to develop simulations of these systems which can accurately predict the performance of alternative designs. ACKNOWLEDGMENTS We would like to recognize the helpful comments on various drafts of this paper given by Daniel Corkill and Lee Erman. REFERENCES II11 c21 c31 r41 [51 [61 c71 Erman, L. D., F. Hayes-Roth, V. R. Lesser and R. Reddy (1980). "The Hearsay-II Speech Understanding System: Integrating Knowledge to Resolve Uncertainty," Computing Surveys, 12:2, June 1980. Fox, M. S. (1979). "Organizational Structuring: Designing Large Complex Software," Technical Report CMU-CS-79-155, Department of Computer Science, Carnegie-Mellon University, Pittsburgh, Pennsylvania. Hayes-Roth, F. and V. R. Lesser (1977>, "FOCUS of Attention in the Hearsay-II Speech Understanding System," In Proceedings of the Fifth International Joint Conference on Artificial Intelligence-1977, P. 27-35, Cambridge, Massachusetts, 1977. Lesser, V. R., J. Pavlin and S. Reed (19801, "First Steps Towards Quantifying the Behavior of Knowledge-Based Systems," Technical Report, Department of Computer and Information Science, University of Massachusetts, Amherst, Massachusetts. Paxton, W. H. (19781, I1 The Executive System," in D. E. Walker, (editor), Understanding Spoken Language, Elsevier, North-Holland, N. Y., 1978. Rosenfeld, A. R. , R. A. Hummel , and s. w. Zucker (19761, "Scene Lab eling by Rel axation Operators," IEEE Transactions on Systems, Man and Cybernetics, SMC-6, pg. 420-433, 1976. Woods, W.A. (1977)) "Shortfall and Density Scoring Strategies for Speech Understanding Control,11 In Proceedings of the Fifth -- International Joint Conference on Artificial Intelligence-1977, p. 18-26, - Cambridge, Massachusetts, 1977. 115

1980

68

65

Representation of Task-Specific Knowledge in a Gracefully Interacting User Interface Eugene Ball and Phil Hayes Computer Science Department, Carnegie-Mellon University Pittsburgh, PA 15213, USA Abstract Command interfaces to current interactive systems often appear inflexible and unfriendly to casual and expert users alike.’ We are constructing an interface that will behave more cooperatively (by correcting spelling and grammatical errors, asking the user to resolve ambiguities in subparts of commands, etc.). Given that present-day interfaces often absorb a major portion of implementation effort, such a gracefully interacting interface can only be practical if it is independent of the specific tool or functional subsystem with which it is used. Our interface is tool-independent in the sense that all its information about a particular tool is expressed in a declarative tool description. This tool description contains schemas for each operation that the tool can perform, and for each kind of object known to the system. The operation schemas describe tne relevant parameters, their types and defaults, and the object schemas give corresponding structural descriptions in terms of defining and derived subcomponents. The schemas also include input syntax, display formats, and explanatory text. We discuss how these schemas can be used by the tool-independent interface to provide a graceful interface to the tool they describe, 1. Introduction Command interfaces to most current interactive computer systems tend to be inflexible and unfriendly. If the user of such a system issues a command with a trivial (to a human) syntactic error, he is likely to receive an uninformative error message, and must re-enter the entire command. The system is incapable of correcting the error in the “obvious” way, or of asking him to retype only the erroneous segment, or of providing an explanation of what the correct syntax really is. Anyone who has used an interactive computing system is only too familiar with such situations, and knows well how frustrating and time-consuming they are, for expert as well as novice users. We are involved in a project to build an interface which will behave in a more flexible and friendly way, one that will inferact graceful/y. As we have described in earlier work [3, 41, graceful interaction involves a number of relatively independent skills including: m the parsing of ungrammatical input, either to correct it or to recognize any grammatical substrings; 1 This research was sponsored by the Defense Advanced Research Projects Agency (DOD), ARPA Order No. 3597, monitored by the Air Force Avionics Laboratory Under Contract F33615-78-C-1551. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed OI Implied, of the Defense Advanced Research Projects Agency or the US Government. o robust communication techniques to ensure that any assumptions the system makes about the user’s intentions are implicitly or explicitly confirmed by the user; o the ability to give explanations of how to use the system or the system’s current state; e interacting to resolve ambiguities or contradictions in the user’s specification of objects known to the system; o keeping track of the user’s focus of attention; e describing system objects in terms appropriate to the current dialogue context. Providing these facilities is clearly a major programming task requiring extensive use of Artificial Intelligence techniques (see [2] for just the flexible parsing aspect). We believe that it is unrealistic to expect the designers of each interactive sub-system (or tool) to implement a user interface with these capabilities. Therefore, instead of constructing a gracefully interacting interface for a single application, we are attempting to build a tool independent system, which can serve as the user interface for a variety of functional sub-systems. The availability of a tool independent user interface would greatly simplify the construction of new computer sub-systems. Currently, even if the new system is not intended to be gracefully interacting but merely to perform according to minimal standards, a large amount of implementatron effort must be devoted to user interface issues. The system designers must decide on a style of interaction with the user, select the general format and detailed syntax of all commands, and provide for the de.tection of illegal input. The command language must then be thoroughly checked to ensure that it does not contain ambiguities or misleading constructions and that likely error sequences will not be misinterpreted and cause unrecoverable system actions. Often, the design can only be completed after an Initial implementation of the system has produced feedback about the usability of the human interface. This design process represents the minimum effort necessary to produce a system that is even usable by a large number of people; if a superior (but still far from gracefully interacting) interface or one which can be used by non-programmers is required. much more work must be expended. Editing facilities, which are required in most interactive systems (at least for correction of typed input), must be fully integrated into the sub-system; compatibility with other editors in common use on the computer must be considered, even though this may lead to difficult interactions with the sub-system command language. Error detection and reporting must be Improved; generating coherent diagnostics for the inexperienced user can be very difficult indeed. Online documentation must be provided, 116 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. including reasonable facilities which allow a user to quickly find the answer to specific (although vaguely expressed) questions. The complexity of this task often means that most of the implementation effort in adding a new tool to a computer system is absorbed by the user interface. Technological trends are aggravating the problem by raising the level of performance expected of an interface. In particular, as high-resolution graphics displays equipped with pointing devices become available, users expect to use menu-selection and other more sophisticated forms of input, and to see system output displayed in an attractive graphical format. The very recent, but growing, availability of speech input and output will intensify this pressure for sophistication. An additional reason for constructing a tool-independent interface is to make the computer system as a whole appear consistent to the user. If the interfaces for different tools use different conventions, then no matter how sophisticated each of them is individually. the user is likely to be confused as he moves from one to another because the expectations raised by one may not be filled by the other. For all these reasons, we are attempting to make our gracefully interacting interface as tool-independent as possible. In the remainder of this paper we outline the system structure we have developed, and go on to give further details about one component of this structure, the declarative format in which information about the tool is made available to the interface, together with sketches of how the tool-independent part of the system uses the information thus represented. 2. System Structure The basis for our system structure is the requirement that the interface contain no tool-dependent information. All such information must be contained in a declarative data base called the tool description. In an effort further to improve portability and reduce duplication of effort between interfaces implemented on different hardware configurations, we have made a second major separation between the device-dependent and -independent parts of the interface. The resulting structure is illustrated in figure 1. User Agent Figure 1. User Interface System Structure The intelligent functions of the interface, those itemized above, are isolated in a tool and device independent User Agenf, which interacts with the tool through a narrow interface that is completely specified by the declarative tool description. Communication between the Agent and the user is not direct, but goes via a device-dependent Front-End, which allows the Agent to specify its output in a high-level device-Independent manner, and which preprocesses the user’s input into a standard, device-independent, format. Communication between the Agent and Front-End is thus restricted to a well-defined format of input and output requests. Display formats in which to realize the tool’s and Agent’s high-level output requests are specified declaratively in the tool description. The basic function of the Agent is to establish from the user’s input what functional capability of the tool the user wishes to invoke and with what parameters he wishes to invoke it. Once this is established, the Agent issues the appropriate request to the tool and reports to the user relevant portions of the tool’s response. To make this possible, the tool description includes a specification of all the operations provided by the tool in terms of their parameters and their types, defaults, etc., plus specifications of all the abstract objects manipulated by the tool in terms of their defining (and descriptive) sub-components. This representation of operations and objects follows Minsky’s frames paradigm [7] in the spirit of KRL [l] or FRL [13]. The representation allows the Agent to follow the user’s focus of attention down to arbitrarily deeply nested aspects of object or operation descriptions to resolve ambiguities or contradictions. This facility depends on the tool to provide resolution of object descriptions into sets of referents. The tool description also specifies the syntax for the user’s input descriptions of the objects and operations. The Agent applies the grammar thus specified to the user’s input (as pre-processed by the Front-End) in a flexible way, providing the kinds of flexible parsing facilities mentioned above. The user may also request information about the tool or other help, and the Agent will attempt to answer the query with information extracted from the tool description, and displayed according to tool-independent rules. Besides requesting that the Front-End output text strings to the user, the Agent may also specify instances of system objects. The Front-End will then drsplay the objects according to a display format specified in the tool description. For the Front-End we are using, whrch operates through a graphics display equipped with a pointing device. this allows the user to refer directly to system objects by pointing. The Front-End reports such pointing events to the Agent in terms of the system object refered to. Typed input is pre-processed into a stream of lexical items, and other pointing events, such as to menus, can also be reported as lexical items. We are also experimenting with a limited-vocabulary, single word (or phrase) speech recognizer, isolated in the Front-End. It’s output can also be reported as a lexical item. This concludes the overview of the system structure. For the remainder of the paper, we will concentrate on the representation employed in the tool-description, and the way the information, thus represented, is used by the remainder of the system. Our examples will be in terms of the tool being used as a test-bed for the development of the Agent and Front-End: a multi-media message system, capable of transmitting, receiving, filing, and retrieving pieces of electronic mail whose bodies contain mixtures of text, speech, graphics, and fax. 3. Representation of task specific information A functional sub-system, or tool, is characterized for the user interface program by a data base which describes the objects it manipulates and the operations it can perform. This fool description is a static information structure (provided by the sub-system implementor) which specrfies everything that lhe User 117 Agent needs to know about the tool. We’ll first give a brief overview of the structure of the tool description and how the Agent uses it to provide an interface to the sub-system. Then the content of the data base will be explained in more detail, with examples showing how this information is utilized in the processing of commands from the human user. The tool description consists of: o Declarations of the data objects used by the tool. These declarations specify the internal structure of each object type defined within a particular sub-system. The tool may also contain references to object types that are defined in a global data base and are used by many different tools (e.g. files, user names, dates and times). The object declaration provides rules for displaying an instance of the object, syntax for descriptions of it in commands, and documentation that can be used to explain its function to the user. o Descriptions of the operations which the tool can perform. Each operation entry specifies the parameters that the Agent must provide to the tool to invoke that action. It also defines the legal syntax for the command, provides some simple measures of its cost and reversability, and supplies a text explanation of its purpose. As mentioned earlier, the primary goal of the Agent is to help the human user to specify sub-system operations to be executed. To carry out this function, it parses the user’s commands (including text, pointing, and possibly spoken input) according to the syntax specifications in the tool description. It decides which operation has been selected and attempts to fill out the parameter template associated with it. This process may involve interpreting descriptions of sub-system objects, negotiating with the user about errors or ambiguities that are discovered, and explaining the meaning of command options, 3.1. Object Descriptions The tool description contains a declaration of each data type that is defined within that sub-system. Data objects which will be manipulated by both the Agent and the tool are represented as lists and property lists (sets of name-value pairs), using the formalism defined by Postel for the communication of Internet messages [9]. This representation is self-describing in that the structure and type of each data element is represented explicity in the object. Thus, complexly structured objects can be transferred between the User Agent and the tool, and the Agent can interpret them according to the information contained in the tool description. For example, the following is the internal representation of a simple message (primitive elements are integers or text strings, and brackets are used to delimit sets of name-value pairs): [ StructureType: Objectlnstance ObjectName: Message Sender: [ PersonName: [ First: John Middle: Eugene Last: Bali] Host: [ Site: CMU Machine: A ] 1 Recipient: [ PersonName: [ First: Phil Last: Hayes ] Host: [ Site: CMU Machine: A ] 1 Copies: [] Date: [ Year:1980 Month:April Day:10 Weekday:Thursday AMPM: AM Hour: 11 Minutes: 16 Seconds: 371 Subject: “Meeting tomorrow?” Body: “Phil, Could we meet tommorrow at 1 pm? -Gene” 1 The structure of a Message is defined in the tool data base by a message schema. This schema declares each legal field in the object and its type, which may be a primitive type like TEXT or an object type defined by another schema in the tool description. The schema also specifies the number of values that each field may have, and may declare default values for new instances of the object. The following is a simplified schema for a message object and some of its components: StructureType: ObjectSchema ObjectName: Message DescriptionEvaluation: ToolKnows Schema: Sender: [ FillerType: Mailbox ] Recipient: [ FillerType: Mailbox Number: OneOrMore ] Copies: [ FillerType: Mailbox Number: NoneOrMore ] Date: [ FillerType: Date ] Subject: [ FillerType: MultiMediaDocument ] Body: [ FillerType: MultiMediaDocument ] After: [ FillerType: Date UseAs: DescriptionOnly ] Before: [ FillerType: Date UseAs: DescriptionOnly ] StructureType: ObjectSchema ObjectName: Mailbox DescriptionEvaluation: ToolKnows Schema: [ PersonName: [ FillerType: PersonName ] Host: [ FillerType: Host ] 11 StructureType: ObjectSchema ObjectName: PersonName DescriptionEvaluation: OpenEnded Schema: [ First: [ FillerType: TEXT ] Middle: [ FillerType: TEXT Number: NoneOrMore] Last: [ FillerType: TEXT ] II StructureType: ObjectSchema ObjectName: Host DescriptionEvaluation: ToolKnows Schema: [ Site: [ FillerType: TEXT Default: CMU ] Machine: [ FillerType: TEXT Default: A ] 11 In addition to defining the structure of an instance of a message object, the schema includes fields which are used by the Agent to interpret descriptions of messages. The DescriptionEvaluation field tells the Agent how to evaluate a description of an object in this class. For example, ToolKnows indicates that the sub-system is prepared to evaluate a description structure and return a list of instances matching the description. A description structure is an instance of the object with special wild card values for some fields and with possible extra DescriptionOnly entries, such as the After field in the example above. Since a description of one object may reference other objects (“the messages from members of the GI project”), the Agent uses the hierarchy defined by the object declarations to guide its evaluation of the user’s commands. Each level generates a new sub-task in the Agent which processes that portion of the description, and is responsible for resolving ambiguities that it may encounter. This structure also makes it possible to follow the user’s focus of attention, since new input may apply to any of the currently active subgoals (“No, only the ones since October” or “No, only the ones at ISI”). 118 Each object declaration also includes information which is used information network at a node representing the data type, which is by the Front-End module to display instances of that object type. connected to Zog frames for other objects referenced by that type. Several different formats may be defined; the tool (or Agent) The user can find information about a related sub-system object selects an appropriate format by name each time it displays an by choosing a link to follow (with a menu selection); that frame is object. The format declaration specifies which fields of the object quickly displayed. The legal syntax for descriptions of the object to display and their layout; it may also provide parameters to the and links to frames for operations which manipulate it are also Front-End which invoke special capabilities (highlighting, font included in the Zog network. In the following example selection) of the display hardware. For example, the following documentation frame for Message, the italicized entries are section of the description for a Message defines two different buttons which can be activated to request more information about styles of message header display. a specific topic: DisplayFormat: [ ShortHeader: // style: From Hayes on 18-Mar [ Text: “From &Sndr& on &Day&” Sndr: [Field: Sender/PersonName/Last FaceCode: Bold] Day: [ Field: Date Style: ShortDay ] 1 FullHeader: // From Eugene Ball on lo-Apr-80 11:16am about ‘Meeting tommorrow? [ Text: “From &Sndr& on &Day& about ‘&Subj&“’ Sndr: [ Field: SenderIPersonName Style: FullName ] Day: [ Field: Date Style: FullDate ] Subj: [ Field: Subject Style: FullSubj ] 11 Messase: (Multi-Media Message System) Each message contains the author and date of origination and is addressed to (one or more) destination mailboxes; copies may optionally be sent to additional destinations. The body of a message may contain uninterpreted text, images, sketches, and voice recordings. Example syntax: ‘messages [from Person] [before/after Date] [about String]’ Detailed syntax Each object declaration also defines the legal syntax that can be used in commands that refer to objects of that type. In order to understand a phrase like “the messages from Phil since March”, the Agent uses the syntax definition associated with the object type Message, which in turn refers to the syntax for other objects like Date and PersonName. In the example below, question marks indicate optronal syntactic elements, asterisks mark fields that can be repeated, slashes indicate word class identifiers, and ampersands mark places where the syntax for other object types is to be expanded. Some syntactic entries also specify correspondences to particular fields in the object; as a command is parsed, the Agent builds description structures which represent the objects referred to in the command. Thus, the phrase “since March” results in an appropriate After clause in the Message description. The grammar defined by these syntax entries is applied to the user’s input in a flexible way [2], so that grammatical deviations such as misspellings, words run together, fragmentary input, etc. can still be parsed correctly. Related object types: Mailbox Display Multi Media Document Operations: Send Date Edit 3.2. Operation Descriptions Each sub-system operation which can be invoked by the Agent is also described by an entry in the tool data base. An operation entry specifies the parameters that the Agent must provide to the tool to have it perform that action. The object type of each parameter is declared and the tool description can optionally indicate that a parameter position may be filled by a set of such objects. In addition, constraints on the legal values of a parameter are sometimes provided, which can help the Agent to avoid requesting an illegal operation. Syntax: [ Pattern: (?/Determiner /MessageHead */MessageCase) Determiner: (the (all ?of ?the) every) MessageHead: (messages notes letters mail) MessageCase: [ StructureType: Operation OperationName: Forward Reversible: false Cost: moderate Parameters: [ Message: [ FillerType: Message Number: OneOrMore ] Recipient: [ FillerType: Mailbox Number: OneOrMore ] Forwarder: [ FillerType: Mailbox MustBe: CurrentUser ] ( [ Syntax: (/From &Mailbox) StructureToAdd: [ Sender: &Mailbox ]] [ Syntax: (/Before &Date) StructureToAdd: [ Before: &Date]] [Syntax: (/After &Date) 1 Syntax: [ Pattern: (/Forward %Message to %Recipient) Forward: (forward send mail (pass on) deliver redeliver) 1 Explanation: “Message Forwarding StructureToAdd: [ After: &Date ]] 1 From: (from (arriving from) (that came from) (/Mailed by)) Mailed: (mailed sent delivered) Before: (before (dated before) (that arrived before)) After: (after since (dated after)) 1 A copy of a message that was delivered to you can be sent to another person with the Forward command. You must specify the message to forward and the destination mailbox. Sample syntax: ‘Forward the last message from Phil to Adams at ISIE”’ Finally, the object description provides information which is used to automatically construct documentation and provide answers to user requests for help. Each object contains a brief text explanation of its structure and purpose in the sub-system, which can be presented to the user in response to a request for information. The documentation is also placed into a Zog [14] The example entry for the forward operation also mcludes a declaration of the legal syntax for the command, and a text entry which will be included in its documentation frame. It also indicates that this command is not reversible (once executed it cannot be undone), and that it is moderately expensive to execute. This information is used by the Agent to select an appropriate style of interaction with the User; for example, irreversible operations will usually require explicit confirmation before the request is given to the sub-system for execution. 119 4. Conclusion References The design and implementation of a good user interface for a computer sub-system is a difficult and time-consuming task; as new techniques for communication with computers (especially high resolution displays and speech) gain widespread use, we expect this task to become even more expensive. However, we also feel that the typical user interface must be made much more robust, graceful, and intelligent. For this goal to be feasible, substantial portions of the interaction with the user must be independent of the details of the application, so that the development cost of the user interface code can be shared by many sub-systems. Therefore, we are designing a generalized User Agent which can be used to control a variety of different sub-systems. The Agent carries on a dialog with the human user; it can understand a variety of different command styles, recognize and correct minor syntactic or spelling errors, supply default values for command arguments based on context, and provide explanations when requested. All of the information that the Agent needs to know about the application system is explicitly stored in a tool description provided by the sub-system implementor. This paper 1. Bobrow, D. G. and Winograd, T. “An Overview of KRL-0, a Knowl&ge Representation Language.” Cognitive Science 1,1 (1977). 2. Hayes, P. J. and Mouradian, G. V. Flexible Parsing. PrOC. of 18th Annual Meeting of the Assoc. for Comput. Ling., Philadelphia, June, 1980. 3. Hayes, P. J., and Reddy, R. Graceful Interaction in Man-Machine Communication. Proc. Sixth Int. Jt. Conf. on Artificial Intelligence, Tokyo, 1979, pp. 372-374. 4- Hayes, P. J., and Reddy, R. An Anatomy of Graceful Interaction in Man-Machine Communication. Tech. report, Computer Science Department, Carnegie-Mellon University, 1979. 5. Kernighan, Brian W. and Ritchie, Dennis M. The C Programming Language. Prentice-Hall, Inc., 1978. 6. Metcalf, Robert and Boggs, David. “Ethernet: Distributed has concentrated on the content of that data base, detailing the information represented there and demonstrating how the Agent can apply it to provide a sophisticated interface to a specific Packet Switching for Local Computer Networks.” Comm. ACM 79, 7 (July 1976), 395404. application system. 7. Minsky, M. A Framework for Representing Knowledge. In The tool description is represented in a unified formalism, which enables us to maintain a single data base which specifies all of the task-specific attributes of a particular sub-system. Because the information is stored in a single format, it can easily be utilized by multiple portions of the interface system. For example, a single syntax description is used to parse user commands, to generate Winston, P., Ed., The Psychology of Computer-Vision, McGraw Hill, 1975, pp. 211-277. 8. Newell, A., Fahlman, S., and Sproull, R.F. Proposal for a joint effort in personal scientific computing. Tech. Rept. , Computer Science Department, Carnegie-Mellon University, August, 1979. explanations of system actions, and to construct documentation Of the options available in the tool. 9. Postel, J. Internet Message Protocol. Draft Internet Experiment Note, Information Sciences Institute, Univ. of The initial implementation of the Agent will provide the user Southern California, April, 1980. interface for the Multi-Media Message System; an electronic mail facility which manipulates messages containing mixtures of text, lo- Rashid, R. A proposed DARPA standard inter-process recorded speech, graphics, and images. The system is being communication facility for UNIX version seven. Tech. Rept. , implemented as a multiple machine, multiple language distributed Computer Science Department, Carnegie-Mellon University, system: screen management (multiple windows) and graphics February, 1980. support are provided in Bcpl [l I] on a Xerox Alto [15] with a high resolution raster display and pointing device: audio recording and 1 1 - Richards, M. BCPL: A tool for compiler writing and systems playback is controlled by a DEC PDP-11 in C [5]; the Front-End programming. Proceedings of the Spring Joint Computer module and User Agent are implemented in C and LISP Conference, AFIPS, May, 1969, pp. 34557-566. respectively, on a VAX-l l/780 running Unix [12]; and the tool (message system) runs in C on the VAX. The system modules 12. Ritchie, D. M. and Thompson, K. “The UNIX Time-Sharing communicate using a message based Inter-Process System.“ Comm. ACM 77, 7 (July 1974), 365375. Communication facility [lo] within Unix, and a packet broadcast network (Xerox Ethernet [6]) between machines. Most of the 13. Roberts, R. B. and Goldstein, I. P. The FRL Manual. A. I. system components are currently running as individual modules, the first version of a single integrated system should be completed Memo 409, MIT Al Lab, Cambridge, Mass., 1977. by June 1980. Because of our goal of a smoothly working, robust, 14. and graceful system, we expect to continue tuning and improving Robertson, G., Newell, A., and Ramakrishna, K. ZOG: A the implementation for at least another year. The system will Man-Machine Communication Philosophy. Tech. Rept. , eventually be moved to a single powerful personal computer, Carnegie-Mellon University Computer Science Department, where we expect it to make substantial contributions to the CMU August, 1977. Spice (Scientifc Personal Integrated Computing Environment [8]) development effort. 1 6 . Thacker, C.P., McCreight, E.M., Lampson, B.W., Sproull, R.F., and Boggs, D.R. Alto: A personal computer. In Computer Structures: Readings and Examples, McGraw-Hill, 1980. Edited by D. Siewiorek, C.G. Bell, and A. Newell, second edition, in press. 120

1980

69

66

AN EFFICIENT RELEVANCE CRITERION FOR MECHANICAL THEOREM PROVING* David A. Plaisted Department of Computer Science University of Illinois Urbana, Illinois 61801 ABSTRACT To solve problems in the presence of large knowledge bases, it is important to be able to de- cide which knowledge is relevant to the problem at hand. This issue is discussed in [l 1. We pre- sent efficient algorithms for selecting a relevant subset of knowledge. These algorithms are presene ed in terms of resolution theorem proving in the first-order predicate calculus, but the concepts are sufficiently general to apply to other logics and other inference rules as well. These ideas should be particularly important when there are tens or hundreds of thousands of input clauses. We also present a complete theorem proving strate- gy which selects at each step the resolvents that appear most relevant. Thisstrategy is compatible with arbitrary conventional strategies such as P - deduction, locking resolution, et cetera. Also, 1 this strategy uses nontrivial semantic information and "associations" between facts in a to human problem-solving processes. I RELEVANCE FUNCTIONS Definition. A support set for a set S of clauses is a subset Ti of S such that S-Ti is con- way similar sistent. A support class for S is a set {Tl,..., Tk] of support sets for S. Definition. A (resolution) proof of C from S is a sequence C1,C2 ,...,C, of clauses in which Cn is C and each clause C i is either an element of S (an input clause) or a resolvent of two preceding clauses in S. (Possibly both parents of Ci are identical.) The length of such a refutation is n. A refutation from S is a proof of NIL (the empty clause) from S. Definition. A relevance function is a func- tion R which, given a set S of clauses, a support class T for S, and an integer n, maps onto a sub- set Rn(S, T) of S having the following property: If there is a length n refutation from S, then there is a refutation from Rn(S, T) of length n or less. Thus if we are searching for length n refutations from S, we need only search for length n refuta- * This research was partially supported by the Na- tional Science Foundation under grant MCS-79-04897. tions from R,(S, T). In fact, the derivation from R,(S, T) will be a subderivation of the derivation from S, for all relevance functions considered here. Thus if there is a length n P1-deduction from S, there will be a P l-deduction of length n or less from R,(S, T), and similarly for other corn plete strategies. Definition. Suppose S is a set of clauses. The connection graph of S, denoted G(S), is the graph whose nodes are the clauses of S, and which has a directed arc from C to C 1 2 labeled (Ll, L2)' if there are literals L 1 c C 1 and L2 E C2 such that Ll and z2 are unifiable. Such graphs have been introduced and discussed in [2]. Note that there will also be an arc labeled (L 2' Ll> from C 2 to Cl in the above case. Definition. A path from Cl to Cn in G(S) is a sequence Cl,C2,..., Cn of clauses of S such that there is an arc from C i to c i+l in G(S), for lli<n. Also, the length of the path is n. Definition. The distance d(Cl, C2) between Cl and C2 in G(S) is the length of the shortest path from Cl to C2 in G(S), and m if no such path exists. Definition. If S is a set of clauses, T is a support class for S, and n is a nonnegative inte- ger, then Q,(S, T) is EC E S: d(C, Ti)(n in G(S) for all Ti in T}, where d(C, Ti) is min{d(C, D):D E Ti}. Intuitively, if d(Cl, C2) is small, Cl and C 2 are "closely related." Also, Q,(S, T) is the clauses that are closely related to all the sup- port sets. Typically we will know that several clauses are essential to prove a theorem, and each such clause by itself can be made into a support set. Definition. A set S of clauses is fully matched if for all C there exists C2cS an a ES for all literals L LicC2 such that Ll an EC a 2 L1 are unifiable. 79 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Definition. R (S, T) is the largest fully matched subset of Q"(S, T). Thus we obtain R (S, T) from Qn(z, T) bynrepeatedly deleting Claus& containing unmatched" literals. This definition is not ambiguous, since if Q (S, T) contains more than one nonempty fully mate ed subset, then Rn(S, R T) is the union of all such subsets. Theorem 1. The function R is a relevance function. That is, if there is a length n refuta- tion from S, and T is support class, then there is a refutation from R (S, T) of length n or less. In fact, there is s&h a refutation from R Pl 6, T) Proof. Assume without loss of generalgty that NIL appears only once in the refutation and that every clause in the refutation contributes to the derivation of NIL. Let S1 be the set of input clauses appearing in the refutation. Then S1 is connected, intersects all the support sets, and has at most n elements. binary trees we can show that Using properties ;f elements. Note that R (S T) Sl has at most r;;l Vance criterion. Thatnisl; is a "globaL" relg- it depends in a non- trivial way on all the input clauses and on inter- actions between all the support sets in T. II EXAMPLES Let S be the Pl P2 -- Pl P2 E Ql p2 Q2 P3 Q7 following Q6 ?% 44 Q5 set of clauses: Q4 P4 LB 44 P3 P4 EL Q7 P4 Here ?i Ql indicates (??, Ql], i.e., E V Ql et cetera. Let Tlbe {{Pl], {P2]], let T2 be {{E]), and T = {Tl, T2). Then R (S, T) = R,(S, T) = R -- (S, T) = $j but R (S, T) =+Pl],{P2}" {Pl P2 P3? {p3,P4], {x}] which is in fact a miAimal'in:on-' sistent subset of S. For a second example, let S be the following: IN(a,box) IN(x,box) 1 IN(x,room) IN(x,room) > IN(x,house) E(x,house) ON(x,box) 1 m(x,box) ON(x,street) ' %?(x,house) IN(x,house) > IN(x,village) G(house,box) AT(house,street) ON(b,box) ON(c,street) m(d,village) Let Tl be {{IN(a,box))) and let T2 be ([IN(x, house)]]. Also, T = {Tl, T2]. R2(S, T) = R3(S, T) = 0 but R Then Rl(S, T) = (S, T) = {(IN(a, box)], (m&,box), IN(x,room) 4, {IN(x,room) IN(x, house)], {IN(x,house))]. This is a minimal'incon- sistent subset of S. Here "box", "room", "house", "a" et cetera are constants and x is a variable. Note that we cannot always guarantee that this re- levance criterion will yield minimal inconsistent . 80 sets as in these examples. is a set of clauses. Let Suppose S I I 4 1 be the length of S in characters when written in the usual way. Let Lits(S) be the sum over all clauses III ALGORITHMS C in S, of the number of literals in C. If S is a set of propositional clauses and IISII = m9 then G(S) may have O(m2) arcs. However, we can construct a modified version G,(S) of G(S) which has the same distances between klauses as G(S) does but which has only O(m) arcs. The idea is to reduce the number of arcs as follows: sup- pose C ,C. are the clauses containing L and D1, . . . . Dk%;e'tha clauses containing z. Then we add a node NL and arcs as follows in Gl(S): The numbers indicate the lengths of the arcs. Sim- larly, there are arcs of the form Di~-Nr,.f3C i2' Although Gl(S) is not a connection graph, and has arcs of length 0 and 1, it preserves distances be- tween clauses as in G(S). Using this modified con- nection graph, we have linear time algorithms to do the following, if S is a set of propositional clauses ,T=C T1 ,...,Tk] is a support class a suppo rt set, andn is a posi tive in teger: 2. Construct G1(S) from S. Find {C c S. d(C, T;> 5 nl. , Ti is 3. Given Q (S, T) for stpport class T, to find Rnn( S, T). Since step 2 must be performed 1 TI times to obtain Q (S, T), the total algorithm to obtain R (S, T) rzquires O(lT/*IIS/ I) time. (Here ) Tl isnthe num- ber of support sets.) The algorithm to find {C E S: d(C,T.) < r$is a simple modification of standard shortestipath al- gorithms. gorithms see [ 31. For a presentation of these standard al- This can be done in linear time because the edge lengths are all 0 and 1. pute R,(S, T) as follows: We com- Definition. If S is a set of clauses, let M(S) be the largest fully matched subset of S. Note that Rn(S, T) = M(Q,(S, T)). The following algorithm M1 computes M(S) for set S of propositional clauses in linear time. This algorithm can therefore be used to compute R,(S, T) if S is a set of propositional clauses. Note that t is a push-down stack. procedure Ml(S); tt- empty stack; for all L such that LEC or ~EC for some C&S do clauses(L) f {C~S:LEC]; - count(L) + I clauses(L)1 od; for all C&S do member(C)T TRUE: for all LEC do ' - if count (L) = 0 then - push C on t; member(C) + FALSE fi od while t not empty do - pop c off t; for all L&C- do count(L) ycount(L) - 1; if count(L) = 0 then - for all C I% E clauses(l) do if mem er(C,) then - - push C on t; member Cl) + FALSE fi t - od fi - - od *;- return((CeS: member(C) = TRUE]); end Ml; If S is a set of first-order clauses, then G(S) can be constructed in O(Lits(S)*I IS I I) time using a linear unification method [4]. This bound results because a unification must be attempted between all pairs of literals of S. The number of edges in G(S) is at most Lits(S)2. Given G(S), we can find {C E S: d(C, T.) < n] in time propor- tional to the number of &dges in G(S) (since all edge lengths are one). Also, given Q (S, T), we can find R (S, T) in time proportiona f to the num- ber of edggs in G(S) by a procedure similar to M above. The total time to find R (S, T) is there-l fore O<I T19cLits(S)2 + I IS) I*LitsTfS)). If 11 Sl I= m then the time to find R (S, T) is O(m2( TI). By considering only the przdicate symbols of literals, the propositional calculus algorithms can be used as linear time preprocessing step to eliminate some clauses from S. An interesting problem is to compute R (S, T) efficiently for many values of n at the sa#ie time. There are methods for doing this, but we do not discuss them here. IV REFINEMENTS A. Connected Components Proposition 1. If there is a length n refu- tation from S, and T is a support class for S, then there is a length n refutation from one of the connected components of R IF1 (S, T). Also, the connected components can be found in linear time [ 31. B. Iteration Definition. If T = {T1,T2, . . ..Tk} is a sup- port class for S and Sl is a subset of S then Th Sl (T restricted to S1) is {Tl n S may be that T//R (S, ,...,Tkn Sl]. It T) # T or t at distances in ii G(Rn(S, T)) arendifferent than distances in G(S). This motivates the following definitions. Definition. R;(S, T) = S, R;(S, T) = Rn(S,T), and if i>l then Ri(S, T) = Rn(Riql(S, T),TlRi-l(S, n T)). Also, Ry(S, T) is the limit of the sequence R;(S, T), R$, T), R;(S, T), *.. l Proposition 2. Rn i+l(S, T) GRi (S, T) for i>l. - Therefore the limit Ri(S, T) exists. Also, Rw(S, T) can be computed in at most I SI iterations oh@ *>. thannth&? Can it be computed more efficiently Theorem 2. If there is a length n refutation from S, and T is a support class for S, then there is a length n refutation from one of the connected components of R m (S, T). $1 Proposition 3. For all i>O there exist n, S and T such that R:+'(S, T) = R: (S, T) # Rn i-l(S, T). Thus this computation can take arbi- trarily long to converge. Proof. Let n = 2, S = {P. 1 Pi+l: l<i<k] U CP i+l 3 P. : l<i<k). -- T. = {Pi: P -- Let T = {+l,T2,T3} where J i+l: i s j(mod3)) U {Picl 1 P. : isj (mod 3)). Then R;(S, T) = @ if 2azk but iy(S, T) # 6 if 2a < k. C. Selecting Support Clauses We now give another approach. Theorem 3. Suppose there is a length n refu- tation from S and T = IT ,...,T 1 is a support class for S. Then there'exist tlauses Ci E Ti, l<i<k, such that -- a) {Cl,C2,...,Ck] C R co (S, {CC,], CC,}, El . . . . {Ck}}) and L b) thzre is a length n refutation from R $1 (S, CCC,}, . . . . k.,H). Thus it is possible to select particular clauses from the support sets and use them to define R. There may be many sets R co (S, {{Cl), I"1 . . . , {C,}}) satisfying condition a) $bove, but they may be smaller than the connected components of R ~0 Pl (S, T). It is possible to construct examples z av- ing this property. Therefore it may help to use the above sets rather than R co (S, T) when search- ing for refutations. rg Another advantage is that it is possible to examine the clauses C. E T. in some heuristically determined order. Fu&herm;re, the 81 above approach is useful when the support sets Ti are not known in advance but are obtained one clause at a time. We now give a recursive procedure "re13" for generating all the sets as in Thereom 3. The idea is to order the Ti so as to reduce the branching factor as much as possible near the beginning of the search. Thus we use the principle of "least commitment." This procedure has as input sets Sl and S of clauses, integer n, and support class T Ri(S, {(D1),..., {Dj}, {Cl}, . . . . {Ck})) having ,...,Ck} as a subset, for Ci e Ti, ere is a length n refutation from S, and T is a support class for S, then there will be a length n refutation from some set output by rel3(lb, rtl, S, T). Definition. If S = {Dl, = {{D,) , . . . . {Dj% . . ..Dj) then Single(S) procedure rel3(S,,n,S,T) S2 -+ <(S, T 6 Single(S1)) if s1 C S2 then if T = fl then output (S2) else - T,+- TG I choose 'I$ e Tl minimizing I T21; for all C E T2 do od: re13(S1U {C),n,S2,Tl-T2) fi;-’ fi; end re13; By searching for such sets Sl, we can some- times obtain much better relevance criteria than by previous methods. The use of centers insures that elements of Sl will be closer together than in pre- vious methods. To implement this method, let S2 be Q p+2,@ T) . For each C E S2, let S3 be RrnT2 (S,{{b}}4).' -I 1-p If S3 intersects all support sets, then it is a candidate set of input clauses for a length n refu- tation. Here S, is a set of possible centers. Note that two clauses of S will have distance at most rtl + 1 in G(S3). No?e also that if n=6 then IF1 = 2 and if n = 10 then r?l = 3. Thus we can get somewhat nontrivial refutations with quite small distance bounds. E. Typing Variables For these relevance criteria to be useful, there must exist clauses C and C of S such that dC+ C2) is large. Howe&r, if z he axiom x=y 1 y=x is in S then two clauses of the form tl = t2 d D and t3 # t4 Tib v D2 will have distance 3 or less. is may cause everything to be close to everything else. To reduce this problem, we propose that all variables be typed as integer, Boolean, list, string, et cetera and unifications only succeed if the types match. Thus the above clauses would not necessarily be within distance 3 if tl and t4 or t2 and t3 have different types. The use of types may increase the number of clauses, since more than one copy of some clauses may be needed. However, the overall effect may still be beneficial. F. Logical Consequences D. Center Clauses By using the idea that graphs have "centers," we can reduce the distance needed to search for relevant clauses by another factor of 2. Theorem 4. Suppose there is a length n refu- tation from set S of clauses, and T is a support class for S. Then there exists a clause C E S and a set S 1 c S having the following properties: 1. ;: There is a length n refutation from S1 Sl is fully matched 4. Sl intersects all the support sets in T C E Sl and for all Cl E Sl, d(C, Cl) 2 +1 in G(Sl). Proof. Let Sl be the input clauses actually used insome minimal refutation from S. Then I s1 I Choose a "center" C of Sl, and note that The preceding ideas can also be applied to derivations of clauses other than NIL from S. Definition. A support set for S relative to C is a subset V of S such that C is not a logical consequence of S-V. A support class for S relative to C is a collection of support sets for S relative to c. For example, if I is an interpretation of S in which C is false, and V is the set of clauses of S that are false in I, then V is a support set for S relative to C. Definition. M(S, C) is the largest subset of S in which all literals are matched, except possi- bly those having literals of C as instances. Definition. Rn(s,T,c) is M(Q,(s,T) 23. Theorem 5. If there is a length n derivation of something subsuming C from S, and T is a support class for S relative to C, then there is a length n derivation of something subsuming C from R (S, T,O. $1 As before, we can introduce R O" (S,T,C) and other relevance criteria. 121 82 Plaisted, D. 5 G. Procedures scribed earlier. To incorporate procedural and heuristic in- formation, we may add clauses expressing the as- sertion A(x) 3 (sy)B(x,y) where A and B are input and output assertions for the procedure and x and y are input and output variables. To account for the fact that heuristics may fail, we assign pro- babilities of truth to clauses. The task then is to find a set S of clauses from which the desired consequence can'possibly be derived, subject to the condition that the product of the probabili- ties of the clauses in Sl is as large as possible. One way to do this is to run many trials, gene- rating relevant subsets of S, where the clauses of S are chosen to be present or absent with the appropriate probability. We then select a rele- vant set of clauses from among those clauses that have been found to be relevant in many of the trials. Note that if procedures are encoded as above, then a short proof may correspond to a so- lution using a few procedure calls, but each pro- cedure may require much time to execute. H. Subgoals If procedures are encoded as above, then each procedure may call the whole theorem prover recur- sively. This provides a possible subgoal mechan- ism. By storing the clauses from all subgoals in a common knowledge base, we may get interesting interactions between the subgoals. By noticing when subgoals are simpler than the original pro- blem in some well-founded ordering, we may be able to get mathematical induction in the system. The use of clauses, procedures, subgoals, and relevance criteria as indicated here provides a candidate for a general top-level control structure for an artificial intelligence system. V A COMPLETE STRATEGY The following procedure attempts to construct a refutation from set S of first-order clauses: procedure refute(s); for d = 1 step 1 until (NIL -- for j = 1 step 1 until (j refl(S, j, d) od od; -- is derived) do - > d) do - end refute; procedure refl(S, i, d); let T be a support class for S; RtR?(S, '0; if R 'is empty then return fi; ~Ru llevel 1 resolvents from R); if NIL E V or d = 1 then return fi; for j = 1 step 1 until(NIL is delved) do - refl(V, j, d - 1) &; end refl; This procedure selects at each step the clauses that seem most relevant and attempts to construct a refutation from them. Similar procedures can be given using other of the relevance functions de- A. Generating Support Sets One way to generate support sets for the above procedure is to let each support set be the subset of S in which specified predicate symbols occur with specified signs. This would yield 2n support sets for n predicate symbols. Of course, it is not necessary to use all of these support sets. A more interesting possibility is to have a collection {I,, 12, . . . , Sandtolet T. Ikl of interpretations of be the set of clauses that are false in I.. 'If I, has a finite domain then T. can be comiuted by ehaustive testing. Otherwisk, special methods may be necessary to determine if a clause C is true in I.. If I. has an infinite domain, a possible he?iristic 3s to let T. be the set of clauses that are false on some fiiite sub- set of the domain. If f is an abstraction mapping or a weak abstraction mapping [5] and I is an interpretation, then{CeS: some clause in f(C) is false in 11 is a support set for S. This approach may allow the use of nontrivial support sets which are easy to compute, especially if all elements of f(C) are ground clauses for all C in S. Note that T may include support sets obtained both syntac- tically and semantically. Although it may require much work to test if C is true in I., this kind of effort is of the kind that humans s&em to do when searching for proofs. Also, this provides a meaningful way of incorporating nontrivial seman- tic information into the theorem prover. The arcs in the connection graph resemble "associations" between facts, providing another similarity with human problem solving methods. REFERENCES [l] Gallaire, H. and J. Minker, eds. Logic and Data Bases. New York: Plenum Press, 1978. -- [2] Kowalski, R., "A Proof Procedure using Connection Graphs". J-ACM 22(1975)572-595. -- [3] Reingold, E. M., J. Nievergelt, and N. Deo. Combinatorial Algorithms: Theory and Practice. 7 Englewood Cliffs, New Jersey: Prentice-Hall, 1977. [4] Paterson, M. S. and M. N. Wegman, "Linear Unification", IBM Research Report 5304, IBM, 1976. [5] Plaisted, D., "Theorem Proving with Abstrac- tion, Part I", Departmental Report UIUCDCS-R- 79-961, University of Illinois, February 1979. 83

1980

7

67

This paper presents the results of research done on the representation of control knowledge in rule-based expert systems.’ It discusses the problems of representing co&o1 knowledge implicitly in object-level inference rules and presents specific examples from a MYCIN-like consultation system called PUFF. As an alternative, the explicit representation of conerol knowledge in sloes of a frame-like data structure is demonstrated in the CENTAUR system. Explicit representation of control knowledge has significant advantages both for the acquisition and modification of domain knowledge and for explanations of how knowledge is used in the expert system. REPRESENTATION OF CONTROL KNOWLEDGE IN EXPERT SYSTEMS Janice S. Aikins Computer Science Department Stanford University Stanford, California 94305 ABSTRACT I INTRODUCTION This paper emphasizes the importance of representing domain-specific control knowledge explicitly and separately from other forms of domain knowledge in expert systems. The particular focus of research on this topic has been MYCIN-like consultaeion systems [S] which represent their domain knowledge in the form of condition-action or production rules. Examples in this paper are taken from the PUFF system [43 which performs consultations in the domain of pulmonary (lung) physiology. The CENTAUR system was created in response to several knowledge representation and control structure problems in the rule-based systems, among which were the problems caused by ehe implicit representation of control knowledge. CENTAUR provides a framework for performing tasks using an hypothesize and match approach [5] to problem solving. This approach focuses the search for new information around recognized patterns of knowledge in the domain, a strategy that was not represented in the rule-based systems. Knowledge in CENTAUR is represented in the form of frame-like structures, called prototypes, which represent the expected patterns of knowledge, and in production rules, which serve as a stylized form of procedural attachment and are used to infer values or “fill in” slots in the prototype. This knowledge of prototypical situations is used for control of the consultation, for explanation of system performance, and also as a guide for acquiring additional knowledge and for modifying the existing knowledge base. ’ This work was supported by the Advanced Research Projects Agency under contract MDA 903-77-C-0322. Computer facilities were provided by the SUMEX-AIM facility at Stanford University under National Institutes of Health grant RR-00785-07. The author is sponsored by the Xerox Corporation under the direction of the Xerox Palo Alto Research Center. CENTAUR’s combination of prototypes and rules results in a knowledge representation that is expressive enough to allow the many kinds of domain knowledge necessary for system performance to be explicitly represented. Control know/edge for the consultation is represented in slots associated with each prototype, separately from the inference rules. Rules are associated with prototypes as the explicit contexts in which the rules are applied. The slots in the prototype specify the function of the attached rules, such as to summarize data already given or to refme an interim diagnosis. Other details of the CENTAUR system and a full discussion of the knowledge representation and control structure problems in the rule-based systems can be found in [I]. II TME PUFF SYSTEM One such rule-based system is the PUFF system which was created using a MYCIN-like framework. PUFF’s domain- specific knowledge is represented by a set of approximately 60 production rules. The “IF” part of the productlon states a set of conditions (the premise clauses) in which the rule is applicable. The action, or “THEN” part of the production, states the appropriate conclusions. The goal in PUFF is to interpret a set of lung function tests performed on a patient, and to produce a diagnosis of pulmonary disease in that patient. Each rule clause is a LISP predicate acting on associative (object- attribute-value) triples in the data base. In PUFF there is a single object, the patient. The attributes (or clinical parameters) are the lung function tests and other information about the patient. The PUFF control structure is primarily a goal-directed, backward chaining of the production rules as it attempts to determine a value for a given clinical parameter. written. A complete description of this mechanism is given in E61. III IMPLICIT CONTROL IN THE RULES P’roduction rules, in theory, are modular pieces of knowledge, each one capturing some “chunk” of domain- specific expertise. Indeed, one of the advantages of using production rules [3] is that there need be no direct interaction of one rule with the others, a characteristic which facilitates adding rules to the knowledge base or modifying existing rules. In practice, however, there are significant interactions among rules. Executing one rule will in turn cause others to be tried when the information needed for the first rule is not already known. Therefore, the order of the premise clauses of a rule affects the order in which other rules are executed. Further, in an interactive system such as PUFF, in which the user is asked for information that can not be inferred by rules, the order of the premise clauses also determines the order in which questions are asked. 121 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. This means of controlling question order by placing premise clauses in a specific order is, in fact, exploited by experts who recognize the need for ordering the questions that are asked, but have only this implicit and indirect mechanism for achieving their goal. In this case, the production rule framework itself becomes a programming language where the rules have multiple functions; some rules represent independent chunks of expertise with premise clauses specified in an arbitrary order, while others serve a controlling function with premise clauses that cannot be permuted without altering the behavior of the system. An example of implicit control knowledge is illustrated by the PUFF rule in Figure I below. This rule invokes other rules in an attempt to determine whether there is Obstructive Airways Disease. (Clause One), and if so, to determine the subtype (Clause Two) and findings associated with the disease (Clause Three). If Clause One were inadvertently placed after either Clause Two or Three, the system’s questions of the user would probe for more detailed information about Obstructive Airways Disease without having confirmed that the disease is present. For example, by reordering the clauses in RULEOOP, PUFF might begin its consultation by asking about the patient’s smoking history, one of the findings associated with Obstructive Airways Disease, and a question that would be inappropriate in a patient without a smoking-related disease. However, this rule contains no explicit indication that the order of the clauses is critical. The problem with implicit representation of control knowledge becomes apparent in working with the knowledge base, either to modify the knowledge or to explain its use in the system. RULE882 ----B-s If: 1) 2) 3) An attempt has been made to deduce the degree of obstructive airways disease of the patient, An attempt has been made to deduce the subtype of obstructive airways disease, and An attempt has been made to deduce the findings about the diagnosis of obstructive airways disease Then : It is definite (1.8) that there is an interpretation of potential obstructive airways disease FIGURE 1. PUFF Rule--Implicit Control Modifying rules is a normal part of system development. Clauses often must be added to or removed from rules in response to perceived errors or omissions in the performance of the system. However, removing or modifying the clauses of a controlling rule can alter the system’s behavior in unexpected ways, since the implicit control knowledge also will be altered. Therefore, modifications can be safely done only by persons intimately familiar with the knowledge base. This factor not only limits the set of people who can make modifications, and of course precludes the success of automatic knowledge acquisition systems in which each rule is considered individually, but it also limits the size of the knowledge base, as even the best of knowledge engineers can retain familiarity with only a limited number of rules at a time. A system’s explanations of its own performance also suffer when information critical to performance is not represented explicitly. The rule-based systems studied generate explanations of why questions are being asked using direct translations of those rules which were being used when the question was asked. (See [2] for details.) There is no distinction made between rules that control a line of reasoning, as opposed to rules that infer a piece of information. However, users of the system should be able to ask both kinds of questions in order to obtain justifications of the system’s reasoning process as well as justifications of its inference rules. The uniform representation of control and inference knowledge in rule-based systems further confuses the user by mixing the two kinds of explanations. IV CONTROL KNOWLEDGE IN CENTAUR Control knowledge about the process of pursuing an hypothesis in CENTAUR is represented in slots associated with each prototype, separate from the inference knowledge which will actually confirm or deny the hypothesis represented as production rules. Each slot specifies one or more LISP clauses, or control tasks, that are executed at specific points during the consultation as defined by a top-level prototype representing the “typical” consultation (the CONSULTATlON Prototype). For the pulmonary function domain, prototypes correspond to specific pulmonary diseases. During a CENTAUR consultation, initial case data suggest one or more disease prototypes as likely matches. Control knowledge in these prototypes then guides the consultation by specifying what information should be sought next. Expected data values in each prototype enable CENTAUR to pinpoint inconsistent or erroneous information during the consultation. Final conclusions are presented in terms of the prototypical situations determined to be present in the case, and any inconsistencies are noted. Thus the system developer can specify “what to do” in a given prototype context as an important part of the knowledge about the domain that is distinct from the inferential knowledge used in the consultation. These control tasks are specified as LISP functions, and the system developer can define any new functions as they are required. For example, Figure 2 shows CENTAUR’s representation of the control knowledge in the PUFF rule shown in Figure I. The control knowledge is represented in two of the control slots associated with the Obstructive Airways D’isease (OAD) prototype. They specify that when OAD is confirmed (the If-Confirmed Slot), the next tasks are to deduce a degree and a subtype for OAD, and, at a later stage in the consultation (when the prototype ACTION slots are executed), to deduce and print findings associated with OAD. If-Confirmed Slot: Deduce the Degree of OAD Deduce the Subtype of OAD. Action Slot: Deduce any Findings associated with OAD Print the Findings associated with OAD FIGURE 2. OAD Prototype Control Slots 122 Prototypes not only represent the domain-specific knowledge of a particular application, but also represent domain-independent knowledge about the operation of the CENTAUR system. At the highest level in CENTAUR, the Consultation Prototype lists the various stages of the consultation (e.g., entering initial information, suggesting likely prototypes, filling in prototypes) in its control slots. The advantages of explicit representation of control knowledge thus extend to control of the consultation process itself. V ADVANTAGES OF THE CENTAUR APPROACH The association of control knowledge with individual prototypes allows control to be specific to the prototype being explored. Thus domain experts can specify a different set of control tasks for each prototypical situation. In the pulmonary domain, for example, the expert proceeds in a different way if he has confirmed OAD rather than some other disease in the patient. Further, because this control knowledge is separate from the inference rules, the expert does not have to anticipate and correct incidental interactions between control and inference knowledge. Representing the entire consultation process itself as a prototype has additional advantages. First, the system designer’s conception of the consultation process is clearly defined for all system users. Second, representing each stage of the consultation as a separate control task allows stages to be added or removed from the consultation process. For example, the Refinement Stage, which uses additional expertise to improve upon an interim conclusion, was omitted during early stages of system development for the pulmonary function problem. “Filling in” a consultation prototype with user- specified options, such as a choice of strategy for choosing the current best prototype (for example, confirmation, elimination, or fixed-order), results in a control structure that can be tailored to the desires of each individual user. The organization of knowledge into prototypical situations allows the user to more easily identify the affected set of knowledge when changes to the knowledge base are desired. Points at which specific control knowledge is used during the consultation are clearly defined, with the result that it is easier to predict the effects of any control modifications that may be made. Explicit representation of control knowledge also facilitates explanations about that knowledge. In addition to the HOW and WHY keywords available in MYCIN, a new keyword, CONTROL, has been defined so that a user of the system can inquire about the control task motivating the current line of reasoning. For example, if the user types “CONTROL” in response to a system question about the patient’s smoking history, the system would respond, The current control task Is to determine the findings associated with OAD. VI SUMMARY This paper has discussed the importance of representing control knowledge explicitly, particularly as it affects knowledge acquisition and explanation in a knowledge-based system. The representation of control knowledge as slots in a prototype in the CENTAUR system demonstrates one feasible approach. Augmenting the rule representation to include rules that function exclusively as control rules might be another. The critical lesson learned from working with the rule-based systems is that the system’s representation structures must be expressive enough to represent control knowledge explicitly, so that it will not be inaccessible to the system and to the knowledge engineer. ACKNOWLEDGMENTS Many thanks to Doug Aikins, Avron Barr, Jim Bennett, Bruce Buchanan, and Bill Clancey for their helpful advice and comments on earlier versions of this paper. REFERENCES [l] Aikins, J. Prototypes and Production W/es: A Knowledge Representation for Computer Consultations. (Forthcoming Ph. D. Thesis), Heuristic Programming Project, Dept. of Computer Science, Stanford University, 1980. [Z] Davis R. Applications of Meta Level Knowledge to the Construct/on, Maintenance and Use of Large Knowledge Bases. STAN-CS-76-552, Stanford Universlty, July 1976. [3] Davis R., and King, J. An Overview of Production Systems. In E. W. Elcock and D. Michie (Eds.), Machine lntelllgence 8. New York: Wiley & Sons, 1977. Pp. SOO- 332. [4] Kunz, J., Fallat, R., McClung, D., Osborn, J., Votteri, B., Nii, H., Aikins, J., Fagan, L., and Feigenbaum, E. A Physiological Rule Based System for lnterpretlng Pulmonary Function Test Results. HPP-78-19 (Working Paper), Heuristic Programming Project, Dept. of Computer Science, Stanford University, December 1978. [5] Newell, A. Artificial Intelligence and the Concept of Mind. In R. Schank and K. Colby (Eds.), Computer Models of Thought and Language. San Francisco: W. H. Freeman and Company,l973. Pp. l-60. f6j Shortliffe, E. H. MYCIN: A Rule-based Computer Program for Advising Physicians Regarding Antimicrobial Therapy Selection. Ph. D. dissertation in Medical Information Sciences, Stanford University, 1974. (A Iso, Computer-Based Medical Consultations: MYCIN. New York: American-Elsevier, 1976. 123

1980

70

68

DELTA-MIN: A Search-Control Method for Information-Gathering Problems Jaime G. Carbonell Computer Science Department Carnegie-Mellon University Abstract The A-MIN method consists of a best-first backtracking algorithm applicable to a large class of information-gathering problems, such as most natural language analyzers, many speech understanding systems, and some forms of planning and automated knowledge acquisition. This paper focuses on the general A-MIN search-control method and characterizes the problem spaces to which it may apply. Essentially, A-MIN provides a best-first search mechanism over the space of alternate interpretations of an input sequence, where the interpreter is assumed to be organized as a set of cooperating expert modules.’ 1. Introduction A present trend in Al is to design large systems as cooperating collections of experts, whose separate contributions must be integrated in the performance of a task. Examples of such systems include HEARSAY-II [4], POLITICS [l], PSI [5], SAM, [3]. The division of task responsibility and grain size of the experts differs markedly. For instance, the latter two systems contain few large-scale experts, which are invoked in a largely predetermined order, while the former two systems contain a larger number of smaller modules whose order of invocation is an integral part of problem solving task itself. In this paper I discuss a new search method, called A-MIN, that incorporates some of the desirable features from best-first search and some properties of gradient search. (Gradient search is locally-optimized hill-climbing.) The primary objective is to make global control decisions based on local knowledge provided by each expert module. No module is required to know either the internal structure of another module, or the overall controling search mechanism. In this way, I depart somewhat from the freerer blackboard control structure of HEARSAY-II, where search was controled by the experts themselves. The module that “shouted loudest” was given control, hence each module had to know when and how loud to shout with respect to other expert modules. In addition, there was a “focus knowledge source” [6] that helped guide forward search. This method acquires its flexibility by placing a substantial amount of global control responsibility on local experts. Moreover, it entails no externally-transparent search discipline. Finally, the primary emphasis is on forward search, not reconsidering wrong decisions in favor of choosing an alternate interpretation. In light of these considerations, I attempted to factor domain knowledge (what the experts know) from search discipline (when to pursue alternate 1 This research was sponsored in part by the (ONR) under grant number NO001 4-79-C-0661. Office of Naval Research paths suggested by different experts), so that each problem may be investigated in its own right. Here, I focus on the search control aspect, and consider the internal structure of each domain expert as a virtual “black box”. To simplify matters, I confine my discussion to tasks whose terminating condition is defined by processing an input sequence to completion without error. This class of problems is exemplified by natural language analysis, where an input sentence is processed left-to-right and the goal state is the formation of a consistent semantic representation of the input. Clearly, this is a satisficing rather than optimizing task [8], in the sense that only the first of potentially many solutions is sought. Since I want the language analysis to give the same parse of the sentence as a human would, the process must be biased to favor reaching the appropriate solution first. This biasing process, based on local decisions made by expert modules is the primary input to the A-MIN search method described below. It must be noted, however, that the left-to-right processing assumption is more restrictive than the HEARSAY paradigm, where “islands of interpretation” could grow anywhere in the input sequence and, when possible, were later merged into larger islands until the entire input sequence was covered [6, 41. 2. An Information-Gathering Search Space Consider a search space for the task of processing a finite sequence of input symbols (such as an English sentence) and producing an integrated representation incorporating all the information extracted from the input (such as a semantic representation of the meaning encoded in the sentence). The problem solver consists of a set of experts that may be applied at many different processing stages, without fixed constraints on their order of application. For instance, in the language analysis domain, one can can conceive of a verb-case expert, a morphological-transformation expert, an extra-sentential referent-identifier expert (or several such experts based on different knowledge sources), a dialog-context expert, an immediate-semantics expert, a syntactic-transformation expert, etc... A robust language analyzer must be capable of invoking any subset Of these and other experts according to dynamically determined needs in analyzing the sentence at hand. NOW, let US back off the natural language domain and consider the general class of problem information-gathering,* spaces to which an cooperating-expert approach appears useful. First, we draw a mapping between the general problem solving terminOlOgy and the expert module approach. The search space outlined below is a considerably constrained version of a general search space. This property is exploited in the A-MIN search method described in the fcllowing Section. 124 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. e The operators in the search space are the individual expert modules. Each module may search its own space internally, but I am concerned only with the macro-structure search space. Each expert has conditions of strict applicability and preference of applicability. The latter are used for conflict resolution decisions when more than one expert is applicable. o A state in the space consists of the total knowledge gathered by invoking the set of experts that caused the state transitions from the initial state to ihe present. This definition has two significant implications: An expert that adds no new knowledge when invoked does not generate a new state; therefore, it can be ignored by the search control. There is a monotonicity property in that each step away from the initial states adds information to the analysis, and therefore is guaranteed to “climb” to a potential final state. (Left-to-right, single-pass natural language analysis can exhibit such monotonic behavior.) e A final state is defined by having reached the end of the input sequence without violating a path constraint, and no expert can add more information (i.e., no transition in the state space is possible). e A path constraint is violated if either a new segment of the input cannot be incorporated or an expert asserts information that contradicts that which is already part of the current state. When this situation arises, directed backtracking becomes necessary. o The initial state is a (possibly empty) set of constraints that must be satisfied by any interpretation of the input sequence. For instance, a dialog or story context constrains the interpretation of an utterance in many natural language tasks. o Each expert can draw more than one conclusion when applied. Choosing the appropriate conclusion and minimizing backtracking on alternatives is where the real search problem lies. Choosing the next expert to apply is not a real problem, as the final interpretation is often independent of the order of application of experts. That is, since information-gathering is in principle additive, different application sequences of the same experts should converge to the same final state. The experts preselect themselves as to applicability. Selecting the expert who thinks it can add the most information (as in HEARSAY-II) only tends to shorten the path to the final state. The real search lies in considering alternate interpretations of the input, which can only be resolved by establishing consistency with later information gathered by other experts. Finally, given the possibility of focused backtracking fro& a dead end in the forward search, less effort needs to be directed at finding the “one and only correct expert” to apply. Search Metho A-MIN is a heuristic search method specifically tailored to the class of search spaces described above. It combines some of the more desirable features of gradient search and best-first search, with the modularized information sources of a cooperating-expert paradigm. Figure 3-O is the search-control strategy algorithm in an MLISP-style form. Subsequently I discuss how A-MIN works, exemplifying the discussion with a sample search-tree diagram. Every expert is responsible for assigning a likelihood-of-correctness value to each alternative in the interpretation it outputs. These values are only used to determine how much better the best alternative is than the next best alternatives, an item of information crucial to the backtrack control mechanism. There is no global evaluation function (the outputs of different experts are not directly comparable - it makes no sense to ask questions like: “Is this anaphoric referent specification better than that syntactic segmentation?“) Nor is there any mechanism to compute differences between the present state and the goal state. (Recall our definition of goal state -- only upon completion of the input processing can the goal state be established.) PROCEDURE A-MIN(initial-state. experts) altlist := NULL globaldelta := 0 state := initial-state input := READ(firsfinput) NEXTOP: IF NULL(input) THEN RETURN(state) ELSE operator := SELECTBEST(APPLICABLE(experts)) IF NULL(operator) THEN input : = READ( next input) ALSO GO NEXTOP ELSE alts := APPLY(operator, state) IF NULL(alts) THEN MARK(operator. 'NOT-APPLICABLE, 'TEMP) ALSO GO NEXTOP bestalt := SELECTMAX(alts) IF Ilaltsll > 1 THEN alts := FOR-EACH alt IN REMOVE(bestalt, alts) COLLECT <'ALT: alt, 'STATE: state, 'DELTA: globaldelta + VALUE(bestalt) - VALUE(alt)> altlist := APPEND(alts, altlist) NEWSTATE: state := MERGE-INFORMATION(state, bestalt) IF NOT(state = 'ERROR') GO NEXTOP ; if no error, continue gradient search, ; else delta-min backup below WHILE state has no viable alternatives DO BEGIN MARK(state, 'DEAD-END, 'PERM) ; delete dead ends state := PARENT(state) ; from search tree END backup-point := SELECTDELTA-MIN(altlist) state := GET(backup-point. 'STATE:) globaldelta := GET(backup-point, 'DELTA:) bestalt := GET(backup-point, 'ALT:) altlist := REMOVE(backup-point. altlist) GO NEWSTATE END A-MIN Figure 3-1: The A-MIN Search-Control Algorithm Let us see how A-MIN can be applied to an abstract example, following the diagram in figure 3-Z The roman numerals in the arcs reflect the order in which they are traversed. At the initial state, expert-4 applies and generates three alternate interpretations of the input. One alternative is ranked as most likely. A “A” value is computed for the remaining alternatives, 2 "lnformetion gathering" is a term coined by Rai Reddy to refer to encoding the difference in confidence that expert-4 had between search problems where progress towards agoalstateischaracterized by them and the most likely alternative. The more sure expert-4 is of accruing and integrating information from outside SOUrCeS. its best choice relative to the other alternatives, the larger the A 125 values. The best interpretation generated by expert-4 is integrated with the initial state constraints and found to be consistent. At this point, a new state has been generated and expert-2 applies to this state generating no new information. More input is read and expert-l applies, generating only one alternative, which is found to be consistent with the information in the present state. In a similar fashion, the rest of the tree in figure 3-xis generated. Up to now, we have witnessed an instance of gradient search, where a different evaluation function is applied at each node (The local evaluation function is, in effect, the expert who generated the likelihood values.) If no error occurs, (i.e., if the interpretation of the input remains consistent) no backup is needed. The likelihood rankings clearly minimize the chance of error as compared to straightforward depth first search. Now, let us consider the possibility of an inconsistency in the interpretation, as we continue to examine figure 3-X Figure 3-2: A-MIN Search Tree With Directed Backup The most likely interpretation generated by expert-6 was found to be inconsistent with the information in the present state. Backup is therefore necessary, given the depth-first nature of the search. But, where do we back up to? Normally, one might consider a depth-first unwinding of the search tree; but, is this the most reasonable strategy? Expert-5 was much less certain in its choice of best alternative than expert-6 (A = 2 vs A = 4). It seems more reasonable to doubt expert-5’s decision. Therefore, one wants to back up to the point where the probability of having chosen a wrong branch is highest, namely to the choice point with the minimal A (hence the name A-MIN). Continuing with figure 3-z we restore the state at expert-5 and incorporate the A = 2 interpretation. It is found to be consistent, and we apply expert-l to the new state. The best interpretation of expert-l leads to error, and backup is again required. Where to now? The minimal A is at expert-l, but this would mean choosing a non-optimal branch of a non-optimal branch. Lack of confidence in the choice from expert-5 should be propagated to the present invocation of expert-l. Hence, we add the two As in the path from the inital state and get the value: A =3, which is greater than the minimal A at expert-4 (A = 2). Therefore, we back up to expert-4. This process continues until a consistent interpretation of the entire input is found (i.e., a goal state is reached), or the search exhausts all viable alternate intepretations. Essentially, A-MIN is a method for finding one globally consistent interpretation of an input sequence processed in a predetermined order. In natural language analysis, the problem is to find a semantically, syntactically, and contextually consistent parse of a sentence. In speech understanding the constraint of formulating legal phonemes and words is added, but the nature of the problem and the applicability of the A-MIN approach remains the same. For instance, A-MIN is an alternate control structure to HARPY’s beam search [7], which also processes a sequence of symbols left to right, seeking a globally consistent interpretation. 4. Concluding Remarks To summarize, the A-MIN method exhibits the following properties: o A-MIN is equivalent to gradient search while no error occurs. Path length (from the initial state) is not a factor in the decision function. o The backtracking mechanism is directed to undo the choice most likely to have caused an interpretation error. This method compares all active nodes in the tree, as in best-first search, but only when an error occurs (unlike best-first search). o Perseverance in one search path is rewarded, as long as the interpretation remains consistent, while compounding less-than-optimal alternate choices is penalized. This behavior falls out of the way in which A values are accrued. o No global evaluation function forces direct comparisons among information gathered by different knowledge sources. Such an evaluation function would necessarily need to encode much of the information contained in the separate experts, thus defeating the purpose of a modular cooperating-expert approach. The A comparisons contrast only the differences between locally-optimal and locally-suboptimal decisions. These differences are computed by local experts, but the comparisons themselves are only between relative ratings on the desirability of alternate decisions. 126 Additional discussion of implementation, analysis, and details of the A-MIN search method may be found in [2], where an effective application of A-MIN is discussed for constraining search in a natural language processing task. 1. 2. 3. 4. 5. 6. 7. 8. References Carbonell, J. G., “POLITICS: An Experiment in Subjective Understanding and Integrated Reasoning,” in inside Computer Understanding: Five Programs Plus Miniatures, R. C. Schank and C. K. Riesbeck, eds., New Jersey: Erlbaum, 1980. Carbonell, J. G., “Search in a Non-Homogeneous Problem Space - The A-MIN Algorithm,” Tech. report, Dept. of Computer Science, Carnegie-Mellon University, 1980. Cullingford, R., Script Application: Computer Understanding of, Newspaper Stories, PhD dissertation, Yale University, Sept. 1977. Erman L. D. and Lesser, V. R., “HEARSAY-II: Tutorial Introduction & Retrospective View,” Tech. report, Dept. of Computer Science, Carnegie-Mellon University, May 1978. Green, C. C., “The Design of the PSI Program Synthesis System, ” Proceedings of the Second International Conference on Software Engineering, October 1976 , pp. 4-18. Hayes-Roth, F. and Lesser, V. R., “Focus of Attention in the Hearsay-II Speech Understanding System,” Proceedings of the Fifth international Joint Conference on Artificial Intelligence, 1977 , pp.. 27-35. Lowerre, B., “The HARPY Speech Recognition System,” Tech. report Computer Science Department, Carnegie-Mellon University, April 1976. Newell, A. and Simon, H. A., Human Problem Solving, New Jersey: Prentice-Hall, 1972. 127

1980

71

69

ON WAITING Arthur M. Farley Dept. of Computer and Information Science University of Oregon Eugene, Oregon ABSTRACT Waiting is the activity of maintaining selected aspects of a current situation over some period of time in order that certain goal-related actions can be performed in the future. Initial steps toward the formalization of notions relevant to waiting are presented. Conditions and information pertin- ent to waiting decisions are defined. Introduction Waiting is the activity of maintaining selected aspects of a current situation over some period of time in order that certain goal-related actions can be performed in the future. Those aspects of the current situation which are maintained serve as preconditions for the desired future actions. Waiting extends from a decision to maintain these preconditions until (attempted) execution of the goal-related actions (or a decision to forego them). The act of waiting is closely associated with the future actions; a problem solver is said to be "waiting to" (do) these actions. The primary function of waiting is to improve the efficiency with which problem solving plans are executed. Waiting attempts to overcome reestablisk ment of preconditions for desired, future actions. Not only may less effort be expended during problem solving, but performance can become more coherent. The problem solver can avoid frequent side- and back-tracking in multiple-goal situations. Waiting also allows for a certain degree of parallelism in problem solving. Waiting can be engaged in simul- taneously with other actions which do not destroy satisfaction of the preconditions being maintained. Waiting is an important and frequent problem solving activity in the real-world. By real-world, we mean an ongoing, continuing, schedule based con- text, within which cooperative, as well as compet- itive, efforts among groups of problem solving systems normally occur. Waiting minimally requires such a context to be effective. A decision to wait implies that other, as yet unsatisfied, precondi- tions of anticipated goal-related actions are expected to be met by means other than direct intervention by the waiting system. Waiting as a problem solving activity has been largely (if not totally) ignored by AI research to date. This is primarily due to the fact that real- world contexts as defined here have only recently been considered. This paper outlines initial steps toward formalisms within which issues of waiting can be addressed. The research represents exten- sions to a knowledge-based problem solving system previously described by the author [3,41. We briefly review important aspects of that system before describing straightforward extensions which aid our understanding of waiting. We conclude by discussing related research and suggesting future work. Knowledge-based Problem Solving The form of its representation of the environ- ment influences all other aspects of a problem solving system. Let a situation of the relevant environment at any point in time (past, present, or future) be represented by a situation state. A situation state consists of a finite set of propo- sitions which are true with respect to the specific environmental situations(s) which the state is said to represent. The current state represents the present environmental situation. A goal state is a situation state which the problem solving system desires the current state to satisfy (i.e. be con- sistent with). A problem exists for a problem solving system when the current state does not sat- isfy constraints specified by a goal state. Proh- lem solving refers to any activity undertaken in attempts to eliminate differences between current and goal states. A problem is solved when differ- ences between its goal state and the current state no longer exist. A knowledge-based problem solving system solves most problems by instantiation and execution of known general solution plans. A general solution plan describes a process which is capable of satis- fying a set of goal states from any of a set of current states. A general solution plan is repre- sented as a rooted, directed, labelled tree of plan states. The root of the tree is the goal state. Plan states of the tree are interconnected by dir- ected arcs, each labelled by an operator. An oper- ator is a description of an action (or process), represented as sets of add, delete, and precondition propositions [71. The operator labelling an arc is capable of transforming the plan state at the tail of the arc into the plan state at the head of the arc. The plan state at the tail satisfies precon- ditions of the operator, while the one at the head reflects the results of additions and deletions associated with the operator. The maximal directed path from any plan state ends at the goal state. A plan state is a situation state whose pro- positions have been partitioned into three compo- nents. For each proposition in the SELFOP compo- nent, the problem solving has one or more operators capable of satisfying the proposition. Furthermore, From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. the system normally expects (prefers) to satisfy a SELFOP proposition itself by executing one or the operators. For each proposition in the OTHEROP component, (costly) operators may exist allowing the system to satisfy the proposition itself, but the system normally expects the proposition to be satisfied by other problem solving systems in the environment. Finally, for each proposition in the NOOP component, neither the system itself nor any other problem solving system has control over sat- isfaction of the proposition (i.e., weather condi- tions). Though the structure of a general solution plan is made clear by its tree, representing the plan as a production system can facilitate its execution. A production system [61 is a collection of condition-action pairs called rules. In the pro- duction system representation of a general solution plan, each plan state serves as the condition part of a rule whose action part is the operator label- ling the arc leaving that plan state. In [41, the author describes how this representation can be useful in a selective approach to coordinating the execution of multiple plans. Rules from general solution plans associated with as yet unsatisfied goals are combined to form one production system. A rule classification scheme is defined which pro- vides for the avoidance of most inter-plan con- flicts while allowing for responsive, efficient problem solving. For example, a goal state is classified as ungS;ounded if it denies satisfaction of conditions from a (critical) rule of another plan (in a conjunctive goal set). Rules from plans for ungrounded goal states are not executed until conflicting plans have been completed. Waiting We are concerned here with the control of wait- ing during plan executions by knowledge-based pro- blem solving systems operating in real-world con- texts. We first consider the question: what situ- ations trigger consideration of waiting? A rule is self-satisfied if all propositions in its SELFOP component are satisfied. Whenever a self-satisfied rule exists during plan execution, a problem solv- ing system may consider waiting, maintaining those conditions of the rule it has satisfied until the rule can be executed. A rule whose conditions are completely satisfied, and thus can be executed during a current system cycle, is classified as ready. Ready rules can suggest waiting. The sys- tem may wait to execute one ready rule, selecting another which leaves the rule ready (through con- flict resolution). During plan execution, a rule corresponding to a plan state whose NOOP component is not satisfied is classified as irrelevant. Such a rule can only fire if uncontrollable aspects of the environment change favorably. An irrelevant rule which is otherwise satisfied is a contingent rule. Being self-satisfied, contingent rules suggest consider- ation of waiting. Any NOOP condition must be satis fied occasionally, otherwise a rule would be fan- tasy and not considered part of a problem solving system's executable plans. An important class of NOOP propositions are those which deal with time. In a general solution plan reflecting real-world time constraints (a scheduled plan), the NOOP component of a plan state will contain one or both of the propositions ISNOWBEFORE(t) and ISNOWAFTER(t), where t is a time specification. The truth of a time proposition is determined relative to the current time, considered to be an ever present aspect of the current state. Time propositions of plan states derive from time constraints placed upon goal state satisfaction. They are propogated to other states (rules) of a scheduled plan, with time parameters adjusted to reflect time estimates for traversed operators. In a scheduled plan, time propositions alone may determine the relevancy status of rules. Rules which are irrelevant solely due to unsatisfied time constraints have their NOOP components classified as early or late, depending upon which time propo- sition is not satisfied. A rule with early NOOP component, but which is otherwise ready, is classi- fied as imminent. Whenever an imminent rule exists during plan execution, the problem solving system may consider waiting while time passes until the rule becomes relevant and can be fired. One source of imminent rules are overestimations of time re- quirements for completed, prior operations from a scheduled plan. For example, you estimate a 30 minute drive to a shopping mall; but it only takes 15; you arrive before the mall opens. The rule by which you would enter the mall is imminent. Another situation which may prompt waiting is the existence of a dependent rule. A relevant, self-satisfied rule corresponding to a plan state whose OTHEROP component is not satisfied is ClaSSi- fied as dependent. Whenever a dependent rule exists during plan execution, the system may consider waiting until the expected (necessary) assistance arrives and the rule can be fired. Dependent rules can often arise during cooperative problem solving efforts. For example, you are painting a house with a friend and turn around, expecting him to hand you a new can of paint, but he hasn't finished opening it. The rule by which you would take the paint is dependent. Finally, rules which have only their SELFOP component satisfied are classified as needy. Needy rules arise frequently when coordinating with pub- lic, problem solving support systems, such as mass transportation. To board a bus, a problem solver must be at a bus stop (SELFOP), at the scheduled time (NOOP), with the bus being at the stop (OTHEROP) . If a problem solver arrives at a stop five minutes early, the rule by which it boards the bus is needy. 129 Given the existence of a self-satisfied rule, what information is pertinent to subsequent waiting decisions? With each self-satisfied rule, the pro- blem solving system can associate three values: a waiting period, a set of compatible rules, and a set of contending rules. The waiting period is an estimate of the length of time before a self- satisfied rule will become ready. A compatible rule is a ready rule which requires less (esti- mated) time than the waiting period, and would not destroy satisfaction of the self-satisfied rule's conditions. A contending rule is a ready rule from another scheduled plan which would destroy the other's self-satisfaction, requires more time than the waiting period, but will become classified late if it is not fired within the waiting period. Determining these two sets of rules would not require dramatic additional computational effort. They only contain ready rules, rules which the system always determines before selecting the next rule to execute. A simplest policy for waiting can be stated in terms of these two sets of rules. If there is a contending rule associated with a self-satisfied rule, do not wait; otherwise wait, firing a com- patible rule, if any exist. Though this may pro- duce effective behavior in many circumstances, a moment's thought suggests further considerations. A goal state is nearby to a self-satisfied rule if time estimates indicate that the system could satisfy the goal state and reestablish conditions of the self-satisfied rule within the expected waiting period. A more complex waiting policy could have a system elect to wait in the face of contending rules, especially when compatible rules and/or nearby goal states allow active waiting. Relative importances of goal states would influence such a policy. Finally, rather than an expected waiting period, a cumulative probability function representing the likelihood that a self-satisfied rule will be able to fire within a given period of time could add further sophistication to waiting policies. Conclusion Research on distributed problem solving has used waiting as a coordination primitive [ll. Coordination is realized in our model through ungrounded goal states and unsatisfied OTHEROP com- ponents. Postponing actions until others have completed their responsibilities may or may not result in waiting as defined here. Research on medical diagnosis has dealt with time duration propositions in rule conditions [2]. These suspend judgement until time lends its support, Again, time-based postponement differs from issues of waiting addressed here. In this paper we have specified formalisms which add waiting to the repertoire of capabilities of problem solving systems dealing with real-world contexts. An (incomplete) set of conditions under which waiting may be reasonably considered are defined, as are aspects of the subsequent waiting decision process. We are continuing investigation of waiting by considering a particular context -- that of satisfying a set of work and personal goals on a given day within an office and city setting. This context has received attention in recent research on plan formulation [51. It appears to be just as fruitful a source of ideas on plan execu- tion and the role of waiting. Meeting schedules and office hours, going shopping, and using public transportation are actions which require frequent, even planned, waiting. A simulation to evaluate various waiting policies is planned. r11 [21 r31 II41 [51 [61 [71 REFERENCES Corkill, D. 'Hierarchical planning in a distributed environment:, Proceedings IJCAI-79, Tokyo, 1979, p. 168-175. Fagan L. etal., "Representation of dynamic clinical knowledge", Proceedings IJCAI-79, Tokyo, 1979, p. 260-262 Farley, A.M. "The coordination of multiple goal satisfaction', Proceedings IJCAIS, MIT, 1977, p. 495. Farley, A.M. "Issues in knowledge-based problem solving" to appear in I.E.E.E. Trans- actions on Systems, Man, and Cybernetics, August, 1980. Hayes-Roth, B. and Hayes-Roth F. "A cog- nitive model of planning", Cogntive Science, 3 (19791, pp. 275-310. Newell, A. and Simon, H.A. Human Problem Solving, Prentice-Hall: Englewood Cliffs, NJ, 1972. Sacerdoti, E.D. "Planning in a hierarchy of abstraction spaces", Artificial Intelligence, 5, 1974, pp. 115-135. 130

1980

72

70

DOU&S E. &pelt Slar@rd University, Stanford, California SRI International, Menlo Park, California ABSTRACT This paper reports recent results of research on planning systems that have the ability to deal with multiple agents and to reason about their knowledge and the actions they perform. The planner uses a knowledge representation based on the possible worlds semantics axiomati7ation of knowledge, belief and action advocated by Moore [5]. This work has been motivated by the need for such capabilities in natural language processing systems that will plan speech acts and natural language utterances [1, 21. The sophisticated use of natural language requires reasoning about other agents, what they might do and what they believe, and therefore provides a suitable domain for planning to achieve goals involving belief. This paper does not directly address issues of language per se, but focuses on the problem-solving rcquircmcnts of a language-using system, and describes a working system, KAMP (Knowledge And Modalities Planner), that embodies the ideas reported herein. I. WI-IAT,A KNOWLEDGE PLANNER MUST DO Consider the following problem: A robot named Rob and a man named John arc in a room that is adjacent lo a hallway containing a clock. Both Rob and John are capable of moving, reading clocks, and talking to each other, and they each know that the other is capable of performing these actions. They both know that they are in the room, and they both know where the hallway is. Neither Rob nor John knows what time it is. Suppose that Rob knows that the clock is in the hal& but John does not. Suppose further that John wants to know what time it is, and Rob knows he does. Furthermore, Rob is helpful, and wants to do what he can to insure that John achieves his goal. Rob’s planning system must come up with a plan, perhaps involving actions by both Rob and John, that will result in John knowing what time it is. We would like to see Rob devise a plan that consists of a choice between two alternatives. First, if John could find out where the clock was, he could go to the clock and read it, and in the resulting state would know the time. So, Rob might tell John wh&re the clock was, reasoning that this information is sufficient for John to form and execute a plan that would achieve his goal. The second alternative is for Rob to move into the hall and read the clock himself, move back into the room, and tell John the time. This research was supported by ihe Defense Advanced Research Projeers Agency under contract NOOO39-79-C-0118 with the Naval Electronic Systems Command The views and conclusions contained in this document are those of the aitthor and should not be interpreted as represenlative of the oficial policies, eilher expressed or implied, of the Defense Advanced Research Projects Agency, or the I/. S. Government. Existing planning mechanisms such as NOAII [6] or SrI:It% [3] arc incapable of dealing with this sort of problem. First, to solve this problem a system must reason cffcctivcly about propositional attitudes such as knorv, believe, and want. Fxisring planning syctcms arc based on knowledge rcprcscntations that arc not adequate for that purpose. Morcovcr, they arc not equipped to handle the integration of the actions of multiple agents into a single plan. In lhc solution to the above problem, the first choice consists of actions by both John and Rob. Rob does an informing act, pnd John moves into the hall and reads the clock. This means that Rob has planned for events to occur that are beyond its immediate control and which involve knowledge about the capabiliticc of another agent. The KAMP system solves problems such as the cxamplc above. It adopts a knowledge representation based on possible worlds semantics, which is capable of representing the knowledge needed for the task. By reasoning about the knowledge and wants of other agents, KAMP dctermincs what courses of action other agents can be expected to take in the filturc, and incorporates those actions into its own plans. II. REPRESENTING KNOWT,EDGE ABOUT BELIEF It is important that a planner be based on a knowledge representation that is adcquatc for representing different kinds of facts about knowing, wanting, believing, etc. For example, it may be necessary to reprcscnt that someone knows the value of a term, without the system itself knowing what that vnluc is, or the system may need to rcprcscnt that a person knows -P as opposed to not knowing whether P. A variety of strategies have been suggcstcd for representing such knowledge, but just representing the knowledge is not sufficient. It is also necessary that the system be able to reason with the knowledge efficiently. Many of the alternatives that have been proposed for rcprcscnting knowledge are fundamentally lacking in cithcr rcprcscntational adequacy or efftcicncy. Moore [5] discusses some of the specific proposals and their shortcomings. The representation which has been selected For the KAMP system is based on Moore’s axiomatization of possible worlds semantics. This approach has a great deal of power to represent and reason ‘efficiently with modal operators, and it is particularly elegant in describing the relation between action and knowledge. Because the design of the planner is largely motivated by the design of the knowledge representation, I will briefly outline Moore’s strategy for representing knowlcdgc about belief and how it relates to action. For comparison with a system that uses a different knowledge representation for planning to influence belief, see Konolige and Nilssun [4]. 131 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. The representation consists of a modal object language that has operators s~uz11 as believe and knovv. This object langauge is translated into a m&a-language that is based on a first order axiomitization of the possible worlds semantics of the modal logic. All the planning and deduction takes place at the lcvcl of the mcta language. In this paper I will adopt the convention of writing the meta language translations of object language terms and predicates in boldface. For example, to represent the fact that John knows P, one asserts that P is true in every possible world that is compatible with John’s knowledge. If K(A, wl, w2) is a predicate that means that w2 is a possible world which is compatablc with what 4 knows in wl, and T(w, P) means that P is true in the possible world M’, and Wo is the actual world, then the statement that John knows P is represented by the formula: (I) VW K(John, W,, w) 1 T(w, P), This states that in any world which is compatablc with what John knows in the actual world, P is true. Just as knowledge defines a relation on possible worlds, actions have a similar effect. The predicate CV RWo(A PI> y. w2) rcprcsents the fact that world w2 is rclatcd to world wl by agent A performing action P in wl. Thus possible worlds can bc used in a manner similar to state variables in a state calculus. Using a combination of the K and R predicates, it is possible to develop simple, elegant axiom schemata which clearly state the relationship between an action and how it effects knowledge. For example, it would bc possible, to jxiomatize an informing action with two simple axiom schemata as follows: (3) VW, Vw2 Va Vb VP R(Do(a, Inform(t), P)), wI, w2) 3 [t/w, Ma, wl, w3) 3 T(w3, 01 (4) VW, Vw2 ‘da Vb VP R(Do(a, Inform(b, P), wl, w2) 3 VW, W, w2, w3> 3 3QC(b, wl, w4> ii R(Do(a, Inform(b, P)), w4, w+] Axiom (3) is a precondition axiom. It says that it is true in all possible worlds (i.e. that it is universally known) that when someone does an informing action, he must know that what he is informing is in fact the case. Axiom (4) says that it is universally known that in the situation resulting from an informing act, the hearer knows that the inform has taken place. If the hearer knows that an inform has taken place, then according to axiom (3) he knows that the speaker knew P was true. From making aat deduction, the hearer also knows P. III. USING THE POSSIBLE WORLDS KNOWLEDGE REPRESENTATION IN PIANNING The possible worlds approach to representing facts about knowledge has many advantages, but it presents some problems for the design of a planning system. Axiomatizing knowledge as accessiblity relations between possible worlds makes it possible for a first order logic deduction system to reason about knowledge, but it carries the price of forcing the planner to deal with infinite sets of possible worlds. Planning for someone to know something means making a proposition true in an infinite number .of possible worlds. The goal wff, instead of being a ground level proposition, as is the case in other planning systems developed to date, becomes an expression with an implication and ‘I universally quantified variable, similar to (1). Another problem arises from the way in which actions are axiomatized. l’fficicnt axioms for deduction require the assumption that people can reason with their knowledge. ‘I’hc cffccts of SOITIC actions manifest themselves through this reasoning process. For example, this occurs in the axiomatization of IWORM giVCn in (3) and (4). Speakers usually execute informing actions to get someone to know something. However, the hearer dots not know what the speaker is informing directly as a result of the inform, but from realizing that the speaker has performed an informing act. (See Searlc [7]) ‘I’hc only effect WC know of for an informing act is that the hcarcr knows that the spcakcr has pcrformcd the act. When KAMP has a goal of somconc knowing something, it must be able to dctcrmine somehow that INI:OKM is a rcasonablc thing to do, even though that is not obvious from the effects of the action. A related problem is how to allow for the possiblity that people can reason with their knowlcdgc. If the system has the goal Know(A, Q), and this goal is not achicvcablc directly, and the system knows that A knows that P > Q, then Know(A, P) should be generated as a subgoal. IIxamples of this sort of situation occur whenever Q is some proposition that is not directly obscrvablc. If P is some observable proposition that entails Q, then the planner can perform some action that will result in knowing P, from which Q can be inferred. This is the basis of planning cxpcriments. Since there is a tradeoff between being able to take full advantage of the logical power of the formalism and being able to efficiently construct a plan, a strategy has been adopted that attempts to strike a balance. The strategy is to have a module propose actions for the planner to incorporate into a plan to achieve the current goal. This module can be thought of as a “plausible move generator”. It proposes actions that are likely to succeed in achicvcing the goal. The system then uses its deduction component to verify that the suggested action actually does achieve the goal. To facilitate the action generator’s search for reasonable actions, the preconditions and effects of actions are collected into SrRrPs-like acfiott summaries. Thcsc summaries highlight the direct and indirect effects of actions that are most likely to be needed in forming a plan. For example, the action summary of the INFORM action would include the hcarcr knowing P as one of the (indirect) effects of 1NI;ORM. The effects of actions as they are represented in the action summaries can be any well formed formula. So it is possible to state effects as implications which can match the implications that arc the meta language translations of the Know operator. Using action summaries is not cquivalcnt to recasting the knowledge rcprescntation as a SrRrPS-like system. The action summaries are only used to suggesf alternatives to be tried. To allow the possibility of agents reasoning with their knowledge, KAMP follows the following process whenever a goal involving a knowlcdgc or belief state is encountered: The system tries to invoke an operator that will achieve the goal directly. If this fails, then the system’s base of consequent rules is examined to find subgoals that could be achieved and that would allow the agent to deduce the desired conclusion. Although this approach has the advantage of simplicity, the number of subgoals that the planner has to consider at each step can grow exponentially. It seems intuitively correct to consider subgoals in the order of the length of inference that the agent has to go through to reach the desired conclusion. The planner needs good criteria to prune or at 132 least postpone consideration of less likely paths is an interesting area for further research. of inference. ’ I‘his IV. UNIVERSAL KNOWLEDGE PRECONI)ITIONS When planning actions on a strictly physical level, there are few if any preconditions which can bc said to apply universally to all actions. However, when dealing with the knowledge required to perform an action, as well as its physical enabling conditions, there are a sufficient number of interesting universal preconditions so that their treatment by the planner as a special case is warranted. Universal knowlege preconditions can be summarized by the statement tl?at an agent has to have an executable description of a procedure- in order to do anything. For example, if an agent wishes to perform an INFORM action, it is necessary for him to know what it is that he is informing, how to do informing, and who the intended hearer’ is. Since thcsc preconditions apply ro all actions, instead of including them in the axioms for each action, the planner automatically sets them up as subgoals in every case. V. THE REPRESk:?\‘l-ATION OF THE PLAN WITHIN I’HE SYSTEM The KAMP planner uses a system of plan representation similar to that of Sacerdoti’s procedural networks [6]. The major difference is that CHOICE nodes (OR-SPL.H'S) arc factored out into a disjunctive normal form. Since a choice may occur within an AND- SPLIT that affects how identical goals and actions are cxpandcd after the choice, all the choices arc factored out to the top level, and each choice is treated as a spearate alternative plan to be evaluated independently, (but perhaps sharing some subplans as “subroutines” with other branches.) As in Sacerdoti’s system, nodes can specify goals to be achieved, actions to bc performed, or they may be “ph;lntoms”, goals which coincidentally happen to already be satisfied. Each node of the procedural network is associated with a world. This world represents the real world at that particular stage in the execution of the plan. At the beginning of the plan, the first node is associated with WO. If the first action is that the robot moves from A to B, then the node rcprcscnting that action would have an associated world W1 and there would be an assertion of the form R(Do(Robot, Movc(A, B)), WO, WL) added to the data base. VI. CONTROL STRUCTURE The planner’s control structure is similar to that of Sacerdoti’s NOAH system. A procedural network is created out of the initial goal. The planner then attempts to assign worlds to each node of the procedural network as follows: First, the initial node is assigned Wo, the initial actual world. Then intcratively, when the planer proposes that a subsequent action is performed in a world to reach a new world, a name is generated for the new world, and an R relation between the original world and the new world is asserted in the dcducer’s data base. Then all goal nodes that have worlds assigned are evaluated, i.e. the planner attempts to prove that the goal is true using the, world assigned to that node as the current state of the actual world. Any goal for which the proof succeeds is marked as a phantom (achicvcd) goal. Next, all the unexpanded nodes in the network that have been assigned worlds, and which are not phantoms, are examined. Some of them may bc high level actions for which a procedure exists to determine the appropriate expansion. These procedures are invoked if they exist, otherwise the node is an unstaisfied goal node, and the action generator is invoked to find a set of actions 133 PI PI [31 [41 151 Kl [71 REFERENCES Appclt, Douglas E., Problem Solving Applied to Nutural Larlgrrage Generafion, proceedings of the Annual Confcrcnce of the Association for Computational Linguistics, 1980. Cohen, Philip, On Knowing Whal 10 Say: Planning Speech Acts, University of Toronto Technical Report #LB, 1978. Fikes, Richard E., and N. Nilsson, STRIPS: A New Approach to the Application of Theorem Proving to Problem Solving, Artificial Intelligence, No. 2, 1971. Konoligc, Kurt and N. Nilsson, Planning in n Mulfiple Agent Environment, Proceedings of the First Annual Conference of the American Association for Artificial Intelligence, August, 1980. Moore, Robert C., Reasoning About Knowledge and Action, Massachusetts Jnstitute of Technology Artificial Intelligence Laboratory Technical Report ‘I’R-???, 1979. Saccrdoti, Earl, A Structure for Plans and 13chaviol; Elscvier --- North-Holland, Inc., Amsterdam, The Netherlands, 1977. Scarle, John, Sgccch Acts, Cambridge University Press, 1969. which might be performed to achieve the goal. If an action is found, it is inscrtcd into the procedural network along with its preconditions, both the universal ones and those specific to the particular action. After the nodes are expanded to the next level, a set of critic procedures are invoked, which can examine the plan for global interactions and take corrective action if needed. This cntirc process is rcpeatcd until the plan has been expanded to the point where every unexpanded node is either a phantom goal or an executable action. VII. CURRENT STATUS OF RESKARCH The KAMP planning system has been implemented and tested on several examples. It has been used to solve the problem of John, Rob and the clock cited earlier in the paper. All the critics of Sacerdoti’s NOM1 have either been implemented or are currently uncicrgoing implcmcntation. Further development of the planner will be dictated by the needs of a language planning and generation system currently under development. It is expected that this langauge generation task will make full use of the unique features of this system. KAMP is a first attempt at dcvcloping a planner that is capable of using the possible worlds semantics approach to representing knowledge about belief. Combining a planner with a very powerful knowledge representation will enable problem solving tcchniqucs to bc applied to a variety of domains such as language generation, and planning to acquire and distribute knowledge, in which they have played a relatively small role in the past. ACKNOWLEDGEMENTS The author is grateful to Barbara Cross, Gary Hendrix and Terry Winograd for comments on earlier drafts of this paper

1980

73

71

Making Judgments Hans J. Berliner Computer Science Department Carnegie-Mellon University Pittsburgh, Pa. 15213 Abstract Reasoning-based problem solving deals with discrete entities and manipulates these to derive new entities or produce branching behavior in order to discover a solution. This paradigm has some basic difficulties when applied to certain types of problems. Properly constructed arithmetic functions, such as those using our SNAC principles, can do such problems very well. SNAC constructions have considerable generality and robustness, and thus tend to outperform hand coded case statements as domains get larger. We show how a SNAC fimction can avoid getting stuck on a sub-optimal hill while hill-climbing. A clever move made by our backgammon program in defeating the World Champion is analyzed to show some aspects of the method. 1 Int reduction Problem solving research and examples usually deal with sequential reasoning toward a conclusion or required response. For such situations, criteria exist that make it possible to identify the correct response and possibly order other responses with respect to their goodness. However, in most domains such a paradigm is not possible because the number of states in the domain is so large that it is next to impossible to describe the properties of an arbitrary state with sufficient accuracy to be able to reason about it. Expertise in such domains appears to require judgment. We consider judgment to be the ability to produce graded responses to small changes in the stimulus environment. In judgment domains several responses may be considered adequate, while reasoned decisions would appear to only be correct or incorrect. The ability to reliably judge small differences in chess positions is what separates the top players from their nearest competitors. Even though a decision procedure exists for determining whether one position is better than another, it is intractable. It is this intractability or the inability to isolate features that can be used in a clear reasoning process that distinguishes the judgment domain from the reasoning domain. The boundary between the two is certainly fuzzy, and undoubtedly changes as new information about any particular domain is developed. It seems that the larger the domain and the less precise the methods of making comparisons between elements of the domain, the less adequate are reasoning techniques. 2 The Problem There are a number of techniques available to allow a program to make comparisons, i.e. to discriminate good from bad from indifferent in selecting among courses of action and among potential outcomes. However, while these techniques are fine for doing simple comparisons, most of them break down with even small additional complexity. Consider the syllogism: 1) The more friends a person has, the happier he is. 2) John has more friends than Fred. Therefore: John is happier than Fred. So far so good. However, adding just a small amount of complexity with the two additional propositions: 3) The more money a person has, the happier he is. 4) Fred has more money than John. makes it possible to derive two contradictory conclusions from the premises. This is a most unsatisfactory state of affairs. Especially so, since recoding the premises into first order predicate calculus does not help either. Neither will using productions or the branching logic of programming languages. For such rcprcsentations, the most likely formulation would be that X will be happier than Y #he is superior in This research was sponsored by the Defense Advanced Research Projects Agency (DOD), ARPA Order No. 3597, monitored by the Air Force Avionics Laboratory Under Contract F33815-78-C-1551. 134 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. all applicable categories. Another formulation would have X happier than Y if he is superior in a majority of categories (with a tie being undefined). Such “voting” techniques can be shown to be deficient if We further increase the complexity of the decision that is to be made. If premises 2 and 4 were restated as: 2a) John has 25 friends and Fred has 20. 4a) Fred has $20,000 and John has $500. Most people would agree that Fred was happier according to our definitions of happiness. Yet, the only machinery available for coming to grips with problems such as this in systems that reason is to produce a large number of additional axioms that contain compound conditions, or to define degrees of difference so that degrees of happiness can be ascertained and summed. The world of reasoning is a world of quantized perception and action. These systems are discrete and do business on a “case” basis. In order to achieve expertise it is necessary to react differentially to states that were formerly considered equivalent. Thus, the number of distinct perceptions and actions gets larger with expertise. This makes it more expensive to find the applicable rule or pattern, and creates difficulty in keeping new rules from interfering in unintended ways with the effects of older rules. Further, the possibility that more than one pattern will match grows as complexity grows, and default conditions are usually defined for states that fail to match specific rules or patterns. This makes adding new knowledge a formidable task for even moderate size domains [3]. So unless, a method is found for automatically appending viable rules to such a system, there seems to be a definite limit on the expertise it can achieve. Because it is easier to pay attention to only a few things at one time, reasoning systems seem to have more of a sub-optimization nature than is necessary in sequential problem solving. The need to solve the top level goal can obscure the fact that it could possibly by solved later with greater facility. For instance, a plan for taking a trip by car could include: 1. Get suitcase 2. Pack clothes in suitcase 3. Put suitcase in car If the raincoat is already in the car, this would involve getting it from the car only to bring it back later inside the suitcase. Conceivably, it would be simpler to bring the packed suitcase to the car and put the raincoat inside it at that time. This shows that goals need not have an immutable hierarchy. Further, there are times when achieving several low level goals is more desirable than achieving the top level goal. In addition to the above there is another problem that exists in domains that interface to the real world, where sensed parameters that have a quasi-continuous character may have to be quantized. Premature quantization of variables loses information and can cause problems when the variable is to be used later for making decisions. For instance, if day/night is a binary variable and it is advantageous to be in day, a program may arrange its problem solving behavior so that it samples the environment just before day turns to night (by the system definition), and, being satisfied with what it finds, pronounces this branch of the solution search as favorable. If it had been forced to continue the branch even a few steps, if would have come to a different conclusion as night was closing in. However, quantization of the relatively continuous day/night variable causes the blenrish effect [2], a behavior anomaly similar to the horizon effect [l], but with the step size of the variable rather than the fixed depth of the search being the culprit. This problem can be prevented by retaining a variable in its quasi-continuous state as long as possible. However, if a variable has a very large range it is impractical to create tests for each value in the range. Resorting to the testing of sub-ranges merely recreates the problem. Thus, discrete treatment of such a variable can cause problems, no matter how it is done. 3 A Better Way Arithmetic functions can do all the above things easily and cheaply if they are constructed in the right way. A polynomial of terms that represent important features in the domain is constructed. We have described our SNAC method of constructing such polynomials and shown [2, 41 that: e It is important that the values of terms vary smoothly. o Non-linearity of terms is extremely important for expertise. e Some method must exist for determining the degree to which each feature is applicable in the present situation. This is done with slowly varying variables that we call application coefficients. The SNAC method also makes it possible to avoid the previously vexing problem of getting stuck on a sub-optimal hill while hill-climbing. Figure 1 shows how getting stuck on a hill is avoided. With non-linear functions, the peaks of hills can be rounded so that retaining the peak becomes less desirable, especially if some other high ground is in view of the searching process. Further, with application coefficients it is possible to change the contour of the hill even as it is being climbed. This is shown in a - c; the arrow showing the location of the current state. As the hill is being climbed, one or more application coefficients that sense the global environment cause the goal of achieving the hilltop to become less important since it is very near being 135 ' * 3 4 5 6 Black Figure 1: Effect of SNAC on Hill Shape achieved. The change in value of the application coefficients causes the contour of the hill to begin to flatten, making the achievement of the summit less desirable, and resulting in the program looking for the next set of goals before even fully achieving the current set. Thus application coefficients can direct progress by reducing the importance of goals that are near being achieved, have already been achieved, or arc no longer important. The above is achieved mathematically as follows: The function p + Y* = C* for -C 2 X 5 C and Y> 0 will produce a semi-circle similar to Figure la. If we now change the function to be >+ A* p= Ce where A 2 1 is an application coefficient (a variable), we can flatten the semi-circle into a semi-elipse of arbitrary flatness. Here, let OLDX be OLDX increases in value. The construction is finalized by only recognizing values of A while OLDX is in the range of (say) -2C to the hill would never seem a desirable thing to do because the program could not tell the difference between getting there when it was far away or already very close. 4 An Example of SNAC Sensitivity The backgammon position in Figure 2 occurred in the Anal game of the match in which my program, BKG 9.8, beat World Champion Luigi Villa in July, 1979. In this position, BKG 9.8 had to play a $1. There are four factors that must be considered here: 1. Black has established a strong defensive points made on the 20 and 22 points. backgame position with 2. In backgame positions timing is very important. Black is counting on hitting White when he brings his men around and home. At such time he must be prepared to contain the man sent back. This can only be done if the rest of his army is not too far advanced so it can form a containing pocket in front of the sent-back man. At the moment Black would not mind having additional men sent back in order to delay himself further and improve his timing. 24 23 22 21 20 19 White 18 17 16 15 14 13 Figure 2: Black to Play a 5,l 3. There is also a possibility that Black could win by containing the man that is already back, but this is rather slim since White can escape with any 5 or 6. However, blockading this man is of some value in case White rolls no 5’s or 6’s in the near future. 4. In case the sole White back man does not escape, there is a possibility of bringing up the remainder of Black’smen not used for the backgame and trying to win with an attack against the back man. In view of the above it is very difficult to determine the right move, and none of the watching experts succeeded in finding it. The most frequently mentioned move was 13-7, which brings a man into the attack and hopes he is hit so as to gain timing (delay one’s inevitable advance). However, BKG 9.8 made the better move of playing 13-8, 3-2, breaking up its blockade somewhat in order to get more attack, and attempting to actively contain the White back man. It did not worry about the increased chance of being hit, as this only helps with later defense. This gives the program two chances to win: If the attack succeeds, and by getting more men sent back, if the attack fails it improves the likelihood of success of its backgame. I have not seen this concept in this form before in books or games. Humans tend to not want to break up the blockade that they have painstakingly built up, even though it is now the least valuable asset that Black has. It is instructive to see how the program arrived at the judgment it made; one that it had never been tested for. Black has 28 legal moves. Ibe top choices of the program were (points of the Scoring pOlYnOmid in parentheses): 13-8, 3-2 (687); 10-5, 3-2 (682); 13-8, 10-9 (672); and 136 13-7 (667). The third and fourth choices were the ones most frequently mentioned by watching experts, thus showing they missed the idea of breaking up the blockade; the thing common to the program’s top two choices. Let us see why it judged the move actually played as better than the third choice (12-17, 15-16). The program considers many factors (polynomial terms) in its judgments and quite a few of these are non-linear. The six factors on which the two moves differed were (points for each and difference in parentheses): 1. Containment of enemy man (177, 131, +46). The move made does hinder his escape more. Containment is always desirable unless one is far ahead and the end of the game is nearing. 2. Condition of our home board (96, 110, -14). It breaks up one home board point. Breaking up the board (points 1 thru 6) is never considered desirable. 3. Attack (37, 21, +16). It is the best attacking move. Attack is desirable unless we are jeopardizing a sure win in the process. 4. Defensive situation (246,260, -14). The move slows White down, thus could reduce the effectiveness of the backgame. 5. Long-term positional (-2, 11, -13). It puts a man on the 2 point, which is undesirable when the game still has a long way to go because it is too far%dvanced to be able to influence enemy men from there. 6. Safety of men (-12, -4, -8). The move made is dangerous. The program realizes this, but also understands that with a secure defensive position such danger is not serious. However, all other things being equal, it would prefer the least dangerous move Thus the better containment and attack are consider-cd to be more important than the weakening of the homcboard, the temporary slowing down of White, the long-term positional weakness, and the safety of the men. The difference between the first and second choice was that in the first choice the attack is slightly stronger. The importance of each of the above terms varies with the situation. In the example, a backgame is established; else the safety term would outweigh the attack term, and BKG 9.8 would not leave two blots in its home board. It does recognize the degree of danger, however, and will not make a more dangerous move unless it has compensating benefits. This is typical of the influence that application coefficients exert in getting a term to respond to the global situation. 5 Perspective We have been employing the SNAC method of making judgments for over two years now, and are struck with its simplicity and power. The happiness example posed earlier is solved trivially in all its forms with SNAC. If the above travel planning problem were solved as a search problem using SNAC functions that measure the economy of effort of the steps used, then undoubtedly SNAC would also do better than sequential planning based on rules, with no evaluation of outcome other than success or failure. At the moment it is difficult to determine what role, if any, SNAC like mechanisms have in human thinking. We have constructed them to simulate lower level “intuitive” type of behavior, and they appear to work admirably in capturing good judgment in the large domain of backgammon. We conjecture that as variables become more and more discrete in character and as criteria for success become more obvious, reasoning gradually replaces such judgment making. At present our backgammon program is being modified to be able to interpret its own functions with the aim of being able to explain its actions, and ultimately being able to identify its failures by type and modifying the culprit functions. 111 PI [31 [41 References Berliner, H. J. Some Necessary Conditions for a Master Chess Program. In Third Inlernational Joinl Conference on Arlificial Intelligence, pages 77-85. IJCAI, 1973. Berliner, H. On the Construction of Evaluation Functions for Large Domains. In Sixth International Joinr Conference on Arbjicial Intelligence, pages 53-55. IJCAI, 1979. Berliner, H. Some Observations on Problem Solving. In Proceeding of Ihe Third CSCSI Conference. Canadian Society for Computational Studies of Intelligence, 1980. Berliner, H. J. Backgammon Computer Program beats World Champion. Arlificial Intelligence 14(l), 1980. 137

1980

74

72

ABSTRACT MULTIPLE-AGENT PLANNING SYSTEMS Kurt Konolige Nils J. Nilsson SRI International, Menlo Park, California We analyze problems confronted by computer agents that synthesize plans that take into account (and employ) the plans of other, similar, cooperative agents. From the point of view of each of these agents, the others are dynamic entities that possess information about the world, have goals, make plans to achieve these goals, and execute these plans. Thus, each agent must represent not only the usual information about objects in the world and the preconditions and effects of its own actions, but it must also represent and reason about what other agents believe and what they may do. We describe a planning system t??at address es these is show how it solves a sample problem . sues and INTRODUCTION Certain tasks can be more advantageously performed by a system composed of several "loosely coupled," cooperating artificial intelligence (AI) agents than by a single, tightly integrated system. These multiple agents might be distributed in space to match the distributed nature of the task. Such systems are often called distributed artificial intelligence (DAI) systems Ll]. We are interested here in systems where the component agents themselves are rather complex AI systems that can generate and execute plans, communicate with each other. make infe rences, and Among the poten tial advantages of such DA1 systems are graceful (fail-soft) degradation characteristics (no single agent need be indispensable), upward extensibi lity (new agents can be added without requiring major system redesign), and communication efficiency (a message- sending agent can plan its "communication acts" carefully, taking into account the planning and inference abilities of the receiving agents). In planning its actions, each agent must consider the potential actions of the other agents. Previous AI research on systems for generating and executing plans of actions assumed a single planning agent operating in a world that was static except for the effects of the actions of the planning agent itself. Examples of such systems include STRIPS [2], NOAH [S], and NONLIN [4]. Several important extensions must be made to planning systems such as these if they are to function appropriately in an environment populated by other planning/execution systems. First, each agent must be able to represent certain features of the other agents as well as the usual information about static objects in the world. Each agent must have a representation for what the other agents "believe" about themselves, the world, and other agents. Each agent must have a representation for the planning, plan-execution, and reasoning abilities of the other agents. These requirements presuppose techniques for representing the "propositional attitudes" believe and want. Second, among the actions of each agent are ~communication actions" that are used to inform other agents about beliefs and goals and to request information. Finally, each agent must be able to generate plans in a world where actions not planned by that agent spontaneously occur. We introduce here the notion of spontaneous operators to model such actions. In this paper, we give a brief summary of our approach toward building DA1 systems of this sort. It should be apparent that the work we are describing also has applications beyond DAI. For example, our multiple agents plan, execute, and understand communication acts in a manner that could illuminate fundamental processes in natural- language generation and understanding. (In fact, some excellent work has already been done on the subject of planning "speech acts" [5-61.) Work on models of active agents should also contribute to more sophisticated and helpful "user models" for interactive computer systems. The development of multiagent systems might also stimulate the development of more detailed and useful theories in social psychology--just as previous AI work has contributed to cognitive psychology. At this early stage of our research, we are not yet investigating the effects of differing "social organizations" of the multiple agents. Our work to date has been focussed on representational problems for such systems independent of how a society of agents is organized. A MULTIPLE-AGENT FORMALISM Each agent must be able to represent other agents' beliefs, plans, goals, and introspections about other agents. Several representational formalisms might be used. McCarthy's formalism for first-order theories of individual concepts and From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. propositions [7] is one possibility, although certain problems involving quantified expressions in that formalism have not yet been fully worked out. Another candidate is Moore's first-order axiomatization of the possible world semantics for the modal logic of knowledge and action [8-g]. Appelt [lo] h as implemented a system called KAMP that uses Moore's approach for generating and reasoning about plans involving two agents. We find Moore's technique somewhat unintuitive, and it seems needlessly complex when used in reasoning about ordinary (nonattitudinal) propositions. Here we develop a representation for each agent based on Weyhrauch's notion of multiple first-order theories and metatheories [ll]. Al: Facts --w-e HOLDING(Al,A) HOLDING(Al,B) CLEAR(B) A pickup(Al,B) HANDEMPTY(A1) CLEAR(B) / putdown(A1,A) HOLDIiG(Al,A) Using Weyhrauch's termi nology, each computer individual is defined by the combination of a first-order language, - a simulation structure or partial model for that language, a set of Facts (expressed in the language), and a Goal Structure that represents a goal for the agent and a plan for achieving it. We assume that each agent has a deductive system (a combination of a theorem-prover and attached procedures defined by the simulation structure) used for deriving new facts from the initial set and for attempting to determine whether goals and subgoals follow from the set of facts. Each agent is also assumed to have a planning system (such as STRIPS) for creating plans to achieve goals. Using a typical "blocks-world" example, we diagram an agent's structure in the following way: Al (agent's name): Facts ----- Goal --s-e HOLDING(Al,A) HOLDING(Al,B) CLEAR(B) Viewed as a computational entity, an agent's structure is typically not static. Its deductive and sensory processes may expand its set of facts, or its planning system may create a plan to achieve a goal. Also, once a plan exists for achieving a goal, the agent interacts with its environment by executing its plan. In this summary, we deal only with the planning processes of agents. In the example above, the occurrence of the goal HOLDING(Al,B) in Al's goal structure triggers the computation of a plan to achieve it. Once generated, this plan is represented in the goal structure of agent Al as follows: Goal m---w Plans are represented by goal/subgoal trees composed of planning operators and their preconditions. We assume a depth-first ordering of the operators in the plan. Now let us introduce another agent, AO. Agent A0 can have the same sort of structure as Al, including its own first-order language, a description of the world by wffs in that language (facts), a simulation structure, a goal structure, and a planner and deducer. Some of AO's facts are descriptions of Al's structure and processes. By making inferences from these facts, A0 can reason about the planning and deductive activities of Al and thus take Al into account in forming its own plans. Also, a structure similar to Al's actual structure, and procedures similar to Al's deducer and planner, can be used as components of AO's simulation structure. Procedural attachment to these "models" of Al can often be employed as an alternative method of reasoning about Al. (Of course, A0 may have an incomplete or inaccurate model of Al.) Because Al's structure is a first-order language (augmented by certain other structures), A0 can use a formal metalanguage to describe it along the lines suggested, for example, by Kleene [12] and developed in FOL by Weyhrauch [Ill. A0 has terms for any sentence that can occur in Al's language or for any of Al's goals or plans; the predicates FACT and GOAL are used to assert that some sentences are in Al's Facts list or goal structure. Consider the following example: assume A0 is holding block A, block B is clear, and A0 believes these facts and further believes that Al believes A0 is holding A and that Al believes B is not clear. A0 would have the following structure: AO: Facts ----- HOLDING(AO,A) CLEAR(B) ;;;;;;;,'HOLDING(AO,A)') /-CLEAR(B) '> We use quote marks to delimit strings, which may have embedded string variables. The denotation of a ground string is the string itself. Thus, the intended interpretation of FACT(Al,'HOLDING(AO,A)') is that the wff HOLDING(AO,A) is part of the facts 139 list of Al (that is, Al "believes" that A0 is holding A). By using the FACT predicate and terms denoting other agents and wffs, any facts list for other agents can be described. (A0 can describe its own beliefs in the same manner.) We purposely use "believe" instead of "know" because we are particularly interested in situations where agents may be mistaken in their representation of the world and other agents. In the above example, AO's opinions about its own and Al's belief about whether or not block B is clear are inconsistent. We avoid formalizing "know" and thus do not take a position about the relationship between knowledge and belief (such as "knowledge is justified true belief"). We can describe some of the usual properties of belief by axioms like FACT(x,p) => FACT(x,'FACT(x,p)'); i.e., if an agent believes p, it believes that it believes p. We do not, however, use an axiom to the effect that agents believe the logical consequences of their beliefs, because we want to admit the possibility that different agents use different procedures for making inferences. In particular, we want to emphasize that the deductive capabilities of all agents are limited. While the static structure of Al is described, for AO, by FACT and GOAL predicates, the action of Al's deductive system and planner can also be axiomatized (for AO) at the metalevel (see Kowalski [13] for an example). This axiomatization allows A0 to simulate Al's deducer or planner by purely syntactic theorem-proving. Thus A0 might use predicates such as ISPROOFfx,p) and ISPLAN(x,p) to make assertions about whether certain proof or plan structures are proofs or plans for other agents (or for itself). In certain cases, A0 can find out if Al can deduce a particular theorem (or if Al can create a plan) by running its procedural model of Al's deducer (or planner) directly, rather than by reasoning with its own facts. This is accomplished by semantic attachments of models of Al's deducer and planner to the predicates ISPROOF and ISPLAN in AO's metalanguage. Semantic attachment thus allows A0 to "think like Al" by directly executing its model of Al's planner and deducer. (Here, we follow an approach pioneered by Weyhrauch [ll] in his FOL system of using semantic attachments to data structures and programs in partial models.) The same kind of attachment strategy can be used to enable A0 to reason about its own planning abilities. The usual problems associated with formalizing propositional attitudes [8,14] can be handled nicely using the FACT predicate. For example, the atomic formula FACT(Al,'CLEAR(A) v CLEAR(B)') asserts the proposition that Al believes that A is clear or B is clear, and is not confused with the formula [FACT(A~,'CLEAR(B)') v FACT(Al,'CLEAR(A)')], which asserts the different proposition that Al believes that A is clear or Al believes that B is clear. Furthermore, semantic attachment methods confer the advantages of the so- called "data base approach" [8] when appropriate. Of particular importance among statements concerning AO's beliefs about Al are those that involve "quantifying in," i.e., where a quantified variable appears inside the term of a FACT predicate. We follow the general approach of Kaplan [15] toward this topic. For example, the sentence (Ex)FAOT(A~,'H~LDING(A~,X)') occurring among AO's facts asserts that Al is holding an identified (for Al) block without identifying it (for AO). AN EXAMPLE We can illustrate some of the ideas we are exploring by a short example. Suppose that there are two agents, A0 and Al, each equipped with a hand for holding blocks. Initially Al is holding a block, A, and A0 wants to be holding A. Suppose that A0 believes these initial facts, but (to make our example more interesting) A0 has no information about whether or not Al itself believes it is holding A. Thus, the initial structure for A0 is: AO: Facts Goal ------ me--- HANDEMPTY(A0) HOLDING(AO,A) HOLDING(Al,A) Let us assume the following planning operators (for both A0 and Al). We use standard STRIPS notation [16]. ('P&D' denotes the precondition and delete lists; 'A' denotes the add list.) putdown(x,b) agent x puts block b on the table P&D: HOLDING(x,b) A: ONTABLE & CLEAR(b) & HANDEMPTY pickup(x,b) agent x picks up block b P&D: CLEAR(b) & HANDEMPTY A: HOLDING(x,b) asktoachieve(x,y,g) agent x gives agent y the P: T goal denoted by string g A: GOAL(y,g) tell(x,y,s) agent x tells agent y the P: FACT(x,s) expression denoted by string s A: FACT(y,s) Agents take into account the possible actions of other agents by assuming that other agents generate and execute plans to achieve their goals. The action of another agent generating and executing a plan is modelled by a "spontaneous operator." A spontaneous operator is like an ordinary planning operator except that whenever its preconditions are satisfied, the action corresponding to it is presumed automatically executed. Thus, by planning to achieve the preconditions of a spontaneous operator, a planning agent can incorporate such an operator into its plan. Let us assume that agent A0 can use the operator "achieve" as a spontaneous operator that models the action of another agent generating and executing a plan: achieve(x,g) agent x achieves goal g by creating and executing a plan to achieve g. PC: GOAL(x,g) & ISPLAN(x,p,g,x) & ISPLAN(x,p,g,AO) D: **the delete list is computed from the plan, p** A: FACT(x,g) FACT(AO,g) The expression ISPLAN(x,p,g,f) is intended to mean that there is a plan p, to achieve goal g, using agent x‘s planner, with facts belonging to agent f. Our precondition for achieve ensures that before A0 can assume that condition g will be spontaneously achieved by agent x, A0 has to prove both that agent x can generate a plan from its own facts and that agent x could generate a plan from AO's facts (to ensure that the plan is valid as far as A0 is concerned). Here are some axioms about ISPLAN that A0 will need: 1) FACT(x,w) => ISPLAN(x,NIL,w,x) (If x already believes w, then x has a plan, namely NIL, for achieving w.) 2) [ISPLAN( X,U,Y,d & PC(Z,Y) & OP(x,z,g)] => ISPLAN(x,extend(u,z),g,x) (If x has a plan, namely u, to achieve the preconditions, y, of its operator, z, with add list containing g, then x has a plan, namely extend(u,z) for achieving g. The functional expression, extend(u,z), denotes that plan formed by concatenating plan z after plan u.) The planning tree in Figure 1 shows a possible plan that A0 might generate using its facts (including axioms about ISPLAN) and operators. The sequence of operators in this plan is {tell(AO,Al,'HOLDING(Al,A)'), asktoachieve(AO,Al,'CLEAR(A)'), achieve(Al,'CLEAR(A)'), pickup(AO,A)j. Note that semantic attachment processes were used in several places in generating this plan. We leave to the control strategy of the system the decision about whether to attempt to prove a wff by semantic attachment or by ordinary syntactic methods. We are now in the process of designing a system for generating and executing plans of this sort. Space prohibits describing some additional features of our system, including its control strategy for generating plans. We plan to experiment with a,, complex of several agents, each incorporating planning systems like that briefly described here. We gratefully acknowledge helpful discussions with Doug Appelt, Bob Moore, Earl Sacerdoti, Carolyn Talcott and Richard Weyhrauch. This research is supported by the Office of Naval Research under Contract No. N00014-80-C-0296. 1 . 2. 3. 4. 5. 6. 7. 8. 9. 10. REFERENCES Sacerdoti, E. D., "What Language Understanding Research Suggests about Distributed Artificial IntelligenceT,, in Distributed Sensor Nets, ppm 8-11. Paper presented at the DARPA Workshop, Carnegie-Mellon University, Pittsburgh, Pennsylvania (December 7-8, 1978). Fikes, R. E. and N. J. Nilsson, "STRIPS: A New Approach to the Application of Theorem Proving to Problem Solving,,, Artificial Intelligence, 2(3/4), pp. 189-208 (1971). Sacerdoti, E. D., A Structure for Plans and Behavior, (New York: Elsevier, 197r Tate, A., "Generating Project Networks,,, in IJCAI-5, pp. 888-893 (1977). Searle, J. R., "A Taxonomy of Illocutionary Acts," in Language Mind and Knowledge, K. Gundersm,(University of Minnesota Press, 1976). Cohen, P. R. and C. R. Perrault, "Elements of a Plan-Based Theory of Speech Acts,,, Cognitive Science, _ 3(3), pp. 177-212 (1979)e McCarthy, J., "First Order Theories of Individual Concepts and Propositions,,, in Machine Intelligence 9, pp.-120-147, -ayes and D. Michie (Eds.), (New York: Halsted Press, 1979). Moore, R. C., "Reasoning About Knowledge and Action,,, in IJCAI-5, pp. 223-227 (1977)' Moore, R. C., "Reasoning About Knowledge and Action,,, Artificial Intelligence Center Techincal Note 191, SRI International, Menlo Park, California'(lY80). 141 13. Kowalski, R., Logic for Problem Solving, (New 15. Kaplan, D., "Quantifying In," in Reference and York: North-Hom,T'79). Modality, L. Linsky (Ed.), pp. 112-144, (London: Oxford University Press, 1971). 14. Quine, W. V. O., "Quantifiers and Propositional Attitudes," in Reference and 16. Nilsson, N. J., Principles of Artificial Modality, L. Linsky (Ed.), pp. 101-111, Intelligence, (Menlo Park: Eoga Publishing (London: Oxford University Press, 1971). co., 1980). Al: HOLDING(AO,A) bkup(AO,A) HANDEMPTY(A0) (initial fact) achieve(Al,'CLEAR(A)') plan axiom {extend(u,z)/pj ISPLAN(Al,p,'CLEAR(Al)',AO) (verified by proc. attach. to a model of Al's planner using AO's facts after substituting for p) GOAL(Al,'C EAR(A)') [ISPLAN(Al,u,y,Al) & PC(z,y) & OP(Al,z,'CLEAR(A)')] asktoachieve(AO,Al, I 'CLEAR(A)') proc. attach. to PC and OP x {'putdown(Al,A)'/z, 'HOLDING(AI,A)'/y~ T ISPLAN(Al,u,'HOLDING(Al,A)',Al) I plan axiom b'IL/uj FACT(Al,'HOLDING(Al,A)') I tell(AO,A1,'HOLDING(A1,A)') FACT(AO,'HOLDING(Al,A)') (verified by proc. attach. to AO's "fact finder".) Figure 1 ' 142

1980

75

73

SCOUT: A SIMPLE GAME-SEARCHING ALGORITHM WITH PROVEN OPTIMAL PROPERTIES Judea Pearl Cognitive Systems Laboratory School of Engineering and Applied Science University of California Los Angeles, California 90024 ABSTRACT This paper describes a new algorithm for searching games which is conceptually simple, space efficient, and analytically tractable. It pos- sesses optimal asymptotic properties and may offer practical advantages over a-6 for deep searches. I. INTRODUCTION We consider a class of two-person perfect information games in which two players, called MAX and MIN, take alternate turns in selecting one out of d legal moves. We assume that the game is searched to a depth h, at which point the terminal positions are assigned a static evaluation function VO* The task iS to evaluate the minimax value, Vh, of the root node by examining, on the average the least number of terminal nodes. SCOUT, the algorithm described in this paper, has evolved as a purely theoretical tool for ana- lyzing the mean complexity of game-searching tasks where the terminal nodes are assigned random and independent values [l]. With the aid of SCOUT we were able to show that such games can be evaluated with a branching factor of P*/(l-P*), where P* is the root of xd+x-1 = 0, and that no directional algorithm (e.g., ALPHA-BETA) can do better. We have recently tested the performance of SCOUT on a 'real' game (i.e., the game of Kalah) and were somewhat surprised to find that, even for low values of h, the efficiency of SCOUT surpasses that of the a-6 procedure [2]. The purpose of this paper is to call attention of game-playing practi- tioners to the potentials of SCOUT as a practical game-searching tool. Section II describes the operation of SCOUT in conceptual terms avoiding algorithmic details. Section III presents, without proofs, some of the mathematical properties of SCOUT and compares them to those of the a-~ procedure. Finally empirical results are reported comparing the performances of SCOUT and a-6 for both random and dynamic orderings. * Supported in part by NSF Grants MCS 78-07468 and MCS 78-18924. II. THE SCOUT ALGORITHM SCOUT invokes two recursive procedures called EVAL and TEST. The main procedure EVAL(S) returns V(S), the minimax value of position S, whereas the function of TEST(S, v, >) is to validate (or refute) the truth of the inequality V(S) > v where v is some given reference value. Procedure: TEST(S, v, >) To test whether S satisfies the inequality V(S) > v, start applying the same test (calling itself) to its successors from left to right: If S is MAX, return TRUE as soon as one suc- cessor is found to be larger than v; return FALSE if all successors are smaller than or equal to v. If S is MIN, return FALSE as soon as one suc- cessor is found to be smaller than or equal to v; return TRUE if all successors are larger than v. An identical procedure , called TEST(S, v, z), can be used to verify the inequality V(S) 1 v, with the obvious revisions induced by the equality sign. Procedure: EVAL(S) EVAL evaluates a MAX position S by first eval- uating (calling itself) its left most successor S,, then 'scouting' the remaining successors, from left to right, to determine (calling TEST) if any meets the condition V(Sk) > V(S1). If the inequal- ity is found to hold for Sk:, this node is then evaluated exactly (calling EVAL(Sk)) and its value V(Sk) iS used for subsequent 'Scoutings' tests. Otherwise Sk iS exempted from evaluation and Sk+1 selected for a test. When all successors have been either evaluated or tested and found unworthy of evaluation, the last value obtained is issued as V(S). An identical procedure is used for evaluating a MIN position S, save for the fact that the event V(Sk) 1 V(S1) now constitutes grounds for exempt- ing S from evaluation. Flow-charts describing both 6 COUT and TEST in algorithmic details can be found in [l]. At first glance it appears that SCOUT is very wasteful; any node Sk which is found to fail a test criterion is submitted back for evaluation. The terminal nodes inspected during such a test may 143 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. (and in general will) be revisited during the eval- uation phase. An exact mathematical analysis, however, reveals that the amount of waste is not substantial and that SCOUT, in spite of some dupli- cated effort, still achieves the optimal branching factor P*/(l-P*), as will be demonstrated in Sec- tion III. Two factors work in favor of SCOUT: (1) most tests would result in exempting the tested node (and all its descendents) from any further evalua- tion, and (2) testing for inequality using the TEST(S, v) procedure is relatively speedy. The speed of TEST stems from the fact that it induces many cutoffs not necessarily permitted by EVAL or any other evaluation scheme. As soon as one suc- cessor of a MAX node meets the criterion u(s ) > v, all other successors can be ignored. EVAL, y I: contrast, would necessitate a further examination of the remaining successors to determine if any would possess a value higher than V(Sk). Several improvements could be applied to the SCOUT algorithm to render it more efficient. For example, when a TEST procedure issues a non-exempt verdict, it could also return a new reference value and some information regarding how the decision was obtained in order to minimize the number of nodes to be inspected by EVAL. However, the analysis presented in Section III, as well as the simulation tests, were conducted on the original version described above. These studies show that, even in its unpolished form, SCOUT is asymptotically opti- mal over all directional algorithms and is somewhat more efficient than the a-6 procedure for the game tested (i.e., Kalah). Recently, Stockman [3] has also introduced an algorithm which examines fewer nodes than a-6. However, Stockman's algorithm requires an enormous storage space for tracing back a large number of potential strategies. SCOUT, by contrast, has storage requirements similar to those of a-6; at any point in time it only maintains pointers along one single path connecting the root to the current- ly expanded node. III. ANALYSIS OF SCOUT'S EXPECTED PERFORMANCE In this section we present, without proofs, some mathematical results related to the expected number of nodes examined by SCOUT and a-B. Addi- tional results, reference [l]. including proofs, can be found in The model used for evaluating these algorithms consists of a uniform tree of height h (h even) and branching factor d, where the terminal positions are assigned random values, independently drawn from a common distribution F. We shall refer to such a tree as a (h, d, F)-tree. Theorem 1: The root value of a (h, d, F)-tree with continuous strictly increasinq terminal dis- tribution F converges, as h -+ CQ (in probability) to the 1-P*)-fractile of F, where P* is the solution 6 of x +x-l = 0. If the terminal values are discrete: v, < v2 < . . . < VM, then the root value converges to a definite limit iff l-P* # F(v.) for all i, in which case the limit is the smallelt vi satisfying l-P* < F(Vi). Definition: Let A be a deterministic algo- rithm which searches the (h, d, F)-game and let IA(h, d, F) denote the expected number of terminal positions examined by A. The quantity: r&b F) = 'im [IA(h, d, F)]l'h h-too is called the branching factor corresponding to the algorithm A. Definition: Let C be a class of algorithms capable of searching a general (h, d, F)-tree. An algorithm A is said to be asymptotically optimal over C if for all d, F, and BEC, rA(& F) 5 rg(d, I=). Definition: An algorithm A is said to be directional if for some linear arrangement of the terminal nodes it never selects for examination a node situated to the left of a previously examined node. Theorem 2: The expected number of terminal positions examined by the TEST algorithm in the orooosition "V(S)> v" for the root gf-a testing (h, h, F)-tree, has a branching v * v* and P*/(l-P*) if v = v* F(v*) = l-P* and P* is the root factor d112 if , where v* sati of xd+x-1 = 0. sfies Theorem 3: TEST is asymptotically optimal over all directional algorithms which test whether the root node of a (h, d, F)-tree exceeds a speci- fied reference v. Corollar 1: Any procedure which evaluates a (h, debust examine at least 2dh/2-1 nodes. Corollary 2: The expected number of terminal positions examined by any directional algorithm which evaluates a (h, d)-game tree with continuous terminal values must have a branching factor great- er or equal to P*/(l-P*). The quantity P*/(l-P*) was shown by Baudet [3] to be a lower bound for the branching factor of the a-6 procedure. Corollary 2 extends the bound to all directional game-evaluating algorithms. Theorem 4: The expected number of terminal examinations performed by SCOUT in the evaluation of (h, d)-game trees with continuous terminal values has a branching factor of P*/(l-P*). Theorem 5: The expected number of terminal examinations performed by SCOUT in evaluating a (h, d, F)-game with discrete terminal values has a branchina factor dl/L, with exceptions only when one of the discrete values, v*, satisfies F(v*) = 1-p*. 144 Corollary 3: For games with discrete terminal values satisfying the conditions of Theorem 5, the SCOUT procedure is asymptotically optimal over all evaluation algorithms. The improvement in efficiency due to the dis- crete nature of the terminal value manifests itself only when the search depth h is larger than log M/log [d(l-P*)/P*], where M is the quantization density in the neighborhood of VO = v*. The branching factor of a-8 is less tractable than that of SCOUT. At the time this paper was first written the tightest bounds on r,,B were those delineated by Baudet [3] giving the lower bound r,-@ 2 P*/(l-P*) (a special case of Ccrollary 2) and an upper bound which is about 20 percent higher over the range 2 5 d I 32. Thus, SCOUT was the first algorithm known to achieve the bound P*/(l-P*) and we were questioning whether a-6 would enjoy a comparable asymptotic performance. More- over, it can be shown that neither SCOUT nor a-6 dominate one another on a node-by-node basis; i.e., nodes examined by SCOUT may be skipped by a-6 and vice versa [l]. The uncertainty regarding the branching factor of the a-0 procedure has recently been resolved [5] Evidently, U-B and SCOUT are asymptotically equiva- lent; r e uals P*/(l-P*) for continuous valued trees a%'dl92 for games with discrete values. P For low values of h the branching factor is no longer an adquate criterion of efficiency and the comparison between SCOUT and a-6 must be done empirically. The following table represents the number of node inspections spent by both algorithms on the game of Kalah (l-in-a-hole version) [2]: It appears that as the search depth increases SCOUT offers some advantage over cl-@. Experiments with higher numbers of stones in each hole indicate that this advantage may deteriorate for large d. We suppose, therefore, that SCOUT may be found useful in searching games with high h/d ratios. Cl1 PI [31 [41 [51 REFERENCES Pearl, J. "Asymptotic Properties of Minimax Trees and Game-Searching Procedures." UCLA- ENG-CSL-7981, University of California, Los Angeles, March 1980, to be published in Artificial Intelligence. Noe, T. "A Comparison of the Alpha-Beta and SCOUT Algorithms Using the Game of Kalah." UCLA-ENG-CSL-8017, University of California, Los Angeles, April 1980. Stockman, G. "A Minimax Algorithm Better Than Alpha-Beta?" Artificial Intelligence 12, 1979, 179-196. Baudet, G. M. "On the Branching Factor of the Alpha-Beta Pruning Algorithm." Artificial Intelligence 10, 1978, 173-199. Pearl, J. "The Solution for the Branching Factor of the Alpha-Beta Pruning Algorithm." UCLA-ENG-CSL-8019, University of California, Los Angeles, May 1980. II Random Orderina II Dvnamic Orderina Search Depth 2 SCOUT 82 a-B 70 % Improvement -17.0 .# J % SCOUT a-6 Improvement 39 37 -5.4 I 3 II 394 I : 380 -3.7 62 61 -1. 4 1173 1322 +11.3 91 96 -1. 5 2514 4198 +40.1 279 336 +17. 6 5111 6944 +26.4 371 440 +15. .45

1980

76

74

Problem Solving in Frame-Structured Systems Using Interactive Dialog Harry C. Reinstein IBM Palo Alto Scientific Center 1530 Page Mill Road Palo Alto, Ca. 94304 ABSTRACT This paper provides an overview of the process by which problem solving in a particular frame-like knowledge-based system is accomplished. The inter-relationship between specialization traversal and entity processing is addressed and the specific role of the user interaction is described. I INTRODUCTION Semantic networks Cl1 and frame-like systems have emerged as powerful tools in a variety of problem domains [Z&l. In many of these systems an initial knowledge base is used to drive an interactive dialog session, the goal of which is the instantiation of the particular knowledge base elements which represent a solution to the problem being addressed. In a system developed at the IBM Scientific Center in Palo Alto [3,41, a dialog is generated from a KRL-based c51 semantic network for the purpose of generating a well-formed definition of a medical sensor-based application program. It is intended that the user of the system be conversant with the problem to be solved by the application but not that they be a computer programmer. The overall logic of this process is the subject of this paper. II THE DIALOG LOGIC --- The ultimate goal of the problem-solving dialog session is the complete instantiation of all entities (knowledge units) relevant to the problem solution. To do this the system must be able to create new work contexts (in our case entities) from existing ones and be able to traverse the specialization hierarchies rooted at these entities to accomplish complete instantiation. The logic governing the interrelationships between these two tasks, and the methods of user interaction tend to characterize frame-based dialog systems. One could, for example, choose to pursue a path of 'least commitment' by processing all relevant references at their highest levels in the specialization hierarchies before attempting deeper specialization traversal [6,71. This approach seems well-suited to problem domains where the solutions are highly dependant on the interaction of constraints between the processed entities. In the case of our application development system it was felt that the solution process would be enhanced if individual entities were completely specialized as they were encountered, with outside references posted as pending work contexts for subsequent processing. This 'greatest commitment' approach seems well-suited to semantic networks in which specialization traversal in one hierarchy provides increasingly more specialized references to other hierarchies, and where the constraints between hierarchies are not potentially inconsistent. An example of this can be seen in the relationship between our SENSOR hierarchy, which provides descriptive knowledge about the analog, digital, and keyboard-entry sensors available in the laboratory and the DATA hierarchy, which describes the kinds of data which can be processed. In these interdependant hierarchies, one finds that BLOOD PRESSURE is measurable only by a subset of PRESSURE SENSORS, and that ARTERIAL BLOOD PRESSURE (a specialization of BLOOD PRESSURE) makes even further specializations on that set. Traversal of either hierarchy will implicitly specialize the related hierarchy. Downward traversal of specialization trees is the main driving force of the dialog system. III MECHANICS OF SPECIALIZATION TRAVERSAL - Entities intended for inclusion in the problem solution are dynamically replicated and inserted into the knowledge structure as an immediate specialization of the entity from which they are copied. As more becomes known about them, these new entities are moved down the specialization hierarchy, always appearing as a specialization of the most constrained model available in the initial knowledge base. These dynamic entities have exactly the same representation as the initial entities and differ from them only in that their constraints can be overwritten in the process of instantiation. If, for example, one of the attributes of a DEVICE is its user supplied name, then the value obtained for that name during the dialog would be placed in the dynamic entity while the corresponding 146 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. entity/attribute in the initial knowledge base is only constrained to be a name. V SUMMARY The mechanism for migrating a dynamic entity down the associated specialization hierarchy may require user interaction. This interaction is accomplished using a video character display with full screen data entry facilities, a light pen, and program function keys. It has been our experience that non-computer-trained users are very sensitive to the level of human factors provided and it is well worth any effort one can make to facilitate interaction. First the user is prompted to supply values for attributes which have been declared in the knowledge base to be user-sourced. (This is equivilent to the 'lab data' assertion in MYCIN c81). Having obtained these, a pattern-matching search is performed to see if specialization is possible. If not, the next step is to attempt specialization by allowing the user to choose from a list of the names of the immediate descendants at the current level in the hierarchy. If the user is unable to select the specialization by name, he or she is interrogated for selected attribute values which, if known, would determine a specialization path. This process continues until a terminus in the hierarchy is reached. During the traversal process any references to other entities must be resolved, and these references generate additional work contexts for the system. It is particularly important that the resolution process be able to determine if the reference should resolve to an already existing dynamic entity or if it should resolve to an entity in the initial knowledge base. Some considerations relevant to this problem are discussed below. Processing entities to their most specialized form is a valid driving function in some knowledge bases, and generating user interaction specifically for this traversal can be a sufficient involvement for the user in the problem solving process. Representing the results of the problem solving session in the same form as, and in direct association with the initial knowledge base has many positive features. Included among these is the ability to use a single search/resolution mechanism to select from either set of entities when building the problem solution. In general, frame-structured knowledge bases, in conjunction with user interaction can provide a powerful problem solving facility. REFERENCES Cl1 Fikes, R. and Hendricks, G. "A Network Based Representation and Its Natural Deduction System" In Proc. IJCAI-77. Cambridge, Massachusetts, August, 1977, pp. 235-245. c21 Waterman, D.A. and Hayes-Roth, F., (Eds.) Pattern-Directed Inference Systems. New York, New York: Academic Press, 1978. [31 Hollander, C.R. and Reinstein, H.C. "A Knowledge-based Application Definition System" In Proc. IJCAI-79. Tokyo, Japan, August, 1979, pp. 397-399. [41 Reinstein, H.C. and Hollander, C.R. "A Knowledge-based Approach to Application Development IV PROCESSING ENTITY REFERENCES for Non-programmers", Report G320-3390, IBM Scientific Center, Palo Alto, California, July 1979. When a reference resolves to a single entity one of three situations prevails: 1) the reference [51 Bobrow, D. and Winograd, T., "An Overview of is to an exactly matching dynamic entity, 2) the m, a Knowledge Representation Language." reference is to an existing dynamic entity which is Cognitive Science 1:l (1977), 3-46. less constrained than desired, or 3) the reference is to an entity in the initial knowledge base. In L61 Martin, N. et al "Knowledge Management for the first of these cases no further processing is Experiment Planning in Molecular Genetics" In Proc. required. In the second case, the more constrained IJCAI-77. Cambridge, Massachusetts, August, 1977, form of the attribute is forced into the dynamic pp. 882-887. entity and a search is performed to see if this new form permits further migration down the [71 Stefik, M.J., Planning With Constraints, Ph.D. specialization hierarchy. In this way values of Thesis, Stanford University, 1980, (available from attributes obtained from specialization down one Computer Science Dept., Stanford University, Report hierarchy can implicitly cause specialization STAN-CS-80-784). acitivity in related hierarchies. In the third case a new dynamic entity would be created. C8l Shortliffe, E.H. MYCIN: Computer-based Medical Consultations, New York, New York: American Elsevier, 1976. 147

1980

77

75

REPRESENTING KNOWLEDGE IN AN INTERACTIVE PLANNER Ann E. Robinson and David E. Wilkins Artificial Intelligence Center SRI International Menlo Park, California 94025 ABSTRACT This note discusses the representation for actions and plans being developed as part of the current planning research at SRI. Described is a method for uniformly representing actions that can take place both in the domain and during planning. The representation accommodates descriptions of abstract (hypothetical) objects. I. INTRODUCTION A principal goal of current planning and plan- execution research at SRI is development. of a planning system that interacts with a person, allowing that person to: (1) explore alternative plans for performing some activity, (2) monitor the execution of a plan that has been produced, and (72 modify the plan as needed during execution. Described here is the knowledge representation being developed. Our research builds' directly on previous planning research and on research in representing the domain knowledge necessary for participating in natural-language dialogs about tasks. In particular, some of our representation ideas are based on the process model formalism described in [2] and [S]. The basic approach to planning is to work within the hierarchical planning paradigm, representing plans in procedural networks, as has been done in NOAH [4] and other systems. Unlike its predecessors, our new system is being designed to allow interaction with users throughout the planning and plan-execution processes. The user will be able to watch and, when desired, guide and/or control the planning process. During execution of a plan, some person or computer system monitoring the execution will be able to specify what actions have been performed and what changes have occurred in the world being modeled. On the basis of this, the plan can be interactively updated to accommodate unanticipated occurrences. Planning and plan-execution can be intermingled by producing a plan for part of an activity and then executing some or all of that plan before working out remaining details. We are extending planning research in several major directions. One of the key directions, the one discussed here, is a method for representing -------- + The research reported here is supported by Air Force Office of Scientific Research Contract F49620-79-C-0188 and by Office of Naval Research Contract N00014-80-C-0300. actions that can take place both in the domain and during planning. Action descriptions (often referred to as operators), procedural networks, and knowledge about domain objects and their interrelationships are represented in the same formalism -- a hierarchy of nodes with attributes. This uniform representation provides the ability to encode partial descriptions of unspecified objects as well as objects in the domain model. Thus, operator descriptions referring to abstract (unbound) objects can be represented in the same formalism as procedural network nodes referring to specific objects in the domain model. (Partial descriptions of unspecified objects will be described here as constraints on the possible values of a variable representing the object.) Operators can be encoded at several levels of abstraction. Each one contains information for planning at the next level of detail. We have already encoded many domain operators for a construction task; planning operators will be encoded shortly. The domain operators provide the planning system with information about producing a plan in the domain. The planning operators provide the planning system with information so it can reason about its own planning process (meta- planning). They also provide a major part of the interface between the planning system and the user, who will be able to direct the planning process via the planning operators. The uniformity of representation for domain knowledge, specific plans of action, and all operators will facilitate both the user's ability to interact with and control the planning system, and the system's ability to incorporate (learn) new operators from plans it has already produced. We will describe the representation in more detail below. 148 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. II. THE FORMALISM The formalism for representing knowledge about actions, plans, and domain objects consists of typed nodes linked in a hierarchy. Each node can have attributes associated with it. There are four node types for representing objects: CLASS, INSTANCE, INDEFINITE, and DESCRIPTION. These will not be discussed in more detail here since they are similar to those occurring in re resentation formalisms such as KRL, FRL, and UNITS -f 51. The node types for representing actions can be grouped into four categories: OPERATOR, for encoding operators; PNET, for representing specific actions (nodes in the procedural network); PLOT, for describing how to expand a given OPERATOR, i.e., a description of an action in greater detail; PNET.ACTION, for encoding plan steps (procedural network nodes) that have been 'executed' and thus represent actions assumed to have occurred in the world being modeled. Nodes can have lists of attributes and can be connected into a hierarchy through CLASS and SUBCLASS links. Attributes of nodes for representing actions include the resources and arguments of the action (i.e., the objects that participate in the action), the action's goal, the action's effects on the domain when it is performed, and the action's preconditions. OPERATOR nodes have a plot attribute which specifies PLOT nodes for carrying out the operator. The PLOT of an operator can be described not only in terms of GOALS to be achieved, but also in terms of PROCESSes to be invoked. (Previous systems would represent a PROCESS as a goal with only a single choice for an action to perform.) The ability to describe operators in terms of both GOALS and PROCESSes will help simplify encoding of operators and will allow the planning system to reason about alternative action sequences more efficiently. Figure 1 shows a sample operator and a PNET it might produce. The figure illustrates the uniformity across different types of nodes in our formalism. The nodes are expressed in the same formalism, and, for the most part, have the same attributes (e.g., resources, shared-resources, arguments, preconditions, purpose) with similar values for these attributes. "Similar values" means that the values refer to the same types of objects -- often the value of an attribute for some node will be more constrained than the value of the same attribute in a corresponding node of a different category. The next two paragraphs illustrate this in detail, after which we describe two instances where the uniformity of the representation is advantageous. Attributes in OPERATOR and PLOT nodes generally refer to variables rather than specific objects, since these are uninstantiated operators that may be instantiated into PNET nodes in different ways during planning. For example, in Figure 1 resource variables meat1 and vegl in the operator FIX.MEAL refer to objects of the meat and vegetable class, respectively. In the expansion of FIX.MEAL, meat1 has been constrained to be a fish (denoted by calling it "fishl") since it was so constrained in the node being OPERATOR expanded. FIX.MEAL RESOURCES. meat 1, vegl - SPLIT GOAL (PREPARED meat 1) RES. meat1 GOAL: (PREPARED vegl) RES: veg 1 f PROCESS: SERVE JOIN - RES meatl, vegl A PNET node which represents usmg FIX.MEAL ma plan. PROCESS. FIX.MEAL . . . 17 RESOURCES. . . 0 flshl veg 1 Expansion of thts PNET node at the next level usmg FIX.MEAL. Figure 1 In our formalism, such variables are described by INDEFINITE nodes with constraints on their possible values. For PNET nodes, attributes frequently refer both to variables (which will often be more constrained in this case) and to completely specified objects. For PNET.ACTION nodes, attributes generally refer to specific objects in the domain model. The system's ability to use INDEFINITE nodes to partially describe objects is important for representing objects with varying degrees of abstractness in the same formalism. Few previous planning systems have used this approach (e.g., NOAH cannot partially describe objects and has different formalisms for describing operators and procedural nets). Stefik's [51 system does allow abstract descriptions and constraints on partially described arguments, but arguments are required to be fully instantiated before the constraints can be evaluated. (See also Hayes-Roth et al. cw The uniformity of representation between PLOT and PNET nodes permits the description of operators as what amounts to generalized fragments of procedural network. This turns problem solving into a process of incremental instantiation. During planning, PNET nodes are incrementally expanded to a greater level of detail by selecting 149 an appropriate operator, determining which of its variables match those in the node being expanded, creating new variable records for those variables not matched, adding any new constraints to these variables, and following the operator's plot description to create new procedural network nodes. The uniformity of representation facilitates this production of PNET nodes from PLOT nodes. Once a plan has been successfully constructed, it may be desirable to save it for subsequent planning activities, incorporating it into the system as a new operator. We expect to develop algorithms for doing this, i.e., producing an operator (with its associated PLOT nodes) from PNET fragments. For each control node and each action- oriented node in a procedural network, a corresponding PLOT node can be easily created for the operator because of the uniformity of representation. The major task remaining in producing an operator would be generalizing the constraints on values for variables in the procedural network nodes into looser constraints in the new operator. An additional uniformity between descriptions of specific actions and operators facilitates the matching of an operator to the node it is to expand. Thus PROCESS and GOAL nodes in a procedural network or plot will have attributes similar to those of the OPERATOR node which represents a more detailed description of their corresponding action. The similarities of representation of all action-oriented nodes facilitates interaction with the user who can talk in the same way about operators, steps in operator plots, and nodes in the procedural network. Similarly, description of actions is facilitated by this uniformity. Organizing the representation as nodes with attributes is, of course, not new and is not essential. The representation could also be expressed in a formal logic (a translation to logic would be fairly straightforward). We have chosen to represent things as nodes with attributes because this blends well with our plans for interaction with the user. III. PARTIAL DESCRIPTION USING CONSTRAINTS Stefik's system [5], one of the few existing planning systems with the ability to construct partial descriptions of an object without identifying the object, contains a constraint- posting mechanism that allows partial descriptions similar to those described above. Our system also provides for partial description using constraints, and extends Stefik's approach in two ways. Unlike Stefik's system, our system permits evaluation of constraints on partially described objects. Both CLASSes and INSTANCES can have constraints. For example, a set can be created which can be constrained to be only bolts, then to be longer than one inch and shorter than two inches, and then to have hex heads. Our system also provides with the for partial descriptions that vary context, thus permitting consideration of alternative plans simultaneously. A context mechanism has been developed to allow for alternative constraints on variables relative to different plan steps. The constraints on a variable's value as well as the binding of a variable to a particular instance (possibly determined during the solution of a general constraint-satisfaction problem) can only be retrieved relative to a particular context. This permits the user to easily shift focus back and forth between alternatives. Hayes-Roth et al. [I] describe the use of a blackboard model for allowing shifting of focus between alternatives. Such focus shifting can not be done in systems using a backtracking algorithm where descriptions built up during expansion of one alternative are removed during the backtracking process before another alternative is investigated. Most other planning systems either do not allow alternatives (e.g., NOAH[4]), or use a backtracking algorithm (e.g., Stefik [5], Tate [6]). IV. CONCLUSION We have described some properties of the knowledge representation developed for our new planning system. Most of the planner is still under development (e.g., critics, reasoning about resources, and search control have yet to be implemented). The central idea discussed here is the uniform representation of the domain operators, planning operators, procedural networks, and knowledge about domain objects. Ways to exploit this uniformity are pointed to. These include a rich interaction with the user, meta-planning, and having the system learn new operators from plans it has constructed. REFERENCES 1. Hayes-Roth, B., F. Hayes-Roth, S. Rosenschein, S. Cammarata, "Modeling Planning as an Incremental, Opportunistic Processll, In Proc. IJCAI-79. Tokyo, Japan, August, 19791 pp. 375- 383. 3. Robinson, A.E., D. Appelt, B. Grosz, G. Hendrix, and J. Robinson, "Interpreting Natural- Language Utterances in Dialogs About Tasks", Technical Note 210, SRI International, Menlo Park, California. March, 1980. 4. Sacerdoti, E., A Structure for Plans and Behavior. Elsevier North-Holland, NewYork. 1977. 5. Stefik, M., Planning With Constraints. Report STAN-CS-80-784, Computer Science Department, Stanford University, Ph.D Dissertation. 1980. 6. Tate, A., "Generating Project Networks", In Proc. IJCAI-77. Cambridge, Mass. August, 1977. 150

1980

78

76

INFEHENCEWITHRECURSIVERULES Stuart C. Shapiro and Donald P. McKay Department of Computar Science State University of New York at Buffalo Amherst, New York 14226 ABSTRACT Recursive rules, such as "Your parents' ances- tors are your ancestors", although very useful for theorem proving, natural language understanding, questions-answe ring and information retrieval systems, present problems for many such systems, either causing infinite loops or requiring that arbitrarily many copies of them be made. We have written an inference system that can use recursive rules without either of these problems. The solu- tion appeared automatically from a technique designed to avoid redundant work. A recursive rule causes a cycle to be built in an AND/OR graph of active processes. Each pass of data through the cycle resulting in another answer. Cycling stops as soon as either the desired answer is produced, no more answers can be produced, or resource bounds are exceeded. Introduction Recursive rules, such as "your parents' ances- tors are your ancestors", occur naturally in inference systems used for theorem proving, question answering, natural language understanding, and information retrieval. Transitive relations, v(x,y,z) [ANCES'lloR(x,y) & ANCESToR(y,z)-+ %%SToR(x,z)], inheritance rules, e.g. ~z(x,y,p) [ISA(x,y) & HAs(y,p) -+ HAS(x,p)l, circular defini- tions and equivalences are all occurrences of recursive rules. Yet, recursive rules present problems for system implemantors. Inference systems which use a %aive chaining" algorithm can go into an infinite loop, like a left-to-right top-down parser given a left recursive grammar [41. Sme systems will fail to use a recur- sive rule more than once, i.e. are incomplete [6,121. Other systems build tree-like data structures (connection graphs) containing branches thelengthofwhichdependonthenlrmberoftimes the recursive rule is to be applied [2,131. Since scme of these build the structure before using it, the correct length of these branches is problematic. Son-e systems eliminate recursive rules by deriving and adding to the data base all implications of the recursive rules in a special pass before normal inference is done [91. The inference system of SNePS [13] was designed ti use rules stored in a fully indexed data base. when a question is asked, the system retrieves --7-s- Theworkwas supported inpartbytheNationa1 Science Foundation under Grant No. MCS78-02274. relevant rules and builds a data structure of processes which attempt to derive the answer frcm the rules and other information stored in the data base. Since we are using a semantic network to represent all declarative information available in the system, we do not make a distinction between "extensional" and "intensional" data bases, i.e. non-rules and rules are stored in the same data base. More significantly, we do not distinguish "base" frm "defined" relations. Specific instances of ANCESTOR maybe storedaswell as a ruledefining ANCESTOR. This point of view contrasts with the basic assurrption of several data base question answering systems [3,8,9]. In addition, the inference system described here does not restrict the left hand side of rules to con- tain only one literal which is a derived relation [3], does not need to recognize cycles in a graph [3,8] and does not require that there be at least one exit frm a cycle [8l. The structure of processes may be viewed as an AND/OR problem reduction graph in which the process working on the original question is the mot, and rules are problem reduction operators. Partly influenced by Kaplan's producer-consumer model [53, we designed the system so that if a process working on some problem is about to create a process for a subproblem, and there is another process already working on that subproblem, the parentprocess canmake useof the extant process and so avoid solving the same problem again. The method we employ handles recursive rules with no additional mechanism. The structure of processes may be viewed as an active connection graph, but, as will be seen below, the size of the resulting structure need not depend on the number of times a recursive rule will be used. This paper describes hm our system handles recursive rules. Aspects of the system not directly relevant to this issue will be abbreviated orcmitted. Inparticular,details of thematch routine which retrieves formulas unifiable with a given formula will not be discussed (but see [lOI). The Inference Systeq The SNePS inference system builds a graph of processes [7,11] to answer a question (derive instances of a given formula) based on a data base of assertions (ground atomic formulas) and rules (non-at&c formulas). Each process has a set of 151 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. registers which contain data, and each process may send messages to other processes. Since, in this system, the messages are all answers to scnte ques- tion, we will call a process P2 a boss of a process PI if PI sends messages to P2. Sane processes, called data collectors, are distinguish& by two features: 1) they can have store than one boss: 2) they store all massages they have sent to their bosses. The stored messages are used for two purposes: a) it allows the data collector to avoid sending the sama message twice; b) it allows the data collector to be given a new boss, which can inmediately be brought up to date by being given all the messages already sent to the other bosses. Four types of processes are important to the discussion of recursive rules. They are called INFER, CHAIN, SWITCH and FILTER. INFER and CHAIN are data collectors, SWITCH and FILTER are not. Four Processes An INFER process is created to derive instances of a formula, Q. It first matches Q against the data base to find all formulas unifiable with Q. The result of this match is a list of triples, <T,-r,W, where T is a retrieved formula called the target. and 'c and 0 are sub- stitutions called ti?get binding and source binding respectively. -- Essentially -c and o are factorea versions of the mst general unifier (mgu) ofQ and T. Pairs of the rqu whose variables are in Q appear in G, while those whose variables areinTappearinT. Any variable in term position is taken from T. Factoring the q-u obviates the need for renaming variables. For example if Q=P(x,a,y) and *P(b,y,x), we would have o=(b/x,x/y) and T={a/y,x/x) (the pair x/x is included to make our algorithms easier to describe). Note that Qa = T-r = P(b,a,x), the variables in the variable position of the substitution pairs of 0 are all and only the variables in Q, the variables in the variable position of 'c are all and only the variables in T, all terms of cr ce from T, and the non-variables in T came from Q. for For each match Q, there are two <TJ,o> that an INFER finds possiblities we shall consider. First, T might be an -assertion in the data base. In this case, G is an answer (Qcr has been derived). If the INFER has already stored 0, it is ignored. Otherwise, CT is stored by the INFER and the pair <Q,o> is sent to all the INFER's bosses. For CT to be a reasonable answer, it is crucial that all its vai&les occur in Q. The other case we shall con- sider is the one in which T is the consequent of some rule of the form Al&...&An1T. (our system allcrws other forms of rules, but consideration of this one will suffice for explaining how we handle recursive rules). In this case, the INFER creates two other processes, a SWITCH and a CHAIN to derive instances of T-r. The SWITCH is made the CHAIN's boss, and the INFER the SWITCH's boss. It maybe the case that an already extant CHADJ maybeused insteadof anewone. This will be discussed below. The SWITCH process has a register which is set to the source binding, O. The answers it receives from the W are substitutions B, signifying that TTB has been derived. SW'PK!H sends to its boss the application 46. the sub- , stitution derived from o by replacing each term t in o by tB. The effect of the SWITCH is to change the answer from the context of the vari- ables of T to the context of the variables of Q. 1n0u.r example, themmightsendthe answer B=k/xl. SWITCH would then send a\B = {b/x,x/y}\ {c/xl= (b/x,c/y) to the INFER, indicating that Q~\B = P(x,a,y)(b/x,c/y) = P(b,a,c) has been derived. The importance of the factoring of the mguof Q andT into the sourcebinding 0 and the targetbinding T - a separation which the SWITCH repairs -- is thattheCHAIN canwork onT in the context of its original variables and report to many bosses, each through its own SWITCH. A CHAIN process is created to use a particu- lar substitution instance, T, of a particular formula, Al&... GAk1T to deduce instances of T-r. Its answers, which will be sent to a SWITCH, will be substitutions B such that T-rB has been deduced using the rule. For each Ai, ISilk, the CHAIN tries to discover if ANT is deducible by creating an INF'ER process for it. However, an INFER process might already be working on Aia. If ~.=TT, the already extant INFER is just what the CHAIN wants. It takes all the data the INFER has already collected, and adds itself to the INFER's bosses so that itwill also get future answers. If a is more general than -r, the INFER will produce all the data the CHAIN wants, but unwanted data as well. In this case the CHAIN creates a FILTER process to standbetween it and the INFER. The FILTER stores a substitution consisting of those pairs of T for which the term is a constant, and when it receives an answer substitution from the INFER, it passes it along to its CHAIN only if the stored substitution is a subset of the answer. For example, if T were (a/x,y/z,b/w} and a were (u/x ,v/z ,v/wl , a FILTER would be created with a sutstitution of {a/x,b/w), insuring that unwanted answers such as (c/x,d/z,b/w) produced by the more general INFER were filtered out. If a is not compatible with T, or is less general than T, a new INFER must be created. However, if a is less general than T, the old INFER might already have collected answers that the new one can use. These are takenby the new DJFERand senttoits bosses. Also, since thenew INFERwillprcduce all the additional answers that the old one would (plus others), the old INFER is eliminated and its bosses given to the new INFER with intervening FILTEXs. The net result is that the same structure of processes is created regardless of whether the mre general or less general question was asked first. A CHAIN receives answers from INFERS (possibly filtered) in the form of pairs <Ai,Bi> indicating that AiPi, an instanceof the antec&entAi,has beendeduced. Whenever the CHAIN collects a set of consistent substitutions (Bi,...,Bn), one for each antecedent, it sends an answar to its bosses consisting of the ccgnbination of Bl,...,Bk (where the ambination of 61 = ~tll/vll,...,tln,/vlnlI ,...,Bk = ctkl/vkl ,...,tknk/vknkl is the mgu of the expressions (vll,...,vlnl,...,vkl,...,vknk) and (tll ,. . . ,tlnl,. . . ,tkl ,...,tknk) [l, p.1871). 152 Recursive Rules Cause Cycles Just as a CHAIN can make use of already exist- ing INFERS, an INFER can r&e use of already existing ms, filtered if necessary. A recursive rule is a chain of the form Al&... &Ak~Bl,Bl&...&Bn~...X, with C unifiable with at least one of the antecedents, Al say. When an INFER operates on Al, it will find that C matches Al, and it may find that it can use the CHAIN already created for C. Since this CHAIN is in fact the INFER's boss, this will result in a cycle of processes. The cycle will produce more answers as new data is passed around the cycle, but no infinite loop will be created since no data collector sends any answernore thanonce. (If an infinite set of Skolem constants is generated, the process will still terminate if the root goal had a finite number of desired answers specified [ll, p.1941). Figure 1 shows a structure of processes which we consider an active connection graph. It is built to derive instances of ANCESToR(William,w) form the rules vCrx,y) [PARENT(x,y)~ANCESToR(x,y) 1 and V‘(x,y,z) [ANcESToR(x,y) & PAREMT(y,z)x ANCESToR(x,z)]. The notation for the rule instances is similar to thatpresented in [31. Note particularly the SWITCH in the cycle which allows newly derived instances of the goal ANCESToR(William,w) to be treated as additional instances of the antecedent ANCESToR(William,y). A similar structure would be built regardless of the order of asserting the two rules, the order of anteaedents inthetwoantecedentrule, the order of execution of the processes, whether the query had zero, either one, or both variables ground, or if the twoanteceiientruleused ANCESTOR for both antecedents. In the SNePS inference system, recursive rules cause cycles to be built in a graph structure of processes. The key features of the inference system which allow recursive rules to be handled are: I) the processes that produce derivations (INFER and CHAIN) are data collectors: 2) data collectors never send the same answer nore than once: 3) a data collector may report to mre than one boss: 4) a new boss may be assigned to a data collector at any time -- it will miately be given all previously collected data; 5) variable contexts are localized, SWITCH changing contexts dynamically as data flows around the graph: 6) FILTERs allow more general producers to be used by less general consumers. 1. chang, C.-L., and Lee, R.C.-T., Symbolic Logic and Mechanical Theorem Proving, Acadtic Press, New York, 1973. 2. chang, C.-L., and Slagle, J.R., Using rewriting rules for connection graphs to prove theorems, Artificial Intelligence 12, 2 (August 1979), 159-180. 3. chang, C.-L., On evaluation of queries contain- ing derived relations in a relational data base. In normal Bases for Data Bases, Gallaire, H., Minker, J. and Nicolas, J. (edS.), Plenum, New York, i980. 4. Fikes, R.E., and Hendrix G.G., The deduction component. In Understanding Spoken Language, Walker, D.E., ed., Elsevier North-Holland, 1978, 355-374. 5. Kaplan, R.M., A multi-processing approach to natural language understanding. Proc. National Computer Conference, AFIPS Press, Montvale, NJ, 1973, 435-440. 6. Klahr, P., Planning techniques for rule selec- tion in deductive question-answering. In Pattem- Directed Inference Systems, Waterman, D.A., and Hayes-Roth, R., eds., Academic Press, New York, 1978, 223-239. 7. McKay, D.P. and Shapiro, S.C., MULTI--A LISP based muitiprocessing system. Proc. 1980 LISP Conference. Stanford Universitv. 1980. 8. Naqvi,'S.A., and Henschen,-L.J., Performing inferences over recursive data bases. Proc. First AAAI Conference, Stanford Univers- 1980. 9. Reiter, R., On structuring a first order data base, Proc. SecondNational Conference, Canadian Society for Computational Studies of Intelligence, 1978, 50-99. - 10. Shapiro, S.C., Representing andlocatingde- duction rules in a semantic network. Proc. Work- shop on Pattern-Directed Inference Systexns. In SIGART Newsletter, 63 (June 1977), 14-18. 11. Shapiro, S.C., The SNePS sepnanticneixork processing system. In Associative Networks: The Representation and Use of Kncwledse bv Camrsuters. F&dler, N.V., ed., Acadenic Press, NW York, 1979, 179-203. 12. Shortliffe, E-H., Computer Based Medical Con- sultations: MYCIN, &rerican Elsevier, New York, 1976. 13. Sickel, S., A search technique for clause interconnectivity graphs, IEEE Transactions on Computers Vol. C-25, 8 (August 1976) I 823-835. ANCZSMR(William,z) PARETJT(William,y) AN(TlZSMR(William, y) Figure 1. 153

1980

79

77

ON PF0VING LAWS OF TBE ALGEBRA OF FP-SYSTE2% INEDINB~LCF Jacek Leszczykcwski Polish Acadeq of Sciences Institute of Computer Science P.O.BOX 22, 00-901 Warszawa PKiN, PQLMD I INTFODUCX'ION J.Backus, in CACM 21/8, defined a class of ap- plicative prograrrunin g systems called FP /functional prograrmcing/ systems in which a user has: l.objects built recursively frcxn atcms, UU /an undefined ele- ment/ and objects by a strict /i.e. a UU-preserving/ "list" operator, 2. elementary functions over obje- cts, 3. tools for building functions out of already defined functions. One can think of a machine support while work- ing with FP systems and proving facts about FP sys- tems as well as facts concerning the functions be- ing defined. The choice of EDINBURGH KF is rather natural because it is an interactive ccmputer sys- tem /implemented in LISP/ for reasoning about fun- ctions /see &I/. It consists of two parts. The first part is a family of calculi each of which is characterized by four factors: 1. type operators /representing domains in the sense of Scott's theo- ry; see m/, 2. constants /representing continuo- us functions/, 3. axicms, 4. inference rules. One of them, PPLAMBDA, is given as the "initial" calcu- lus, and other calculi may be built by users as e.x- tensions of existing calculi. The second part is a high level prograrmxin g language ML which is fully higher order and is strongly typed. Its polymorphic types make it as convenient as typeless languages. This paper is a short report on the application of EDINBUZH LCF to proving the laws of the algebra of FP systems listed by Backus in [ll. Actualiy, we generalized FP-systems and the laws are formula- ted in stronger fonn than it was done by Backus. We briefly describe /sec.II/ the style of proving with the system, then /sec.III/ ccxtnnent the strtegies u- sed in the proofs giving only their specifications. The summing up remarks will be given in sec.IV. r@- re detailed report on the project is given in D>. II STYLEOFPROVING As mentioned there are inference rules associa- ted with each of the calculi of the system; the in- ferece rules of PPLAMBDA are primitive, and derived rules may be prcgramned by users. The inference ru- les are represented as ML-functions taking theorems /being a data type in ML/ as arguments and giving values which are theorems as well; an example is the computational induction rule INDUCT /it is the Scott induction rule; for more detailes see c21/. We could prove our theorems applying the inference rules in appropriate order but it is not a convenient style of proving. We base our proofs on partial subgoaling meth- ods, called tactics; these mean that given a formula to be proved-t to transform it into "simpler" formulae /which in turn have to be proved/ and the proof justifying the "transformation". The system can support this kind of proving via predefined ty- pes: goal, tactic and proof, defined as follatiJs: goal = form # siqset f form list proof = tlxn list -7 thm tactic = goal -> goal list# proof The first element of the Cartesian product defining the type goal is for stating the formula which is going to be proved, the third one for listing assum- ptions, the second one is /frcm the user's point of view/ an abstract type consisting of simplification rules; these are /possibly conditional/ equivalen- ces of terms to be used as left-to-right rewriting rules. We shall explain new the use of the subgoaling methods. Let us define when a theorem A> f' /A'- hypotheses, f'- conclusion/ achieves a goal f,ss,A. This is the case when, up to renaming of bound va- riables, f' is identical with f and each member of A' is either in A or is in the hypotheses of a member of the simplification set ss . Then, we say, a theorem list achieves a goal list if the first element of the theorem list achieves the first element of the goal list, etc. Thus, a tactic T will work "properly" if for any goal g and any goal list gl and any proof p such that T(g) = gl,p if we have a theorem list thml which achieves the goal list gl , then p(thml) will achieve the goal g - An important special case is when gl is empty for then we need only apply p to the empty theorem list - i.e. evaluate p(ni1) to obtain a theorem achieving our original goal g . We shall use two of the standard tactics. The first is SIMPTAC; applied to any goal (w,ss,A) , SIMPTAC produces a singleton list of goals E(w', ss ,Afl andproof p , where w' is the simplification of w by rewriting rules in ss and p justifies all the simplifications made. In the case where w' is a tautology the goal list is null. The second one is CONDCASESTAC ; for any goal (w,ss,wl) it 84 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. finds a term t of the type tr /truth values do- main/ which is free in w and occurs as the con- ditional expression and then produces three subgo- als with formula w as the their first element and the simplification sets extended by the new assump- tions /caseswhen t equal UU, true and false respectively/. NW we introduce the last tool for desigining proofs: tacticals - mechaninsms for ccxnbining tactics to form larger ones. We shall use only one of the standard tacticals of the system called THEN For any tactic Tl and T2 the canposed tactic Tl THEN T2 applies Tl to the'goal and then applies T2 to all resulting subgoals produced by Tl ; the proof function returned is the ccmposition of the proof functions produced by Tl and T2 . III CQMMENTSONTHEPROOFS There are tree Foups of the proofs of the laws of the FP-system algebra listed by Backus in 09. The first one is based on SIMPTAC ; for exam- ple, to prove II.1 of fl) we used the tactic: AFTAC THEN SIMPTAC where AFTAC is one of the pro gramned tactic and is specified as follms: AFTAC ("u =<v",ss,wi) = I"!X. u x = v x",ss,wl-J , p where: ! - universal quantification = < stands for equality or inequality of terms X- anewvariable The second group of proofs is based on COND- CASESTAC . The ccxrrposed tactic used in the proofs was: APTAC THEN CONDCASESTAC THEN SIMFTAC . For example in proofs of the laws: II.2 and 11.3.1 presented in PI . The third group of proofs involves the use of the ccmputational induction rule INDUCT in an in- derectway. The proofs are based on the programmed tactic INDTAC which is an :'inverse" of the struc- tural induction rule over the type of lists which in turn is derived frcm the ccxnputational induction rule INDUCT. The specification of INDATC is the following: where: X- anew variable ss/ is ss extended by the asswtion w simplified by ss P uses the structural induction ru- le on lists derived form INDUCT nil and cons are the constants over lists. Let us presentoneof the proofs using INIYTAC Suppose we want to prove: "ToALL(G o F)= (ToALL G)o @CALL F) 11 where : 0 - the ccmposition of functions ToAILis takes two arguments:H-being a function and a list and produces a list of the results and the fol- lowing axicms are satisfied: "!H. TbALL H UU = UU II "!H. ToALL H nil = nil It "!H.!X.!L. ToALL H (cons X L)= cons(HX)(ToALLHL)i' . Frcxn the shape of subgoals produced by INDTAC we knm that we need to simplify the formulae by the above axicxns as well as the definition of the ccxnposition operator. Suppose we created the simp- lification set ss including the desired rewriting rules; new we can specify our goal in the following way: g = "ToALL (+G o F)= (Jl?oALL G)o (TbALL F)",ss,nil L&T = APTACTHENINDTACTHENSIMPTAC bea tactic we want to use to tackle our problem. If we apply our tactic T to our goal g we get a pair as result. Let us store the value on the variables P and gl ; we can do it in ML in the following way: gl,p := Tg It terns out that our tactic was fully successful and the system respond that gl is the empty list. Thus I see sec.11, we can apply p to the empty theorem list and the produced value is the theorem we wanted to prove. In general cases when we are not so clever, the subgoal list is not empty and we have to apply another tacit to solve the subgoals. Iv . FINAL- As we mentioned in sec.7: the aim of the paper was to present an application of EDINBURGH ID But the aim of any application itself is to find general tactics which can be used in totally diffe- rent examples. Why? Because EDINBURGH LCF is es- sentially a tool placed samewhere between a theorem prover and a proof checker; this is why we cannot rely on the built in strategies which is the case with the theorem provers but we are interested in looking for general purpose ones which when found canbeprogramnedin ML andanbeused to tackle other goals. The tactic used to prove the laws of the FP-system algebra are of a ganeral use and we- re inwolved in proving properties of the functions defined in FP-systems; see [9J. The generalization 85 of FP systems is briefly described in 12 arad presented in mr>re detailed version in v 9 . Exam- ples of more powerful and ccmplex tactics can be found in t8) . Let us canpare EDINBUXHLCF with other im- plemented systems. On one hand the explicite pre - sence of the logic of the systemmakes DXNBURGH LCF "hL11Tkzn-oriented" and easy to extend. On the other hand, for example, the Bayer-Moore theorem prover /see L3]/ is very efficient in its use of built in strategies, but difficult to extend ; by contrast the need to conduct all inferences through the basic inference rules /as ML-procedu- res/ , which appears necessary if we wish to allaw users to extend the system reliably by prqxxrming leads to sm inefficiency in LCF. This we have found tolerable and indeed it can be reduced sig - nificantly by direct implementation of ML /which at present is ccmpiled into LISP / . Another way of making the system quick is by running it on rrcul- tiprocessor machines which is done for example at Royal Technical University , Stockholm, Sweden . For a nice general ccmparison of these two system see [5J . The EDINBURGH ICF style of proving which con sists in solving the problems by means of program- med proof strategies seems to be natural. It took the athor 2 months to be able to work with the system. This style fits pretty well to doing large proofs with machine assistance. By this we mean neither that a large proof is submitted step by step and merely checked by the machine /see D] /, nor that the system discovers the large proof by itself, but that the problem may be split into sma- ler parts, each of which is tackled semiautcmatica- ly by a subgoaling method. A nice example of such application of EDINBURGH XF is the ccmpiler co- rrectness proof presented in E41 . I wish to thank Awa Cohn, Mike Cordon, Robin Milner and Chris Wadsworth for their friendly help duringmy stay inFdinburgh andespecially Robin Milner for his support while preparing the draft version of the paper. REFERENCES BackusJ. lrCan programming be liberated frcxn the von Neumann style? A functional prcgram- ming style and its algebra of programs", C&-m ACM 21,8 , 1978 . Bird R., "Programs and Machines; an introduc- tion to the theory of camputation", Wiley 1976 Boyer R.S., mre J S., "A ccmputational Logic" Academic Press, New York 1979 . Texas, 1979 . Cohn A., "Remarks on Machine Proof", manusc:- ript, 1980 . Gordon M., Milner R., Wadsworth C., 'EDINBURGH Ia?" , Springer Verlag , 1979 . van Benthem Jutting L.S., "Checking Landau's 'Grundlagen'in the AUTCMATH system', Tech. Hoghschule, Eidhoven, The Netherlands, 1977 . IeszczyXmski J., ~~AI-I experiment with EDINBURGH LCF" , Proc. CAUE-5. , Les Arsc, France , 1980 . Leszczylmski J., "Theory of FP systems in EDINBUIGH LCF", Internal Report, Ccanp. Sci. Dept., Edinburgh University, Edinburgh, Scot- land, 1980 . Milner R., "LCF: a way of doing proofs with a machine", Proc. MFCS-8 Symposium, OlcEnouc, Chechoslcrwakia, 1979 . Milner R., "Implementation and application of Scott's logic for ccxnputable functions", Proc. ACM Conference on Proving Assertions abouT Programs, SIGPLAN Notices , 1972 . Leszczy3xwski J., "EDINBUZH IXF supporting FTI? systems", Proc. Annual Conference of the Geselschaft fur Informatik, Universitat des Saarlandes, 1980 . [41 Cohn A., "High level proof in LCF", Proc. 4th Workshop on Autcmated Deduction , Austin , 86

1980

8

78

Mapping Image Properties into Shape Constraints: Skewed Symmetry, Affine-Transformable Patterns, and the Shape-from-Texture Paradigm John R. Kender and Takeo Kanade Computer Science Department Carnegie-Mellon University Pittsburgh, Pa. 15213 1. Introduction Certain image properties, such as parallelisms, symmetries, and repeated patterns, provide cues for perceiving the 3-D shape from a 2-D picture. This paper demonstrates how we can map those image properties into 3-D shape constraints by associating appropriate assumptions with them and by using appropriate computational and representational tools. We begin with the exploration of how one specific image property, “skewed symmetry”, can be defined and formulated to serve as a cue to the determination of surface orientations. Then we will discuss the issue from two new, broader viewpoints. One is the class of Affine-transformable patterns. It has various interesting properties, and includes skewed symmetry as a special case. The other is the computational paradigm of shape-from-texture. Skewed symmetry is derived in a second, independent way, as an instance of the application of the paradigm. This paper further claims that the ideas and techniques presented here are applicable to many other properties, under the general framework of the shape-from-texture paradigm, with the underlying meta-heuristic of non-accidental image properties. 2. Skewed Symmetry In this section we assume the standard orthographic projections from scene to image, and a knowledge of the gradient space (see [4]). Symmetry in a 2-D picture has an axis for which the opposite sides are reflective; in other words, the symmetrical properties are found along the transverse lines perpendicular to the symmetry axis. The concept skewed symmefry is introduced by Kanade [l] by relaxing this condition a little. It means a class of 2-D shapes in which the symmetry is found along lines not necessarily perpendicular to the axis, but at a fixed angle to it. Formally, such shapes can be defined as 2-D Affine transforms of real symmetries. Figures 2-1 (a)(b) show a few key examples. Stevens [5] presents a good body of psychological experiments which suggests that human observers can perceive surface orientations from figures with this property. This is probably because such qualitative symmetry in the image is often due to Figure 2-1: Skewed symmetry (b) real symmetry in the scene. Thus let us associate the following assumption with this image property: “A skewed symmetry depicts a real symmetry viewed from some unknown view angle.” Note that the converse of this assumption is always true under orthographic projection. We can transform this assumption into constraints in the gradient space. As shown in Figure 2-1, a skewed symmetry defines two directions: let us call them the skewed-symmetry axis and the skewed-transverse axis, and denote their directional angles in the picture by (r and /?, respectively (Figure 2-l(c)). Let G = (p,q) be the gradient of the plane which includes the skewed symmetry. We will show that p’*cos*(~) - q’*sin*(y) = -cos(a-/?) (1) where p’ = pcosX + qsinX q’ = -psinX + qcosX h = (a + /3)/2. Thus, the (p,q)‘s are on a hyperbola. That is, the skewed symmetry defined by (Y and fl in the picture can be a projection of a real symmetry if and only if the gradient is on this hyperbola. The skewed symmetry thus imposes a one-dimensional family of constraints on the underlying surface orientation (p,q). 3. Affine-Transformable Patterns In texture analysis we often consider small patterns (texel= texture element) whose repetition constitutes “texture”. Suppose we have a pair of texel patterns in which one is a 2-D Affine transform of the other; we call them a pair of Affine-transformable patterns. Let us assume that “A pair of Affine-transformable patterns in the picture are projection of similar patterns in the 3-D space (i.e., they can be overlapped by scale change, rotation, and translation)“. Note that, as in the case of skewed symmetry, the coiiverse of this assumption is always true under orthographic projections. The above assumption can be schematized by Figure 3-l. Consider two texel patterns P, and P, in the picture, and place the origins of the x-y coordinates at their centers, respectively. The transform from P2 to P, can be expressed by a regular 2x2 matrix A = (aij). PI and P2 are projections of patterns P’, and P’, which are drawn on the 3-D surfaces. We assume that P’, and P’, are small enough SO that we can regard them as being drawn on small planes. Let us denote the gradients of those small planes by G, = (~1 ,ql) and G2 = (p2,q2), respectively; i.e., P’, is drawn on a plane-z=p,x+q,yandP’20n-z=p2x+q12y. 4 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Now, our assumption amounts to saying that P’, is transformable from .P’, by a scalar scale factor u and a rotation matrix R = (zynsz iz,“d,“,*). (W e can omit the translation from our consideration, since for each pattern the origin of the coordinates is placed at its gravity center, which is invariant under the Affine-transform). Thinking about a pattern drawn on a small plane, -z= px + qy, is equivalent to viewing the pattern from directly overhead; that is, rotating the x-y-z coordinates so that the normal vector of the plane is along the new z-axis (line of sight). For this purpose we rotate the coordinates first bv cp around the y-axis and then by 8 around the x-axis. We have the following relations among v, 8, p, and q: sincp = p/m cosql = 1 /J&i (2) sine =9/&7&i c0se = J;;T;-i/ The plane which was represented as -z= px +qy in the old coordinates is, of course, now represented as -z’ =O in the new coordinates. Let us denote the angles of the coordinate rotations to obtain P’, and P’, in Figure 3-l by (cp, ,8,) and (‘p2,e2), individually. The 2-D mapping from P’i (xl-y’ plane) to Pi (x-y plane) can be conveniently represented by the following 2x2 matrix Tt which is actually a submatrix of the usual 3-D rotation matrix. Ti =( cyo -Si~W~“) Now, in order for the schematic diagram of Figure 3-1 to hold, what relationships have to be satisfied among the matrix A = (at)), the gradients Gi = (pilqi) for i = 1,2, the angles (pi, ei) for i = 1,2, the scale factor u, and the matrix R ? We equate the two transforms that start from P’, to reach at P,: one following the diagram counter-clockwise P’,->P*->P, , the other clockwise P’,->P’,->P,. We obtain AT2 = T, aR. By eliminating u and a, and substituting for sines and cosines of pi and Bi by (2), we have two (fairly complex) equations in terms of pi, qi, and the elements of A. We therefore find that the assumption of Affine-transformable patterns yields a constraint determined solely by the matrix A. The matrix is determined by the relation between P, and P, observable in the picture: without CT/?= CT ( coti -5ind Sind ~0x4 > Figure 3-1: A schematic diagram showing the assumptions on Affine transformable patterns. knowing either the original patterns relationships (a and R) in the 3-D space. (P’, and P’*) or their The Affine transform from P, to P, is more intuitively understood by how a pair of perpendicular unit-length vectors (typically along the x and y coordinate axes) are mapped into their transformed vectors. Two angles (a and /I) and two lengths (T and p) can characterize the transform. Components of the transformation matrix A = (aij) are represented by: a,, =7cosa a12=Pcos/? (3) a21 = Tsina a2* = psi@ Let us consider the case that a and p are known, but T and p are not. Using (3), eliminate T and p. Then, we obtain (Pl cosa + q 1 sinaNp, cosp + q, sir@) + cos(a-P) = 0 which is exactly the same as the hyperbola (1). . The Shape-from-Texture Paradigm This section derives the same skewed-symmetry constraints from a second theory, different from the Affine-transformable patterns. The shape-from-texture paradigm is briefly presented here; a futler discussion can be found in [3]. The paradigm has two major portions. In the first, a given image textural property is “normalized” to give a general class of surface orientation constraints. In the second, the normalized values are used in conjunction with assumed scene relations to refine the constraints. Only two texels are required, and only one assumption (equality of scenic texture objects, or some other simple relation) to generate a well-behaved one-dimensional family of possible surface orientations. The first step in the paradigm is the normalization of a given texel property. The goal is to create a normalized texture property map (NTPM), which is a representational and computational tool relating image properties to scene properties. The NTPM summarizes the many different conditions that may have occurred in the scene leading to the formation of the given texel. In general, the NTPM of a certain property is-a scalar-valued function of two variables. The two input variables describe the postulated surface orientation in the scene (top-bottom and left-right slants: (p,q) when we use the gradient space). The NTPM returns the value of the property that the textural object would have had in the scene, in order for the image to have the observed textural property. As an example, the NTPM for a horizontal unit line length in the image summarizes the lengths of lines that would have been necessary in 3-D space under various orientations: at surface orientation (p,q), it would have to be m. More specifically, the NTPM is formed by selecting a texel and a texel property, back-projecting the texel through the known imaging geometry onto all conceivable surface orientations, and measuring the texel property there. In the second phase of the paradigm, the NTPM is refined in the following way. Texels usually have various orientations in the image, and there are many different texel types. Each texel generates its own image-scene relationships, summarized in its NTPM. If, however, assumptions can be made to relate one texel to another, then their NTPMs can also be related; in most cases only a few scenic surface orientations can satisfy both texels’ requirements. Some examples of the assumptions that relate texels are: both lie in the same plane, both are equal in textural 5 measure (length, area, etc.), one is k times the other in measure, etc. Relating texels in this manner forces more stringent demands on the scene. If enough relations are invoked, the orientation of the local surface supporting two or nrore related texels can be very precisely determined. What we now show is that the skewed symmetry method is a special case of the shape-from-texture paradigm; it can be derived from considerations of texel slope. To normalize the slope of a texel, it is back-projected onto a plane with the postulated orientation. The angle that the back-projected slope makes with respect to the gradient vector of the plane is one good choice (out of many) for the normalized slope measure. Under perspective, the normalized value depends on the image position and camera focal length; under orthography it is much simpler. Using the construction in Figure 4-1, together with several lemmas relating surfaces in perspective to their local vanishing lines, slope is normalized as follows. Assume a slope is parallel to the p axis; the image and gradient space can always be rotated into such a position. (If rotation is necessary, the resulting NTPM can be de-rotated into the original position using the standard two-by-two orthonormal matrix.) Also assume that the slope is somewhere along the line y = ys, where the unit of measurement in the image is equal to one focal length. Then, the normalized slope value--the normalized texture property map -- is given by Es - Ys(P2 + cl211 ~dtl + P2 + s*>1. This normalized value can be exploited in several ways. Most important is the result that is obtained when one has two slopes in the image that are assumed to arise from equal slopes in the scene. Under this assumptions, their normalized property maps can be equated. The resulting constraint, surprisingly, is a simple straight line in the gradient space. Under orthography, nearly everything simplifies. The normalized slope of a texel becomes q / [P/(1 f P2 + q*>1. (4) lt is independent of Y,; in effect, all slopes are at the focal point. Consider Figure 2-1 (a). Given the angle that the two texels form (/?-a), rotate the gradient space so that the positive p axis bisects the angle. Let each half-angle be 6, so 6 = (/I-a)/2. Calculating the normalized value of either slope is obtained directly from the standard normalized slope formula, corrected for the displacement of + S and -6 respectively. That is, for the slope at the positive 6 orientation, instead of formula (4), we use the Figure 4-1: Back-projecting an image slope onto a plane with gradient (p, q). formula under the substitution pcosS + qsin6 for p, for q. We do similarly for the slope at -8. -psin6 qcosd The fact that the normalized slopes are assumed to be perpendicular in the scene allows us to set one of the normalized values equal to the negative reciprocal of the other. The resultant equation becomes p2cos2&q2sin26 = sin*&cos*S = -cos26. This is exactly the hyperbola (1) with 26 =/l-a. 5. Conclusion The assumptions we used for the skewed symmetry, the Affine-transformable patterns, and texture analysis can be generalized as “Properties observable in the picture are not by accident, but are projections of some preferred corresponding 3-D properties.” This provides a useful meta-heuristic for exploiting image properties: we can call it the meta-heuristic of non-accidental image properties. It can be regarded as a generalization of general view directions, often used in the blocks world, to exclude the cases of accidental line alignments. Instances that can fall within this meta-heuristic includes: parallel lines in the picture vs. parallel lines in the scene, texture gradients, and lines convergent to a vanishing point. One of the most essential points of our technique is that we related certain image properties to certain 3-D space properties, and that we map the relationships into convenient representations of shape constraints. We explicitly incorporate assumptions based either on the meta-heuristic or on apriori knowledge of the world. The shape-from-texture paradigm provides a computational framework for our technique. In most part of our discussion we assumed orthography. Similar--though more involved and less intuitive--results can be obtained under perspective projections. This work is further same title as this paper. .discussed in a technical report with PI [31 [41 6 References Kanade, T. Recovery of the 3-Dimensional Shape of an Object from a Single View. Technical Report CMU-CS-79-153, Computer Science Department, Carnegie-Mellon University, Oct., 1979. Kender, J.R. Shape from Texture. PhD thesis, Computer Science Department, Carnegie-Mellon University, 1980. Mackworth, A.K. Interpreting Pictures of Polyhedral Scenes. Artificial intelligence 4(2), 1973. Stevens, K.A. Surface Perception from Local Analysis of Texture and Contour. Technical Report AI-TR-512, MIT-AI, 1980.

1980

80

79

WHAT SHOULD BE COMPUTED IN LOW LEVEL VISION SYSTEMS William B. Thompson Albert Yonas University of Minnesota Minneapolis, Minnesota 55455 ABSTRACT Recently, there has been a trend towards developing low level vision models based on an analysis of the mapping of a three dimensional scene into a two dimensional image. Emphasis has been placed on recovering precise metric spatial information about the scene. While we agree with this approach, we suggest that more attention be paid to what should be computed. Pschophysical scaling, adaptation, and direct determination of higher order relations may be as useful in the per- ception of spatial layout as in other perceptual domains. When applied to computer vision systems, such processes may reduce dependance on overly specific scene constraints. L* Introduction The following is a position paper directed at several aspects of low-level visual processing. The current trend towards focusing on the determi- nation of exact, three-dimensional form in a scene is questioned. Both analysis of representative scene domains and experience with human vision sug- gest that less precise form properties may be suf- ficient for most problems. Several computational issues are also briefly discussed. 2. Alternate Approaches to "Low-Level" Analysis - -- Computer vision systems have traditionally been divided into segmentation and interpretation components. A multiplicity of image features have been investigated in the hope that they would facilitate the partitioning of an image into regions corresponding to "objects" or "surfaces" in the original scene. Only after this two- dimensional segmentation operation was completed would procedures be applied in an attempt to deter- mine the original three-dimensional structure of the scene. Recently, an alternative approach to implementing the lower levels of a computational vision model has been developed. Its basic premise This research was supported in part by the Nation- al Science Foundation under Grant MCS-78-20780 and by the National Insitute of Child Health and Human Development under Grants HD-01136 and HD-05027. is that the determination of three-dimensional structure is such an integral part of the scene description processes that it should be carried out at all levels of the analysis [1,2,3,4,51. Proponents of this approach usually employ a well structured methodology for developing computational models of form perception: Precisely describe a carefully con- strained scene domain. Identify important scene properties. Determine the function which maps these scene properties into an-image. Develop computationally feasible mechan- isms for recovering the "important" scene properties from the image. Great emphasis is placed on determining what scene properties are computable, given a set of con- straints on the scene. Scene properties normally considered essential to the analysis include object boundaries, three- dimensional position, and surface orientation. In many cases, the determination of these features requires that properties such as surface reflec- tance and illumination must also be found. A key distinction between the classical techniques and this newer approach is that in the latter, analysis procedures are developed analytically from an understanding of how scene properties affect the image rather than ad hoc assumptions about how image properties might relate to scene structure. The representational structures which have been used to implement form based analysis have, for the most part, been iconic. The features represented are almost always metric properties of the corresponding point on the surface: distance from the observer, orientation with respect to either the observer or a ground plane, reflectance, incident illumination, and so on. To determine relative effects (eg. which of two points is farther away), absolute properties are compared. The determination of these metric scene pro- perties requires that the possible scenes be highly constrained. Usually, the analysis depends on res- trictions both on the types of objects allowed and on the surface properties of the objects. For From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. example, a "blocks world" assumption (or alter- nately, the assumption of a "Play-Doh" world made entirely of smooth surfaces) might be made. In addition, it is commonly assumed that surfaces are all lambertian reflectors and that, for a given surface, the coefficient of reflectance is constant. Illumination is often limited to a sin- gle distant point source, possibly coupled with a diffuse illuminator. Secondary illumination effects are usually presumed to be negligible. 20 Absolute Scene Properties Are Not Always Needed -m The proponents of form based analysis presume the need for finding exact shape properties of a scene. They concentrate on investigating how con- straints on scenes affect what properties are com- putable and how they can be determined. We suggest that more attention be paid towards what properties should be computed. We argue that for a wide variety of problem areas, absolute metric informa- tion about scene shape is not required. Instead, relative properties such as flat/curved, convex/concave, farther-away/closer, etc. are both sufficient and easier to compute. Most tasks involving description of a visual environment depend on generalized shape properties. In fact, much effort has been spent searching for shape characterizations that embody those relation- ships useful for description but not the enormous amount of irrelevant detail contained in any representation based on specific position. Even in task domains such as object manipulation and obsta- cle avoidance, precise positional information is frequently not necessary. Both these task areas contain significant sub-problems involving object identification - a descriptive task often possible with approximate and/or relative information about shape. Even when actual position is needed, feed- back control can be used to minimize the need for highly accurate positional determinations. A second argument for emphasizing the diffi- culty of determining metric properties comes from our experience with human perception. The psycho- logical literature contains many references to the effects of the scaling process that relates the physical domain to the psychological [6,71, the effects of adaptation to stimulation [8], and the effects of practice on variable error [91. By investigating the competence of the human visual system in determining primitive shape effects, we can gain insight into sufficient (but not neces- sary) properties for more complex analysis. In our own work on perception of surfaces, preliminary results from one set of experiments seem relevant to the development of computational models. We synthesized a frontal view of a surface the profile of which is shown in figure 1. Lighting was assumed to be a combination of a single distant point source and a perfectly diffuse source. A simple reflectance model was used and secondary illumination effects were not considered. A series of synthesized images was produced with the inten- tion of examining the perception of single displays and the ability to determine differences between displays. The "object" in our images was an ellip- soid with semi-axes A, B, and C. (A was in the horizontal direction as seen by the viewer, B was in the vertical direction, and C was in the direc- tion along the line of sight.) The object was presented against a black background, and thus no cast shadows were present. In one set of experi- ments, A and B were held constant producing a cir- cular occluding contour. Subjects were asked to estimate the value of C for a number of different displays, with true values of C ranging from one half of A to four times A. On initial trials, sub- jects tended to see the same shape independently of the actual value of C. On subsequent trials, per- formance improved, but with a significant, sys- tematic underestimation of the true value. As a final note, when subjects were asked to qualita- tively describe the changes in the scene as C was varied, they often indicated that they felt that the change was due to differences in illumination, not shape. It is certainly premature to make any defini- tive conclussions from our results. Nevertheless, we suggest the following conjecture: Subjects appear to see a specific shape (as opposed to sim- ply a "round" object); however, the metric proper- ties they estimate for that shape are not neces- sarily consistent with the "true" values. The sub- jects do appear to be better at ranking displays based on different values of C. 4* --- Non-metric Scene Properties We suggest that requiring specific, accurate determination of scene properties may be unneces- sarily restrictive. Less precise and/or purely relative qualities are sufficient for many situa- tions. By concentrating on these characteristics, we may be able to significantly relax the con- straints under which our computational vision models must operate. Finally, human vision is often quite inaccurate in determining metric values for these same properties. Rather than indicating a deficiency in human vision, this suggests that alternative (and presumably more useful) charac- teristics are being computed by the human per- ceiver. Two approaches to structuring computer vision models based on these observations seem relevant. First of all, it may be possible to directly com- pute properties of interest, rather than deriving them from more "primitive" characteristics (see [lO,lll). For example, we might look for ways of estimating surface curvature that do not depend on first determining depth and then taking the second derivative. A second possibility is to presume that esti- mation of shape properties is subject to the same scaling processes as most other perceptual phenomena. Thus, our model would estimate some non-linear but monotonic transformation of charac- teristics such as depth. The transformations would be adaptive, but in general not known by higher level analysis procedures. Thus, the precise metric three-dimensional structure can not be recovered. For many tasks, the scaled values are sufficient and the need for highly constrained, 8 photometric analysis of the image is reduced. With luminence gradient, illumination direction, and appropriate standardization, precise scene proper- surface curvature. For a given gradient, knowing ties may be determined. Without standardization, either illumination or curvature allows determina- relative characteristics are still computable. tion of the other. The model must be able to Ordinal relationships are detrminable over a wide account for this symmetry. range while quantatative comparisons are possible over a more limited range. (eg. it may be possible 6. Conclusions to judge that A is twice as far as B but not that C is 100 times as far as D.) When attempting to construct computational models of low-level vision systems, we need to pay 50 Computational Models as much attention to what should be computed as we do to how it is computed. We may investigate this Recently, much attention has been focused on problem in at least three ways. The first is a using parallel process models to specify the compu- computational approach: we can determine what is tational structure of low-level vision systems. An computable given a set of constraints about the image is partitioned into a set of neighborhoods, scene and the imaging process. The second is an with one process associated with each region. The ecological approach: we catalog the range of prob- processes compute an estimate of scene properties lem domains in which our system is expected to corresponding to the region using the image function and then determine the primitive scene features in the region and whatever is known about properties needed for analysis. The third is meta- surrounding scene structure. The circularity of phorical: study a working visual system (eg. form estimation for one point depending on the form human) in order to determine which low-level scene of neighboring points can be dealt with in several properties it is able to perceive. These proper- ways. A variable resolution technique may be ties then define a sufficient set for analysis. employed. First large, non-interacting neighbor- hoods are used. Then, progressively smaller neigh- Much current work focuses on estimating exact borhoods are used, each depending on scene proper- positional information about a scene* We argue ties computed using previously analyzed, larger that in many cases, these metric properties cannot regions. (Marr's stereo model is an example [121.) be easily determined. Even more importantly, how- Alternately, an iterative technique can be used to ever, they often need not be determined. Simple find crude estimates of scene properties and then relative properties may be sufficient for analysis those values are fed back into the process to pro- and be much easier to compute. duce more refined estimates. (Examples include many "relaxation labeling" applications [131.) In either case, the determination of absolute scene properties usually requires a set of boundary values - image points at which the scene con- BIBLIOGRAPHY straints allow direct determination of the proper- ties. The computational process must then pro- pagate these constraints to other image regions. [II D. Marr, "Representing and computing visual The robustness of these parallel process information", Artificial Intelligence: An MIT -- models may be significantly increased if they are Perspective, P.H. Winston and R.H. Brown, only required to compute relative properties. The ed-, PP. 17-82, 1979. need for accurately propagating scene information is greatly reduced. Furthermore, photometric [21 H.G. Barrow and J.M. Tennenbaum, "Recovering analysis of the image will usually not be required. intrinsic scene characteristics from images," For instance, general characteristics of the inten- in Computer Vision Systems, A.R. Hanson and sity gradient may be all that is required for E-M. Riseman, eds., New York: Academic Press, analysis. As an example, for a reasonably general 1978. class of scene types, a discontinuity in the luminence gradient will usually correspond to a [31 B. Horn, "Obtaining shape from shading shadow, an occlusion boundary, or the common boun- information," in The Psychology of Computer dary between two surfaces. Continuous but non-zero Vision, P.H. Winston, ed., New York= McGraw- gradient values indicate either surface curvature Hill, 1975. or illumination variation. In neither case is the actual magnitude of the gradient required. [41 S. Ullman, The Interpretation of Visual Motion, Cambridge: MIT Press, 1979: Finally, many of the problems in low-level vision are underspecified. No single "correct" [51 K-A. Stevens, "Surface perception from local solution exists because insufficient information is analysis of texture and contour", Ph.D. available to derive the original scene properties. Thesis, MIT, Feb. 1979. Thus, computational models must either naturally embody default assumptions or allow for ambiguous 161 GOT. Fechner, Elemente der Psychophysik, representations. (There is reason to expect that Leipzig: Breitkopf and Hartel, 1860. both approaches are useful.) Even more important, (Reissued Amsterdam: Bonset, 1964.) the control structures used by the models must not impose any arbitrary input/output assumptions. For example, consider again the relationship between 9 [73 S.S. Stevens, "Perceptual magnitude and its measurement," in Handbook of Perception, &. II, Psychophysical JudgemeT;;& Measurement, Carterette and Friedman, eds., New York: Academic Press, 1974. [al H. Helson, Adaptation Level Theory, New York: Harper, 1964. Dl E.J. Gibson, Perceptual Learning and Develop- ment, New York: Appleton-CenGy-Crofts, 1969. [lo] J.J. Gibson, The Senses Considered as Visual -- Systems, Boston: Houghton Mifflin, 1966. [ll] J.J. Gibson, The Ecological Approach to Visual Percept=, Boston: Houghton Miffli. 1979. [121 DO Marr and TO Poggio, "A theory of human stereo vision," MIT AI Lab. MEMO 451, Nov. 1977. [131 A. Rosenfeld, R. Hummel, and S. Zucker, "Scene labeling by relaxation operations," IEEE Trans. Systems, &, PP and Cybernetics, vol. 6, ppe 420-433, June 19x .’ 1, - . 'i 7 0 . /I I' P Figure 1. 10

1980

81

80

INTERPRETING LINE DRAWINGS AS THREE-DIMENSIONAL SURFACES Harry G. Barrow and Jay M. Tenenbaum Artificiai Inteiiigence Center SRI Internationai, Menio Park, CA 94025 ABSTRACT We propose a computationai modei for interpreting iine drawings as three-dimensionai surfaces, based on constraints on iocai surface orientation aiong extremai and discontinuity boundaries. Specific techniques are described for two key processes: recovering the three- dimensionai conformation of a space curve (e.g., a surface boundary) from its two-dimensionai projection in an image, and interpoiating smooth surfaces from orientation constraints aiong extremai boundaries. INTRODUCTION Our objective is the deveiopment of a computer modei for interpreting two-dimensionai iine drawings, such as Figure 1, as three-dimensionai surfaces and surface boundaries. Line drawings depict intensity discontinuities at surface boundaries, which, in many cases, are the primary source of surface information avaiiabie in an image: i.e., in areas of shadow, compiex (secondary) iiiumination, or specuiar surfaces anaiytic photometry is inappropriate. Understanding how iine drawings convey three- dimensionaiity is thus of fundamentai importance. Given a perspectiveiy correct iine drawing depicting discontinuities of smooth surfaces, we desire as output arrays containing vaiues for orientation and reiative range at each point on the impiied surfaces. This objective is distinct from that of eariier work on interpretation in terms of object modeis (e.g. [I]) and more basic. No knowiedge of piants is required to understand the three-dimensionai structure of Figure 1, as can be demonstrated by viewing fragments out of context (through a mask, for exampie). Ambiguity and Constraints The centrai probiem in perceiving iine drawings is one of ambiguity: in theory, each two- dimensionai iine in the image couid correspond to a possibie projection of an infinitude of three- dimensionai space curves (see Figure 2). Yet peopie are not aware of this massive ambiguity. When asked to provide a three-dimensionai interpretation-of an eiiipse, the overwheiming response is a tiited circie, not some bizarrely twisting curve (or even a discontinuous one) that has the same image. What assumptions about the scene and the imaging process are invoked to constrain to this unique interpretation? * This research was supported by funds from DARPA, NASA, and NSF. We observe that aithough aii the iines in Figure 1 iook fundamentaiiy aiike, two distinct types of scene event are depicted: extremai boundaries (e.g., the sides of the vase), where a surface turns smoothiy away from the viewer, and discontinuity boundaries (e.g., the edges of the ieavss), where smooth surfaces terminate or interswt. Each type provides different constraints on three-dimensionai interpretation. At an extremai boundary, the surface orientation can be inferred exactiy; at every point aiong the boundary, orientation is normai to the iine of si ht and to the tangent to the curve in the image d. A discontinuity boundary, by contrast, does not dirsctiy constrain surface orientation. However, its iocai curvature in the image does provide a statisticai constraint on the three- dimension& tangent of the corresponding space curve. The iocai surface normai is constrained oniy to be orthogonai to this tangent, and is thus free to swing about it as shown in Figure 3. The abiiity to infer 3-D surface structure from extremai and discontinuity boundaries suggests a three-step modei for iine drawing interpretation, anaiogous to those invoived in our intrinsic image modei [2]: iine sorting, boundary interpretation, and surface interpoiation. Each iine is first ciassified according to the type of surface boundary it represents (i.e., extremai versus discontinuity). Surface contours are interpreted as three-dimensionai space curves, providing reiative 3-D distances aiong each curve; iocai surface normais are assigned aiong the extremai boundaries. Finaiiy, three-dimensionai surfaces consistent with these boundary conditions are constructed by interpoiation. (For an aiternative model, see Stevens [3].) This paper addresses some important aspects of three-dimensionai recovery and interpoiation (see [l] and [4] for approaches to iine ciassification). INTERPRETATION OF DISCONTINUITY BOUNDARIES To recover the three-dimensionai conformation of a surface discontinuity boundary from its image, we invoke two assumptions: surface smoothness and generai position. The smoothness assumption impiies that the space curve bounding a surface wiii aiso be smooth. The assumption that the scene is viewed from a generai position impiies that a smooth curve in the image resuits from a smooth curve in space, and not from an accident of viewpoint. In Figure 2, for exampie, the sharpiy receding curve projects into a smooth eiiipse from oniy one viewpoint. Thus, such a curve wouid be a highiy improbabie three-dimensionai interpretation of an eiiipse. 11 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. The problem now is to determine which smooth space curve is most iikeiy. For the speciai case of a wire curved in space, we conjectured that, of aii projectiveiy-equivaient space curves, humans perceive that curve having the most uniform curvature and the ieast torsion [2]; i.e., they perceive the space curve that is smoothest and most planar. Consistent findings were reported in recent work by Witkin [5] at MIT on human interpretation of the orientation of pianar ciosed curves. Measures of Smoothness - The smoothness of a space curve is expressed quantitativeiy in terms of intrinsic characteristics such as differentiai curvature (k) and torsion (t), as weii as vectors giving intrinsic axes of the curve: tangent (T), principai normai (N), and binormai (B). A simpie measure for the smoothness of a space curve is uniformity of curvature. Thus, one might seek the space curve corresponding to a given image curve for which the integrai of k' (the spatiai derivative of k) was minimum. This aione, however, is insufficient, since the integrai of k' couid be made arbitrariiy smaii by stretching out the space curve so that it approaches a twisting straight iine (see Figure 4). Uniformity of curvature aiso does not indicate whether a circuiar arc in the image shouid correspond to a 3-D circuiar arc or to part of a heiix. A necessary additionai constraint in both cases is that the space curve corresponding to a given image curve should be as pianar as possibie, or more preciseiy that the integrai of its torsion shouid aiso be minimized. Integrai 1 expresses both the smoothness and pianarity of a space curve in terms of a singie, iocaiiy computed differentiai measure d(kB)/ds: d(kB/ds)*ds = (kg2 + k2t2)ds (1) Intuitiveiy, minimizing this integrai corresponds to finding the three-dimensionai projection of an image curve that most cioseiy approximates a pianar, circular arc, for which k' and t are both everywhere zero. Recovery Techniques A computer modei of this recovery theory was impiemented to test its competence. The program accepts a description of an input curve as a sequence of two-dimensionai image coordinates. Each input point, in conjunction with an assumed center of projection, defines a ray in space aiong which the corresponding space curve point is constrained to iie. The program can adjust the distance associated with each space curve point by siiding it aiong its ray iike a bead on a wire. From the resuiting 3-D coordinates, it can compute iocal estimates for curvature k, intrinsic axes T, N, and B, and the smoothness measure d(kB)/ds. An iterative optimization procedure then adjusts distance for each point to determine the configuration of points that minimize the integrai in (1). The program was tested using input coordinates synthesized from known 3-D space curves so that resuits couid be readiiy evaiuated. Correct 3-D interpretations were produced for simpie open and ciosed curves such as an eiiipse, which was interpreted as a tiited circle, and a trapezoid, which was interpreted as a tiited rectangie. However, convergence was siow and somewhat dependent on the initiai choice of z-vaiues. For exampie, the program had difficuity converging to the "tiited-circie" interpretation of an eiiipse if started either with aii z-vaiues in a piane paraiiei to the image piane or aii randomized to be highiy nonpianar. SURFACE INTERPOLATION Given constraints on orientation aiong extremai and discontinuity boundaries, the next probiem is to interpoiate smooth surfaces consistent with these boundary conditions. The probiem of surface interpoiation is not pecuiiar to contour interpretation, but is fundamentai to surface reconstruction, since data is generaiiy not avaiiabie at every point in the image. We have impiemented a soiution for an important case: the interpoiation of approximateiy uniformiy-curved surfaces from initiai orientation values and constraints on orientation [6]. The input is assumed to be in the form of sparse arrays, containing iocai estimates of surface range and orientation, in a viewer-centered coordinate frame, ciustered aiong the curves corresponding to surface boundaries. The desired output is simpiy fiiied arrays of range and surface orientation representing the most iikeiy surfaces consistent with the input data. These output arrays are anaiogous to our intrinsic images [2] or Marr's 2.5D sketch [7]. 12 For any given set of input data, an infinitude of possibie surfaces can be found to fit arbitrariiy weii. Which of these is best (i.e., smoothest) depends upon assumptions about the nature of surfaces in the worid and the image formation process. For exampie, surfaces formed by eiastic membranes (e.g., soap fiims) are constrained to minimum energy configurations characterized by minimum area and zero mean curvature; surfaces formed by bending sheets of ineiastic materiai (e.g., paper or sheet metai) are characterized by zero Gaussian curvature; surfaces formed by many machining operations (e.g., pianes, cyiinders, and spheres) have constant principai curvatures. Uniformiy Curved Surfaces We concentrate here on surfaces that are iocaiiy sphericai or cyiindricai (which have uniform curvature according to any of the above criteria). These cases are important because they require reconstructions that are symmetric in three dimensions and independent of viewpoint. Many simpie interpoiation techniques faii this test, producing surfaces that are too fiat or too peaked. An interpoiation aigorithm that performs correctiy on sphericai and cyiindricai surfaces can be expected to yieid reasonabie resuits for arbitrary surfaces. Our approach expioits an observation that components of the unit normai vary iineariy across the images of surfaces of uniform curvature. Consider a three-dimensionai sphericai surface, as shown in Figure 5. The radius and normai vectors are aiigned, and so from simiiar figures we have: Nx = x/R, Ny = y/R, Nz = z/R . A simiiar derivation for the right circuiar cyiinder is to be found in [6]. The point to be noted is that for both the cyiinder and the sphere, Nx and Ny are iinear functions of x and y, and Nz can be derived from Nx and Ny. An Interpoiation Technique - We have impiemented an interpoiation process that expioits the above observations to derive the orientation and range over a surface from boundary vaiues. It uses paraiiei iocai operations at each point in the orientation array to make the two observabie components of the normai, Nx and Ny, each vary as iineariy as possibie in both x and y. This couid be performed by a standard numericai reiaxation technique that repiaces the vaiue at each point by an average over a two-dimensionai neighborhood. However, difficuities arise near surface boundaries where orientation is discontinuous. We decompose the two-dimensionai averaging process into severai one-dimensionai ones, by considering a set of iine segments passing through the centrai point, as shown in Figure 6a. Aiong each iine we fit a iinear function, and thus estimate a corrected vaiue for the point. The independent estimates produced from the set of iine segments are then averaged. Oniy the iine segments that do not extend across a boundary are used: in the interior of a region, symmetric iine segments are used (Figure 6a) to interpoiate a centrai vaiue; at boundaries, an asymmetric pattern aiiows vaiues to be extrapoiated (Figure 6b). The interpoiation process was appiied to test cases in which surface orientations were defined around a circuiar outiine, corresponding to the extremai boundary of a sphere, or aiong two paraiiei iines, corresponding to the extremai boundary of a right circuiar cyiinder. Essentiaiiy exact reconstructions were obtained, even when boundary vaiues were extremeiy sparse or oniy partiaiiy constrained. Resuits for other smooth surfaces, such as eiiipsoids, seemed in reasonabie agreement with human perception. Current work is aimed at extending the approach to partiaiiy constrained orientations aiong surface discontinuities, which wiii permit interpretation of generai soiid objects. REFERENCES 1. K. Turner, "Computer Perception of Curved Objects Using a Teievision Camera," Ph.D. thesis, Department of Machine Inteiiigence and Perception, University of Edinburgh, Edinburgh, Scotiand (1974). 3. K. Stevens, "Constraints on the Visuai Interpretation of Surface Contours," A.I. Memo 522, M.I.T., Cambridge, Massachusetts (March 1979). 4. I. Chakravarty, "A Generaiized Line and Junction Labeiing Scheme with Appiications to Scene Anaivsis." IEEE Transactions on Pattern ” , -- Anaiysis and MachTInteiiigence, Voi. PAMI- 1, No. 2,(Aprii,79). 59 A. Witkin, Department of Psychoiogy, M.I.T., Cambridge, Massachusetts (private communication). 6. H. G. Barrow and J. M. Tenenbaum, "Reconstructing Smooth Surfaces from Partiai, Noiss Information," Proc. ARPA Image Understanding Workshs.S.C.,LosAngeies, Caiifornia (Fam. 7. D. Marr, "Representing Visuai Information," in Computer Vision Systems, A. Hanson and E. Riseman, eds.(Academic Press, New York, New York, 1978). 13 FIGURE 1 LINE DRAWING OF A THREE-DIMENSIONAL SCENE (Surface and boundary structure are dis- tinctly perceived despite the ambiguity inherent in the imaging process.) FIGURE 2 THREE-DIMENSIONAL CONFORMATION OF LINES DEPICTED IN A LINE DRAWING IS INHERENTLY AMBIGUOUS (All of the space curves in this figure project into an ellipse in the image plane, but they are not all equally likely FIGURE 3 AN ABSTRACT THREE-DIMENSIONAL SURFACE CONVEYED BY A LINE DRAWING (Note that surface orientation is constrained to one degree of freedom along discontinuity boundaries.) FIGURE 4 AN INTERPRETATION THAT MAXIMIZES UNIFORMITY OF CURVATURE y AXIS x AXIS FIGURE 5 LINEAR VARIATION OF N ON A SPHERE (a) symmetric (b) asymmetric FIGURE 6 LINEAR INTERPOLATION OPERATORS interpretations.) 14

1980

82

81

SHAPE ENCODING AND SUBJECTIVE CONTOURS’ Mike Brady, W. E. L. Grimson Artificial Intelligence Laboratory, MIT. and D. J. Langridge Division of Computing Research, CSIRO, Canberra, ACT, Australia. 1. Int reduction Ullman [15] has investigated the shape of subjective contours (see for example [7]. [4]. [5], (121). In fact, the work is more generally ap- plicable to other cases of pcrccptual shape completion, in which the visual system is not constrained by actual physical intensity changes. Examples include patterns foimcd from dots line drawings and alphabetical characters. and incomplctcly Ullman proposes that subjective contc,.us consist of two circles which meet smoothly and which arc tangential to the contrast bound- aries from which they originate. The foim of the solution derives from a number of premises, one of which Ullman calls “the locality hypothesis”. ‘Ihis is “based in part on cxpcrimcntal obscrvntions, and partly on a theoretical consideration” ([I 51 ~2). ‘I’hc “cxpcrimcntal ob- servation” rcfcrrcd to is the following: suppose that A’ is a point near A on the filled-in contour AB as shown in Figure 1. If the process by which AB was constructed is applied to A’B, it is claimed that it gcneratcs the portion of AJ3, bctwccn A’ and B. Let us call this property “cxtcnsibility”. Ullman argues that cxtcnsibility, togcthcr with the propcrtics of isotropy, smoothness, and having minimal integral cur- vature, logically entails a solution consisting of two circles which meet smoothly. In the fit11 version of this paper, we analyze the two-cti solution and formulate the condition for minimal integral curvature. This can bc solved by any descent method such as Newton-Raphson. A program has been written which computes the minimum integral curva- ture two-circle solution given the boundary angles 4, 8, and AB, and which returns as a result the point at which they meet and at which the curvature is discontinuous (the ‘knot point’). A tortuous version of this simple proof and program recently appcarcd in [13]. We then show by example that the two circle solution is not in fact extensible. Figure 1. Ullman’s extensibility property. If the process by which AD was constructed is applied to A’B, portion of AB between A’ and B. it is claimed that it generates the lntcrestingly, Knuth [S]. in his discussion of mathematical typog raphy, initially proposes that “most plcasing curves” should be ex- tensible, isotropic, cyclically symmetric (that is, the solution should not depend on the order of the contributing points), smooth, and be generable by a local (in his case four point) process. In fact, Knuth ([8]. Lemma 1) shows that these requirements arc mutually inconsistent, and hc argues that it is most reasonable to drop the assumption of ex- tensibility. He chooses the mathematically convenient process of cubic spline fitting as the basis of his character typography. This is an in ing choice in the light of what follows here. We conclude from the above that either Ullman’s solution is rect, despite the erroneous justification he gives, in which case should like to be able to offer an alternative derivation, or else incorrect, in which case we should like to propose an alternative. For a longer discussion of the various suggestions regarding completion which have appeared in the litcraturc, the reader is rc to [ll. Often one finds that they are singularly lacking in justifi beyond the vague claim that they seem to generate “rcasonablc” sham On a more precise note, Pavlidis([ 111, chapters 7 and 8) surveys the many versions of polygonal approximation, spline fitting, and Fourier domain descriptors which have been proposed in the considcra tern recognition literature Two criteria have tended to domit selection of curves for shape completion, namely mathcmatic tability and computational efficiency, the latter implicitly assuming some level of computational power on conventional computers. The enor- mous computational power of the human perceptual system advocates caution in the application of such arguments. Indeed, a somewhat different approach is to base the sclcction of a class of curves on an analysis of some problem that the human visual system is trying to solve, and to isolate and examine the constraints within which that problem is posed. Several examples of this approach have app over the past few years, mainly in the work of Marr and his cd- laborators (see [9]). The work of Grimson, described briefly in Section 2. is of this sort. Ullman ([I61 section 3.3) makes some prclimi comments about the biological feasibility of pcrccptual computations, One of the difficulties which surrounds the choice of a class of curves in the specific case of subjective contours is that the differences between altcrnativc choices is often quite small at the angular extent of most examples of subjective contours. This in turn means that the issue is very difficult to resolve by psychophysical experimentation of the sort used by Ullman ([lS] page 3). In [I], we pursue an alternative approach, which was inspired by the second section of Ullman’s paper [ 151. He develops a local algorithm to compute the two circle solution which minimizes integral absolute 15 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. curvature. This naturally suggests dispensing with the cxtcnsibility as- sumption cntircly, just as Knuth did, and proceeding to find a solu- tion which minimizes some “pcrformancc index” related to curvature n, such as abs(n) or $. In order to test this idea, we apply the ideas of modern control theory (see [14],[2]). close approximation to curvature when the gradient fi is small. This is Since IC and abs(K) are non-conservative, WC consider minimizing $. We develop the Hamiltonian and show that it leads to a particularly intractable diffcrcntial equation namely, a reasonable condition to impose in the case of point which we argue at greater length in PI. subjective contours, a WC proceed to set up the Hamiltonian as usual. The boundary conditions arc as follows: (1) z = 0, y = 0, w = 4 (2) z = t, y = 0, w = -0 = yt. where w = 2, ?~n = 9, and the plant equations are: Yl = w WI =u (= &) (1) while the given by pcrformancc index is $. ‘I’hc Clamiltonian state function is where h is the Lagrange multiplier, and u is a constant of integration. This almost certainly does not have a closed form analytical solution. One possible iinc of approach at this juncture would be to base a local U2 Setting $$& equal to zero in the usual way, and solving 8$ = XI and g = -pl, leads to computation on one of the known techniques for the numerical solution of ordinary differential equations. Although shooting methods [3] are an obvious method on which yet explored the idea. to base such a local computation, WC have not ~=AxZ-p, In order to proceed, WC suggest in Section 2 that the shape com- pletion problem considcrcd hcrc is a two-dimcnsionnl analoguc of the problem of interpolating a three-dimensional surface, for example from the rclativcly sparse set of disparity points gcncratcd by the Marr- Poggio [IO] stereo algorithm. ‘I’his problem has rcccntly been studied where p is a constant of integration. This integrates easily to yield the cubic solution 3 2 y+-$+ux+T, where u and r are further constants of integration. Inserting the bound- ary conditions enables us to solve for h, p, u, and r. We get finally by Grimson [6]. He proposes that the pcrformancc index should be a semi-norm and shows that many of the obvious pcrformancc indices rclatcd to curvature, such as $ and abs(lc), do not have this property. He notes that the quadratic variation, defined by f2,* in two dimensions, is not only a semi-norm but is a close approx imation to curvature when the gradicn t fi is small (subscripts indicate partial dcrivativcs). ‘This is a rcasonablc condition to i::lpose in the case of subjcctivc contours. Accordingly, WC set LIP the 1 lomiltonian for the quadr‘rtic variation and show that it lcads to a cubic, which rcduccs to a parabola in the case of equal angles. This is particularly intcrcsting in view of the comments x3 x2 y = t;l(tanfI - tanf$) + t(tan$ - 2 tan@ + ztan0. In Figure 2, WC showed Ullman’s solution for the case 0 = 30”, 4 = 20’. In Figure 3 WC show the curve generated by our method, and in Figure 4 WC overlay the two. Clearly, the diffcrencc between the solutions is quite small at the angular extent shown in the figures. (Further examples arc given in [I]). As WC pointed out in the Introduction, this limits the usefulness of the kind of psychophysical experimentation used by Ullman to decide which solution is adopted by humans. made about Knuth’s work above. 2. Minimizing quadratic variation In order to proceed, we suggest that the shape completion problem In cast the angles 4 and 0 are equal, solution reduces to a parabola, namely the cubic term is zero and the considered here is a two-dimensional analoguc of the problem of inter- polating a three-dimensional surface, for example from the relatively 2 y= -Z-tanB+ztanf?. t sparse set of disparity points generated by the Marr-Poggio [lo] stereo algorithm. More generally, WC suggest that the process by which subjec- tive contours are gcncratcd results from the “mis-application” to two- The apex is at (1, 6 tan 8) and the focal length is &. Hence the focus is below the lint AB so long as 0 < 4, which is normally the case for subjective contours. dimensional figures of a process whose main purpose is the interpola- tion of three-dimensional surfaces. This idea requires some justification, which space here does not permit (see [l]). It also gives a different perspective on subjective contours, and leads to some fascinating, tes- table hypotheses. In the cast 0 = 4, it is straightforward to compute the difference between the parabola gcncrated by our method and the circle predicted by Ullman, which is The three-dimensional interpolation problem has recently been studied by Crimson [6] in the context of human stcreopsis. He observes that a prcrcquisite to finding the optimal fimction satisfying a given set of boundary conditions (namely that they should all pass through the given set of sparse points), is that the functions be comparable. Translating this into mathematics, hc proposes that the performance in- dex should be a semi-norm. Most importantly, hc shows that many of the obvious performance indices related to curvature, such as $ and Using the approximation (1 - abs(K), do not have this property. He notes that the quadratic variation, defined by f2,, in two dimensions, is not only a semi-norm but is a 16 this reduces to [4] Coren “Subjective contours and apparent depth,” Psychol. Review 79 (1972) 359-367. v - 4 II= -ztanB+ wz2tan30. t The diffcrencc bctwcen the solutions is essentially given by the second term whose maximum is at z = 4. Dividing by the horizontal extent t of the curve, the relative difference is bounded by w. For 8 < 1 this is negligible. 3. Acknowledgements This report describes research done in part at the Artificial Intelli- gence Laboratory of the Massachusetts lnstitute of Technology. Support for the laboratory’s artificial intelligence research is provided in part by the Advanced Research Projects Agency of the Dcpartmcnt of Defense under Office of Naval Research contract N00014-75-C-0643. The authors would like to thank the following people for valuable discussions at various stages of the research described here: John Greenman, Pat Hayes, David Marr, Slava Prazdny, Chris Rowbury. Robin Stanton, and Shimon Ullman. * This paper will appear as MIT AI memo 582. Here intelligibility and content have been sacrificed in the intcrcsts of strict page limit brevity. 4. References [I] Rrady, Crimson, and Langridge “Shape encoding and subjective con- tours,” (1980) to apEear. 1[2] Hryson and Ho Applied oplimal confrol, Ginn, Waltham MA, 1969. [3) Come and de Boor Elementary numerical analysis , McGraw Hill, Tokyo, 1972. [5) I;risby and Clatworthy “Illusory contours: curious cases of simul- tancous brightness contrast., 3 ” Perceprion 4 (1975), 349-357. [6] Grimson Computing shape using a theory of human stereo vision, Ph.D. ‘I’hcsis, Department of Dept. of Mathematics, MIT, 1980. [7] Kanisza “Subjective contours,” Sci. Artrer. 234 (1976), 48-52. ]$I Knuth “Mathematical typography,” Bull. Amer. M&h. Sot. (new series) 1(1979), 337-372. 19) Marr Vision, Freeman, San Francisco, 1980. [LO) Marr and Poggio “A theory of human stereo vision,” Proc. R. Sot. Lond B 204 (1979) 301-328. [Ill Pavlidis Slrucrural Partern Recognition , Springer Verlag, Berlin, 1977. ]I4 Rowbury Apparent contours, Univ. of Essex, UK,, 1978. ]13] Wutkowski “Shape completion,” Computer Graphics and Image Processing 9 (1979), 89-101. 1141 Schultz and Melsa State functions and linear control systems , McGraw Hill, New York, 1967. [IS] Ullman “Filling the gaps: The shape of subjective contours and a model for their generation,” Biof. Cyb. 25 (1976) l-6. 1161 Ullman “Relaxation and constrained optimisation by local processes,” Compuler Graphics and Image Processing IO (1979). 115-125. . . . . . . . . . Figure 2. Ullman’s solution for the case of boundary angles 0 = 30”, 4 = 20”. Figure 3. The solution generated by the method proposed in this paper. The boundary conditions are the same as those of Figure 2. ........ ....... ........ .... ..... .... ..... ... ... ...... . ... . ......... . . . . ........... . , ............. .............. . . ........... . . . . . Figure 4. The solutions of Figure 2 and Figure 3 are overlayed to demonstrate the difference between them. 17

1980

83

82

A STATISTICAL TECHNIQUE FOR RECOVERING SURFACE ORIENTATION FROM TEXTURE IN NATURAL IMAGERY Andrew P. Witkin Artificial Intelligence Center SRI International, Menlo Park, CA 94025 ABSTRACT A statistical method is reported for inferring the shape and orientation of irregularly marked surfaces using image geometry. The basis for solving this problem lies in an understanding of projective geometry, coupled with simple statistical models of the contour generating process. This approach is first applied to the special case of surfaces known to be planar. The distortion of contour shape imposed by projection is treated as a signal to be estimated, and variations of non-projective origin are treated as noise. The resulting method is next extended to the estimation of curved surfaces, and applied successfully to natural images. The statistical estimation stratefl is then experimentally compared to human perception of orientation: human observers' judgements of tilt correspond closely to the estimates produced by the planar strategy. I INTRODUCTION Projective geometry lawfully relates the shape of a surface marking, the orientation of the surface on which the marking lies, and the shape of the marking's projection in the image: given any two, the third can usually be recovered. This paper addresses the problem of recovering the shape and orientation of an irregularly marked surface from the shapes of the projected markings. Of the three constituents of the projective relation, one--projected shape--is given, and another-- surface orientation--is to be recovered. Since two of the constituents must be known to recover the third, the problem has no solution unless something is assumed about the unprojected shapes of the surface markings. For instance, if those shapes were known exactly, recovering surface orientation would usually be a straightforward exercise in projective geometry. More interesting and more general is the case of a surface about which nothing specific is known in advance. To recover surface orientation in this case, some assumption must be made about the geometry of surface markings that is general enough to apply to a broad range of surfaces, powerful enough to to determine a solution for surface orientation, and true enough to determine the right solution. To meet these requirements, the inference of surface shape will be treated as a problem of statistical estimation, combining constraints from projective geometry with simple statistical models of the processes by which surface markings are formed. The distortion imposed on shapes by projection will be treated, quite literally, as a signal, and the shapes themselves, as noise. Both the "signal" and the "noise" contribute to the geometry of the image, and statistical models of the noise permit the projective component to be isolated. II ESTIMATING THE ORIENTATION OF PLANAR SURFACES The estimation problem will first be considered subject to the artificial restriction that the surface is known to be planar. While not realistic, this limited case provides the groundwork from which more general methods will be developed. The shape of a surface marking is the shape of its bounding curve. The shapes of curves are naturally described by tangent direction as a function of position. Since tangent direction in the image is subject to projective distortion, this representation is suitable for estimating projective distortion. The mapping between a tangent direction on a surface and the corresponding direction in the image is readily expressed as a function of surface orientation, by means of simple projective geometry. By examining a region of the image, a distribution of projected tangent directions may be collected. The form of this distribution depends in part on the distribution of directions over the corresponding surface region, but undergoes systematic distortion depending on the orientation of the surface: the projection of an inclined shape is foreshortened, i.e. compressed in the direction of steepest inclination (the tilt direction.) The amount of compression varies with the angle between the image plane and the plane of the shape (the slant angle.) The effect of this distortion on the distribution of tangent directions may be illustrated by the simple case of a circular marking. The distribution of tangent directions on the original circle, measured by arc length, is 1 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. uniform. The orthographic projection of a circle is an ellipse, whose minor axis lies parallel to the tilt direction, and whose eccentricity varies with the slant angle. The distribution of tangent directions on an ellipse is not uniform, but assumes minima and maxima in the directions of the minor and major axes respectively. The degree of nonuniformity increases with the eccentricity of the ellipse. In other words, projection systematically "pushes" the image tangents away from the direction of tilt. The greater the slant angle, the more the tangents are "pushed." This systematic distortion of the tangent distribution is the projective "signal" that encodes surface orientation. Its "phase" (the direction toward which the tangents gravitate) varies with surface tilt, and its "amplitude" (th e amount of distortion,) with surface slant. To estimate the phase and amplitude of the projective signal is to estimate surface orientation. Statistical estimation of the projective component requires a model of the "noise", i.e. of the expected distribution of tangent directions prior to projective distortion. With no prior knowledge of the surface, there is no reason to expect any one tangent direction to be more likely than any other. In other words, it is natural to assume that all tangent directions on the surface are equally likely. (Note that the amount of projective distortTon increases with surface slant, effectively increasing the signal-to-noise ratio. Therefore, as slant increases, the exact form assumed for the noise distribution becomes less critical.) Together with the geometric relation, this simple statistical model defines a probability density function for surface orientation, given a set of tangent directions measured in the image. The surface orientation value at which this function assumes a maximum is the maximum likelihood estimate for surface orientation, given the model. The integral of the function over a range of surface orientations is the probability that the actual orientation lies in that range. Conceptually, the distribution of tangent directions in the image can be projected onto an arbitrarily oriented planar surface. We seek that orientation for which the distribution of projected tangent directions is most nearly isotropic. The estimator was first applied to geographic contours: projections of coastlines drawn from a digitized world map. This choice of data circumvents the problem of contour detection, and allows the actual orientation to be precisely controlled. The overall accord between estimated and actual orientation was excellent, and, equally important, the confidence measures generated by the estimator effectively distinguished the accurate estimates from the inaccurate ones. I- -1 --- Figure 1 - Two photographs of roughly planar surfaces, and the orientation estimates obtained from them. The estimated orientations are indicated by ellipses, representing the projected appearance a circle lying on the surface would have if the maximum likelihood estimate were correct. III EXTENTION TO CURVED SURFACES To apply the methods developed in the planar case to curved surfaces without additional assumptions, it would be necessary to obtain at each point in the image a measure of the distribution of tangent directions. But such a local measure is never available, because the density of the contour data is limited. On the other hand, a distribution can be taken at each point of the data in a surrounding region, as small as possible, but large enough to provide a reasonable sample. This spatially extended distribution may be represented as a three dimensional convolution of the image data with a summation function. The same technique was then applied to natural To understand how such a distribution should images, using zero-crossing contours in the be applied to estimate surface disposition, it is convolution of the image with the Laplacian of a Gaussian [I], [2]. helpful to distinguish the intuitive, perceptual While the veridical notion of surface orientation from the strict orientations were not independently measured, the definition of differential geometry. It may be maximum likelihood estimates are in close accord argued that surface orientation is not a unique with the perceived orientations (see Figure 1). 2 property of the surface, but must be regarded as a function of scale. The scale at which orientation is described corresponds to the spatial extent over which it is measured. Thus, by measuring orientation over a'large extent, the surface is described at a coarse scale. The scale at which the,surface is estimated therefore depends on the spatial extent over which the distribution is computed. Since that extent must be sufficiently large compared to the density of the data, the density effectively limits the spatial resolution of the estimate. On this view, close parallels can be drawn to Horn's [3] method for inferring shape from shading. The local estimator for orientation is a geometric analogue to the photometric reflectivity function. The strategy was implemented, and applied to natural images. Contours were extracted as in the planar case, and the spatially extended distribution approximated by a series of two- dimensional convolutions with a "pillbox" mask. The estimated surfaces were in close accord with those perceived by the human observer (see Figure 2). IV RELATION TO HUMAN PERCEPTION A psychophysical experiment was performed to examine the relation between human observers' judgments of the orientations of curves perceived as planar, and the estimates obtained from the estimation strategy outlined above. A series of "random" curves were generated using a function with pseudorandom parameters. Although such curves have no "real" orientation outside the picture plane, they often appear inclined in space. Observer's judgments of orientation were obtained by matching to a simple probe shape. The judgments of tilt (direction of steepest descent from the viewer) were highly consistent across observers, while the slant judgments (rate of descent) were much more variable. Orientation estimates for the same shapes were computed using the planar estimator, and these estimates proved to be in close accord with those of the human observers, although the shapes had no " real" orientation. While no conclusion may be drawn about the mechanism by which human observers judge orientation from contours, or about the measures they take on the image, this result provides evidence that the human strategy, and the one developed on geometric and statistical grounds, are at the least close computational relatives. 1. 2. 3* REFERENCES Marr, D. C & Poggio, T., "A computational theory of human stereo vision," Proc. Roy. Sot. Lond., Vol. 204, pp.301-328 (1979) Marr, D. C & Hildreth, E., "A Theory of edge detection," MIT AI Memo 518 (1979) Horn, B. K. P, "Understanding image intensities," Artificial Intelligence, Vol. 21, No. 11, pp.201-231 (1977) Figure 2 - A complex image, and the estimate obtained. The ellipses represent the appearance the estimated surface would have were it covered with circles. Note that the estimate correctly distinguishes the highly slanted foreground from the nearly frontal background. The upward pitch of the right foreground is also detected. 3

1980

84

83

INFORMATION NEEDED TO LABEL A SCENE Eugene C. Freuder Dept. of Mathematics and Computer Science University of New Hampshire Durham, NH 03824 ABSTRACT I analyze the information content of scene labels and provide a measure for the complexity of line drawings. The Huffman-Clowes label set is found to contain surprisingly little additional information as compared to more basic label sets. The complexity of a line drawing is measured in terms of the amount of local labeling required to determine global labeling. A bound is obtained on the number of lines which must be labeled before a full labeling of a line drawing is uniquely deter- mined. I present an algorithm which combines local sensory probing with knowledge of labeling constraints to proceed directly to a labeling analysis of a given scene. "r INTRODUCTION Huffman [4] and Clowes [2] developed (inde- pendently) a basic labeling scheme for blocks world picture graphs. Given a basic labeling set: + (convex), - (concave), -+ (occluding, region on arrowhead side), and a standard set of simplifying restrictions on scene content and viewpoint, the physically realizable junction labelings are just those shown in the last column of Fig. 1. Waltz [5] explored the richer label sets obtained by including additional information in the labels (and loosening the scene restrictions). In this paper I explore weaker, cruder label sets. I identify three stages of scene labels, of which the standard set is the third and richest. I then explore the increase in information content embodied in each successive stage. Rather sur- prisingly I find that there is very little real information gain as we move from stage to stage. A first stage scene labeling may well determine a unique second stage labeling. If it does not, it will come quite close to doing so, and I am able to identify precisely the additional information that is necessary and sufficient to complete the second stage labeling. Similar results are obtained for the transition from Stage II to Stage III. These results supply some theoretical insight into the nature and strength of the basic line labels and physical constraints. I go on in Section III to analyze the amount of information required to obtain a Stage I labeling. The information is measured in terms of the number of line labels which must be determined in order for labeling constraints to unambiguously STAGE I STAGE II STAGE III Fork A 2 - - 2 2 A 2 2 2 A + + + Arrow I I T T 2 - +I+ 2 - T +i Fig. 1. Junction labelings ot each stage imply a unique labeling of the entire line draw- ing. In practice the required line labels can be obtained by local sensory probing of the physical scene. I obtain a bound on the number of labels required to imply a full labeling of an arbitrary line drawing. Finally I discuss an algorithm that effectively combines sensory probing for labels with knowledge of labeling constraints. The algorithm proceeds directly to a full labeling, reflecting a presented physical scene, while requiring neither a complete sensory scan for every line label nor a consideration of all possible physical realizations of the line drawing. 18 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. II LABELING STAGES The standard label set is a refinement of a cruder categorization of lines as representing a physical edge of either one or two visible faces. I consider three labeling stages. In Stage I the only labels are the numbers 1 and 2, indicating the number of associated faces. In Stage II, the number 1 is replaced by occlusion labels (+) indicating which side the single face is on. They will be termed Stage II labels. A Stage II label- ing will be one that utilizes -f and 2 labels. In Stage III the number 2 is replaced by + and - as the distinction is made between convex and concave edges. The labels + and - will be termed Stage III labels. - A Stage III labeling is one that utilizes +, - and -f labels. At Stage I there are only 9 distinct junction labels. At Stage II the L labelings are differ- entiated, at Stage III the fork and arrow labels are differentiated. Fig. 1 shows the physically realizable labelings at each stage. T labelings are added at each stage, but notice that fork and arrow labelings do not increase in moving from Stage I to Stage II, and the number of L labelings does not increase in moving from Stage II to Stage III. Thus we really do not know any more about forks and arrows at Stage II than we do at Stage I, nor more about L's at Stage III than at Stage II. Once we have labelled a fork 2,1,1 for example, we really know already that it can be labelled 2,+,-t. My interest in Stage ,J labeling was aroused by the work of Chakravarty [1] who utilized inform- ation about the number of regions associated with lines and junctions, in connection with a more elaborate labeling scheme. A. The Picture Graph -- A blocks world line drawing is, of course, a graph. For the purposes of our analysis we will modify picture graphs by "separating" T junctions, removing T junctions by pulling the shafts away from the crossbars. After labeling a scene separated in this fashion the T junction labelings are easily recovered by rejoining the T junctions to form the original scene. The separation reflects the fact the information does not pass through a T junction , and will permit us to identify independent segments of the scene as connected components of the (separated) picture graph. The segments are independent in the sense that each can be labeled independently, a label in one segment can have no bearing on possible labelings for the other segment. The connected components of a graph are the maximal connected subgraphs, where a graph is connected if there is a chain of edges between any two vertices. B. Stage I to Stage II Theorem 1. Given a picture graph with a Stage I labeling (separated at T junctions). Further separate the graph by separating L junc- tions that have a 2 label on one line, i.e. pulling the two sides of each such L apart to remove the junction. The Stage I labeling uniquely deter- mines a Stage II labeling on all connected compon- ents of the resulting graph except those consist- ing solely of l-labeled lines, none of which is a crossbar of a T. A unique labelinq for the ex- ceptions may be determined by specifying the Stage II label of a single line in each such component. c31 l For proofs of the theorems in this paper see C. Stage II to Stage III -- Theorem 2. Given a picture graph with a Stage II labeling (separated at T junctions). The Stage II labeling uniquely determines a Stage III labeling on all connected components except those consisting solely of 2-labeled lines. A unique labeling for the exceptions may be determined by specifying the Stage III label of a single line in each component, III OBTAINING A STAGE I LABELING Given labels for a sufficient number of lines the physical constraints on the labeling process will imply a unique labeling for the remainder of the scene. A bound on this "sufficient number" will provide a bound on the complexity and poten- tial ambiguity of the picture graph, and on the effort required to label it. The limitations on physically realizable labelings summarized in Fig. 1 easily give rise to a set of "implication rules" for completing a junction labeling given the labelings of one or two of its lines. 2. These rules are listed in Fig. Note first that labels for shafts of arrows and crossbars of T's can be derived immediately (2's and l's respectively), without any previous labels. Labels for two of the lines of a fork or arrow imply the third. (Thus, in effect, a single line label, other than for the shaft, determines an arrow labeling.) => 2 I\ I I => T p => yp yp => ‘;1‘ Fig. 3. Implication rules. 19 I will say that the labelina of a subset of picture graph iines implies the labeling of the entire graph, if repeated application of the implication rules of Fig. 2;'starting with the given subset, leads to a complete, unique labeling of t,he graph. The labeling of a subset of lines is sufficient if the labeling implies the labeling of the graph. A subset of lines is sufficient if any consistent lamoftheubset is suffi- cient. The minimal number of lines in a suffi- cient subset will be called the sufficient number of the picture graph. -- The sufficient number must, of course, be determined relative to a specified label set. We will be dealing with sufficiency for Stage I labeling in this paper. In Section A I obtain an upper bound on the sufficient number of a picture graph. In [3] I discuss means of obtaining sufficient sets of lines (and thus, bounds on sufficient numbers) for individual graphs; I modify one of these methods to provide a test for the sufficiency of a set of lines or line labels, and a labeling algorithm. This algorithm is discussed in Section B. Note that deduction can "propagate": we may deduce a label for a line which in turn permits deduction of the label for an adjoining line, etc. There is room for heuristic tuning in choosing which lines to probe for labels. Fig. 3 demonstrates thi on a simple scene. The fort labeled L line to probe for get away with a single senso the choice need not be entir reasonable to suspect that t be 2-labeled.) m + + I I+ I w.2 a. s labeling algorithm uitous choice of a 2- labeling permits us to ry probe. (Actually ely fortuitous. It is he "bottom lines" will A. A Bound on the Sufficient Number of a Picture ---- ---- Graph Theorem 3. The sufficient number of a picture graph is no more than the number of forks and arrows plus the number of connected components of @ => l@l O=>probed the graph separated at T's and L's. The bound provided by the theorem is tight, in the sense that I can exhibit a picture graph with sufficient number equal to the given bound. A simple triangular figure may consist of all occluding lines, or one of the lines may have a 2 label. Knowing the labeling of two lines is not sufficient to imply the label of the third for all possible labelings. Thus, the sufficient number Fig. 3. Labeling algorithm. REFERENCES [1] Chakravarty, I. A generalized line and junction labeling scheme with applications to scene analysis, IEEE Trans. PAMI- (1979) 202-205. is three, which equals the number of forks and arrows (0) plus the number of connected components in the graph separated at T's and L's (3). (In general the sufficient number of a picture graph will be considerably less than the bound provided by the theorem). PI c31 B. A Labeling Algorithm II41 Algorithm: 1. Obtain any labels that can be deduced by repeated application of the implication rules of Figure 2. (N t o e arrow shafts and T crossbars can be labeled immediately.) II51 2. While unlabeled lines remain: 2.1. Pick an unlabeled line and "probe" the physical scene to determine its label. (This information could be obtained from visual, tactile, or range finding data.) Clowes, M.B. On seeing things, Artificial Intelligence 2 (1971) 79-116. Freuder, E. On the Knowledge Required to Label a Picture Graph. Artificial Intelli- Qence, in press. Huffman, D.A. Impossible objects as nonsense sentences, in: Meltzer, B. and Michie, D. (Eds.), Machine Intelligence 5 (Edinburgh Univ. Prminburgh, 1971) 295-324. Waltz, D. Understand ing line drawings of scenes wit h shadows, in: Winston, P.H. (Ed.), The Psychology (McGraw-Hill, New York, 1975 19-91. of Computer Vision 2.2. Deduce any further labels that can be obtained by repeated applications of the implication rules. 20

1980

85

84

INTERPRETIVE VISION AND RESTRICTION GRAPHS Rodney A. Brooks and Thomas 0. Binford Artificial Intelligence Laboratory, Computer Science Department Stanford University, Stanford, California 94305 ABSTRACT We describe an approach to image interpretation which uses a dynamically determined interaction of prediction and observation. We provide a representational mechansim, built on our geometric modeling scheme which facilitates the computational processes necessary for image interpretation. The mechanism implements generic object classes and specializations of models, enables case analysis in reasoning about incompletely specified situations, and manages multiple hypothesized instantiations of modeled objects in a single image. It is based on restriction nodes and quantified variables. A natural partial order on restriction nodes can be defined by comparing the satisifiability of their constraints. Nodes are arranged in an incomplete restriction graph whose arcs represent relations of nodes under the partial order. Predictions are matched to descriptions by finding maximal isomorphic subgraphs of a prediction graph and an observation graph 183 subject to a naturally associated infimum of restriction nodes being satisifiable. In this manner constraints implied by local two dimensional matches of image features to predicted features are propagated back to the three dimensional model enforcing global consistency. I INTRODUCTION A. Image Interpretation A descriptive process is one which takes an image and produces a description of image features and their relations found within that image. A predictive process is one which uses models of objects expected in an image to predict what features and their relations will be present in the image. We have constructed a working model-based vision system called ACRONYM 183. The interaction between prediction and description is stronger than in previous systems. A central aspect of ACRONYM is that it interprets images at the level of three dimensional models. Here we describe a new layer of representation built on the ACRONYM system. We briefly describe a mechanism for reasoning about incompletely specified geometric situations. All this has been implemented. We describe how matching of two dimensional features can be mapped back to constrain geometric uncertainties in three dimensions in order to obtain a *three dimensional understanding of an image. This aspect of the new system is still being implemented (June 1980). 21 B, The ACRONYM System ACRONYM itself covers a wider range of tasks than vision (see 143, [lo]). The reader is referred to 181 for a complete overview of the system and a description of geometric models based on generalized cones as the nodes of a subpart tree. We give here a brief overview of ACRONYM’S vision related modules. A user interacts with the system via a high level modeling language and an interactive editor to provide three dimensional descriptions of objects and object classes which are viewpoint independent, and to partially mode1 general scenes. The result is the object graph. A library provides useful prototypes and a graphics module provides valuable feedback to the user. A rule-based module, the predictor and planner, takes models of objects and scenes and produces the prediction graph which is a prediction of the appearance of objects expected within the scene. It predicts observables in the image over the expected range of variations. It provides a plan, or instructions, for lower level descriptive processes and the matcher to find instances of the objects within the image. The process of prediction and planning is repeated as first coarse interpretations are found, more predictions are carried out, and finer, less ambiguous interpretations are produced. The descriptive aspect of ACRONYM is currently provided by the edge mapper IS] which describes monocular pictures as primitive shape elements (ribbons) and their spatial relationships. It is goal-directed and is thus programmed by the predictor and planner. The observation graph is the result. As with the predictor and planner, the edge mapper may be invoked many times during the course of an interpretation as finer levels of description become desirable. ACRONYM will incorporate stereo and advanced edge-based description modules (Baker [23 and Arnold and Binford [II). This will provide three dimensional cues directly. The matcher interfaces description and prediction. It finds maxima1 subgraphs of the observation graph isomorphic with subgraphs of the prediction graph, which also meet global consistency requirements. In the new implementation the matching process is mapped back to three dimensional models. Such higher level understanding ensures global consistency and enables deductions about three dimensional structures from a single monocular image. The matcher re-invokes the predictor and planner and the edge mapper to extend the two graphs which it is matching in the context of successfully completed submatches. This provides direction to both prediction and From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. description and reduces each must consider. dramatically the number of possibilities Lowe 191 has implemented a system which determines parameters of models including articulations from correspondences of a match. This module provides predictions from a tentative interpretation to guide detailed verification. We have concentrated on two classes of images in our development work; aerial images of airport scenes, and scenes of industrial workstations. Together they provide a wide range of challenges and force us to look for general methods and solutions to problems, since they are sufficiently dis-similar that special purpose solutions will fail on one of the two. II REPRESENTATION A. Requirements We have chosen to describe the world to ACRONYM in terms of three dimensional models of objects, and their expected three dimensional spatial relationships ([Sl contains details). The representation is a coarse to fine description of objects as subpart trees of generalized cones 133. In this paper we are concerned with ways to represent variations within models, and how to keep track of multiple inconsistent instances which may arise during image interpretation. Thus what follows does not rely on generalized cones as the representational primitive. In structured situations the exact dimensions of objects may be known in advance, but their orientation may be uncertain. For instance it may be known that a bin is full of a particular type of part, but the parts may have been dropped in with arbitrary orientations. In an industrial automation domain such descriptions may already be available from a Computer Aided Design data-base. In less structured environments, not even the dimensions of objects will be known exactly. For instance for some aerial image interpretation tasks it is desirable to represent both the class of wide bodied passenger jet aircraft, and particular types of aircraft such as a Boeing 747, and even more particular models such as a 747-B. Thus it is necessary to represent constrained variations in shape, size and even structure (e.g. different aircraft have different engine configurations), and constrained variations in spatial relations between objects. Consider also that an F-l 11 aircraft can have variable wing geometry. A manipulator arm has even more complex variations in spatial relations between its subparts. The appearance of an object may change qualitatively, rather than merely quantitatively, over the allowed variations in its size, shape, structure or orientation relative to the camera. Thus it will often be necessary to carry out some case analysis in prediction of appearance, and put further constraints on models, and to keep such mutually confliciting hypotheses in the prediction graph, until such time as they can be confirmed or denied by descriptive processes. The prediction graph represents case as combinations of components instead of explicit enmueration of all cases. As interpretation of an image proceeds, constraints on the exact values of variations within a model will be derived from the matches made in the image. However there may be multiple instances of a modeled object within the image. Parts on a conveyor belt will have different orientations. Aircraft at a passenger terminal will have different lengths and wing spans. Thus multiple instances of objects must be representable in the interpretation graph. B. Representing Variations In the following discussion we will consider the problem of modeling both the generic class of wide-bodied passenger jet aircraft, and specific wide-bodied passenger jet aircraft, such as the Boeing 747, Lockheed L-101 1, McDonnell-Douglas DC-10 and the Airbus Consortium A-300. We will then discuss a wider situation where such aircraft are on runways and taxiways, and there are undetermined variables in the camera model. We need to represent both variations in size (e.g. different aircraft subclasses will have different fuselage lengths), and variations in structure (e.g. different aircraft subclasses will have different engine configurations). In both cases we want to represent the range of allowable variations. We consider the broader problem of quantification of sets. Furthermore, there will sometimes be interdependencies between these variations (e.g. a scaling between fuselage length and wing span). Node: FUSELAGE-CONE NAME : SIMPLE-CONE SPINE: 20005 SWEEPING-RULE: CONSTANT-SWEEPING-RULE CROSS-SECTION: 20004 Node: Z0005 NAME: TYPE: LENGTH: SPINE STRAIGHT FUSELAGE-LENGTH Node: CONSTANT-SWEEPING-RULE NAME : SWEEPING-RULE TYPE: CONSTANT Node: 20004 NAME : TYPE: RADIUS: CROSS-SECTION CIRCLE FUSELAGE-RADIUS Generalized cone representation Figure 1. of fuselage. The primitive representational mechanism used in ACRONYM is that of units and slots. Objects are represented by units, as are generalized cones, cross-sections, sweeping-rules, spines, rotations and translations to name the more important ones. Figure 1 shows four units with their slots and fillers from a particular ACRONYM model. They describe the generalized cone representing the fuselage of the generic wide-bodied passenger jet aircraft. Note that units are referred to as “Nodes” because they are nodes of the Object graph of figure I. The NAME slot is a distinguished slot which all units possess. It describes the entity represented by the unit and corresponds to the SELF slot of KRL units E51. Units identified by “Z” followed by a four digit number are those which were given no explicit identifier by the user who modeled the object. The modeling language parser has generated unique identifiers for 22 The value of a slot is given by its filler. Slot fillers may be explicit, such as “2” or “STRAIGHT”. They can also be symbolic constants in the same sense as constants are used in programming languages such as PASCAL. Such fillers are fine for representing specific completely determined objects and situations. Slots may be filled by a quantifier, or any evaluable expression involving quantifiers. A quantifier is an identifier with constraint “FUSELAGE-LENGTH” and system (quantification). “FUSELAGE-RADIUS” are examples of such quantifiers in figure 2. The following constraints might be imposed upon FUSELAGE-LENGTH and FUSELAGE-RADIUS when modeling the class of wide-bodied passenger jet aircraft: Node: JET-AIRCRAFT NAME 8 OBJECT SUBPARTS: (STARBOARD-W I NG PORT-W I NG FUSELAGE 1 QUANTIFIERS: (F-ENG-QUANT ENGINE-LENGTH ENGINE-RADIUS WING-ATTACHMENT ENG-OUT ONE-WING-SPAN WING-SWEEP-BACK WING-LENGTH WING-RATIO WING-WIDTH WING-THICK) Node : STARBOARD-W I NG (s 40.0 FUSELAGE-LENGTH) (s FUSELAGE-LENGTH 70.0) fs 2.5 FUSELAGE-RADIUS) (s FUSELAGE-RADIUS 3.5) (s 15.6 (QUOTIENT FUSELAGE-LENGTH FUSELAGE-RADIUS)) NAME : OBJECT SUBPARTS: ( (SP-DES F-ENG-QUANT e STARBOARD-ENGINE)) CONE-DESCRIPTOR: STARBOARD-WING-CONE These constrain the range of allowable length and radius, and express a lower bound on the ratio of length to radius. Quantifiers express allowable variations in dimensions of objects and in the structure of objects. Figure 2 gives the complete subpart tree for a model of generic wide-bodied passenger jet aircraft. For brevity, not all the slots of the OBJECT units are shown here. The QUANTIFIERS slot is explained later. The SUBPARTS slot of an OBJECT unit is filled with a list of subparts giving the next level of description of the object. Entries in the list can be simple pointers to other OBJECT units (e.g. JET-AIRCRAFT has three substructures: STARBOARD-WING, PORT-WING and FUSELAGE). They can also be more complex such as the single entry for the subparts of STARBOARD-WING, which speiifies a quantification of subparts called STARBOARD-ENGINE. In this case the quantification is the quantifier F-ENG-QUANT. Note that PORT-WING has a quantification of PORT-ENGINES as subparts, which is represented by the same quantifier F-ENG-QUANT. This explicitly represents one aspect of the symmetry of the aircraft: it has the same number of engines attached to each wing. Constraints on this quantifier and on R-ENG-QUANT, the number of rear engines might be: Node: STARBOARD-ENGINE NAME I OBJECT CONE-DESCRIPTOR8 PORT-ENGINE-CONE Node : PORT-W I NG NAME r OBJECT SUBPARTS: ( (SP-DES F-ENG-QUANT . PORT-ENGINE) 1 CONE-DESCRIPTOR: PORT-WING-CONE Node: PORT-ENGINE NAME : OBJECT CONE-DESCRIPTOR: PORT-ENGINE-CONE Node: FUSELAGE NAME : OBJECT SUBPARTS: (RUDDER STARBOARD-STABILIZER PORT-STABILIZER) QUANTIFIERS: (STAB-ATTACH STAB-WIOTH STAB-THICK STAB-SPAN STAB-SWEEP-BACK STAB-RAT 101 CONE-DESCRIPTOR: FUSELAGE-CONE (5 1 F-ENG-QUANT) ts 2 F-ENG-QUANT) (s 0 R-ENG-QUANT) (I 1 R-ENG-QUANTI (b 3 (PLUS F-ENG-QUANT R-ENG-QUANTI) These say that there must be either one or two engines on each wing, -zero or one at the rear of the aircraft, and if there are two on each wing then there are zero at the rear. Symmetry of size (such as length of the wings) can likewise be represented by using the same quantifier as a place holder in the appropriate pair of slots. Node : RUDDER NAME : SUBPARTS : OBJECT ( (SP-DES R-ENG-QUANT . REAR-ENGINE)) CONE-DESCRIPTORI RUDDER-CONE Node: REAR-ENG I NE NAME : OBJECT CONE-DESCRIPTORr REAR-ENGINE-CONE Node: STARBOARD-STABILIZER NAME : OBJECT Our compjete model for a generic wide-bodied passenger jet aircraft has 28 quantifiers describing allowable variations in size and structure. CONE-DESCRIPTOR: STARBOARD-STABILIZER-CONE Node: PORT-STABILIZER NAME t OBJECT CONE-DESCRIPTORr PORT-STABILIZER-CONE Subpart tree of generic Figure 3. passenger jet. 23 C. Representing Classes It should be clear that to model a subclass of wide-bodied passenger jet aircraft we need only provide a different (more restrictive) set of constraints for the quantifiers used in ehe general model. To model a specific type of aircraft we could force the constraints to be completely specific (e.g. (= FUSELAGE-LENGTH 52.8)) ). Thus we will not need to distinguish between specialization of the general model to a subclass, or an individual. Given that subclasses use different sets of constraints, the problem arises of how to represent multiple subclasses simultaneously. We introduce a new type of node to the representation: a restriction node. These are the embodiment of specialization. A restriction node has a set of constraints associated with it. If a set of values can be found for all the quantifiers mentioned in the constraints such that all the constraints are simultaneously satisifed, then we say the restriction node is satisifiable. A partial order can be defined on restriction nodes by saying that one restriction node is more restrictive than another if its set of sets of satisfying values is a subsee of that of rhe second node. Different views of the generic model. Figure 4.- For the example of the generic wide-bodied passenger jet aircraft the constraints are associated with some restriction node, GENERIC- JET-AIRCRAFT say. To represent the class of 747s a more reskictive node can be included; e.g.: Node: BOEING-747 NAME: RESTRICTION SUPREMA: (GENERIC-JET-AIRCRAFT) TYPE: MODEL-SPECIALIZATION CONSTRAINTSI <list of constraints> It is constructed by taking the constraints associated with the GENERIC-JET-AIRCRAFT restriction node, and merging in additional constraints to specialize to a BOEING-747. A model is always accessed in Ihe context of a restriction node. Thus when reasoning about the generic class of wide-bodied aircraft, the predictor and planner will access the JET-AIRCRAFT model and base its reasoning on the constraints given by the GENERIC-JET-AIRCRAFT restriction node. When reasoning about Boeing 747s it will base its reasoning about the JET-AIRCRAFT model on the constraints given by the BOEING-747 restriction node. Figure 4 conveys the flavor of viewing the JET-AIRCRAFT through different restriction nodes to see different models. (And in fact, the drawings of the two types of aircraft were produced by ACRONYM from the indicated restriction nodes.) In modeling subclasses, restriction nodes typically form a tree rather than a graph. D. Representing Spatial Relations Affixments are coordinate transforms between local coordinate systems of objects. They are comprised of a rotation and a translation. Sometimes affixments vary over an object class. For instance the in generic wide-bodied passenger jet aircraft model the position along the fuselage at which the wings will be attached will vary with particular types of aircraft. Articulated objects are modeled by variable affixments. Variable affixments can also be useful for modeling spatial relationships between two objects - for instance an aircraft is on a runway. We represent a vector as a triple (a,b,c) where a, b and c are scalars. We represent a rotation as a pair <v,m> where v is a unit vector, and m a scalar magnitude. An affixment will be written as a pair (r,t) where r is a rotation and t a translation vector. We will use some special vectors also: 4 Q and f. We use * for the composieion of rotations, and QP for the application of a rotation to a vector. In ACRONYM we use the quantifier mechanism to represent affixments which describe a class of coordinate transforms. This gives symbolic representations for rotations and translations. Consider the problem of representing ehe fact that an aircraft is somewhere on a runway. Suppose the runway has its x axis along its length, the y axis perpendicular at one end, and the positive z direction vertically upward. Suppose that the coordinate system for the aircraft has its x axis running along the spine of the fuselage and has its t axis skyward for the standard orientation of an aircraft. Then to represent the aircraft on the runway we could affix it with the affixment: (<f, ORI>, (JET-RUNWAY-X, JET-RUNWAY-Y, 011 where ORI, JET-RUNWAY-X and JET-RUNWAY-Y are quantifiers with the following constraints: (I 0 JET-RUNWAY-X) (ls JET-RUNWAY-X RUNWAY-LENGTH) (5 0 JET-RUNWAY-Y) (< JET-RUNWAY-Y RUNWAY-WIDTH) Notice that OR1 is unconstrained. The aircraft is constrained to be on the runway, in the normal orientation for an aircraft (e.g. not upside down), but it does not constrain the 24 direction in which the aircraft is pointed. If we wished to constrain the aircraft to approximately line up in the direction of the runway we could include a constraint on the quantifier ORI, allowing for some small uncertainty. In general, constraints on a single quantifier may derive from different coordinate systems. III PREDICTION AND MATCHING The left most rotation corresponds to a rotation in the image plane and can be ignored when predicting image shape - i.e. shape is invariant with respect to rotation in the image plane. The right most rotation expression is applied directly to the cylindrical tool. But it is a rotation about the x axis, which is the linear axis of a cylinder in our representation 181 and the appearance of a cylinder is invariant with respect to a rotation about its linear axis. Thus for shape prediction we need only consider: <y^,TILT> A. Using Constraints Prediction consists of examining the model in order to find features which are invariant over the range of variations in the model Es], or over a sufficient range to allow a small number of cases. A mixture of reasoning directly about the model, and reasoning about the constraints is needed to find invariants. If ‘TILT is sufficiently constrained (as in this example) it may be possible to predict the shape directly. The prediction takes the form of expected image features, their relations, and what constraints local matches to such features produce on the three dimensional model. See section III-C for an example. But note here that the prediction is a conjunction of expected features. B. Adding Constraints Electric Screwdriver and Holder Figure 5. If TILT in the above example is not sufficentiy constrained there may be more than one qualitatively different shape possible for the cylinder (e.g. an end view of a cylinder is quite different from a side view). If so it is necessary to make a disjunction of predictions. Note however, that ail views need not be explicitly expanded - they can still share much structure. Consider the electric screwdriver holder and electric screwdriver in figure 5. This is a display of an ACRONYM model of a tool for the Stanford hand-eye table. The position and orientation (about the vertical axis) are not known. Neither are the exact camera pan and tilt known. Under these conditions the expression for the orientation of the screwdriver tool relative to the camera, as obtained directly from the model, is: cx^, TI LT>YK<Q, (- PAN) .*<Z?, 3n/Z>r~<y^, 3n/2>*I a<f,ORI >* I *<(j, 3n/2>r#<y^, n/2>*<& n/2>*1 *I *I *I Each prediction is associated with a new, more restrictive restriction node. It is obtained by adding a new constraint which restricts the model sufficiently to necessitate only a single prediction. Figure 5 gives an indication of the structure of a local prediction, with two different cases considered. Not indicated in that diagram are arcs between the feature predictions which specify relations which should hold between instances of those features in the image. R..LrlCLlm u&m corm PredlCLlon where I is the identity rotation, PAN and TILT are quantifiers associated with the camera orientation, and OR1 is the unconstrained orientation of the screwdriver holder about the vertical axis. We have implemented a set of rules in ACRONYM which simplify such expressions to a canonical form, using identities for re-ordering products of rotations. The details can be found in 173. The canonical form aids in the detection of invariants. E.g. the above expression is transformed to the equivalent expression: <f,3n/Z>r<~,TILT>*<~, (+ PAN (- ORi))> W/l/ L oca I Graph Strut ture [or feeture Predlctlms cone appearance predfct ton. Figure 5 Consider the problem of predicting the appearance of the cylinder in the image. We outline below the chain of reasoning intended for ACRONYM. (The previous predictor and planner rule set carried out a slightly less general but still powerful line of reasoning. For instance from the knowledge that the aircraft is on the runway ACRONYM deduces its image in an aerial photograph is determined up to a rotation about the vertical plus a translation. The rules necessary for a class of computations including the simple example below will be implemented over summer 1980.) C. Geclerating Restrictions During Matching In this section we give an example of an image prediction which generates restriction nodes during matching. We predict the appearance of a ribbon generated by the fuselage of figure I in an aerial image. A prediction says what a match of a local feature must imply about both the observed feature, and the aspect of the object whose image generated that feature. Checking this implication (deciding whether the new restriction node is satisfiable) provides an acceptance test 25 for a hypothesized match. For a projective imaging system the observed distance m between two points distance 1 apart on the ground is given (approximately) by: Cl m E -- h where c is a constant dependent on the focal distance of the camera and h is the height of the camera above ground. Thus if the camera is at height HEIGHT (a quantifier) and Ml and M2 are the measurements obtained for length and width from a hypothesized match of a ribbon in the observation graph, then the following constraints must hold if the match is correct: 1. (- Ml (QUOTIENT (TIMES CAM-CONST FUSELAGE-LENGTH) HEIGHT) 1 2. (= H2 KtUOTIENT (TIMES 2 CAM-CONST FUSELAGE-RADIUS) HEIGHT) 1 In general Ml and M2 will not be given exactly by the observation graph, rather an interval estimate will be supplied. Thus they can be represented by quantifiers with constraints on their upper and lower bounds. If CAM-CONST is known in advance its numeric value can be substituted into the constraints generated by the match. If it too is a quantifier, then it is just one more constrained unknown in the system. At the time of hypothesizing a match, a new restriction node is generated by adding*constraints 1 and 2 to the const aints of the restriction node associated with the prediction (see figure 5). If the new node can be shown to be unsatisfiable then the match is rejected. The following is not meant to indicate a proposed reasoning chain for ACRONYM. Rather it is illustrative of how constraints can imply that a hypothesized match is incorrect. Suppose for some hypothesized match, where CAM-CONST is known to be 100.0, the observed Ml lies between 4.0 and 5.0. Then given the constraints on the fuselage size of section II-B, the height of the camera must be between 800.0 and 1750.0 due to constraint 1 above. If this is inconsistent with a priori constraints on HEIGHT the match can be rejected. In fact a priori constraints on HEIGHT may also put further restrictions on the possible range for FUSELAGE-LECNTH. Similarly measurement M2 and constrairit 2 above will lead to restrictions on HEIGHT. If these restrictions are inconsistent with the 800.0 to 1750.0 bounds already obtained the match should be rejected. D. From Local to Global After each phase of local matching, the matcher combines local matches into more global interpretations. This involves finding consistent subgraphs of matches. Previously consistency has only concerned the existence of arcs describing relations between matched ribbons. With the introduction of constraints on quantifiers during the ribbon matching process, these too must be checked for consistency. Constraints on a quantifier at different HYPOTHESIS-MATCH restriction nodes may actually refer to different quantities in the scene. For instance each potential match for an aircraft may have constraints on FUSELAGE-LENGTH and on HEIGHT. When combining the matches for aircraft to produce an interpretation of the image, there is no reason to require that the constraints on FUSELAGE-LENGTH at these different nodes be mutually consistent. Different instances of wide-bodied passenger jet aircraft will be different lengths. However ail the constraints on HEIGHT should be mutually consistent, as there is only one HEIGHT of the camera. Sometimes when constraints on quantifiers actually correspond to different quantities in the world, it may be that these quantities should have the same value. For instance the ENGINE-LENGTH for the port and starboard engines correspond to physical measurements of different objects in the world. However since aircraft are symmetric the constraints given by the matches on possible values of ENGINE-LENGTH for each engine should be consistent. Thus when clumping the local matches for an aircraft the ENGINE-LENGTH constraints from each submatch should be checked for consistency. If they are not consistent the particular set of local matches should be rejected as inconsistent. A slot is provided in object units to represent which quantifiers matched at a lower level should be held consistent for interpretation of an object. This is the QUANTIFIERS slot as shown in figure 2. As the matcher is combining local matches it looks up the subpart tree. Any quantifier mentioned in a QUANTIFIERS slot of any ancestor of the object has its constraints copied into the restriction node for the new more global node. As each constraint is introduced it is checked for consistency. This process is not quite straightforward. Sometimes a constraint on a quantifier involves another quantifier which is not being brought into the new match. Such is the case of FUSELAGE-LENGTH and HEIGHT in the example of the previous section when a global interpretation is being made involving many aircraft. Each aircraft provides a constraint on HEIGHT, but each is in terms of the instance of FUSELAGE-LENGTH of the individual aircraft. One solution is to generate a new unique identifier for the quantifier which is not to be constrained by new constraints imposed. Its role is to ensure the continued satisifiabiiity of the local match in light of new global constraints on other quantifiers involved in that match. Other solutions exist, which may result in constraint inconsistencies being missed in return for much simplified constraint analysis. IV REMARKS We have described a single part of ACRONYM and have ignored many important issues involved in the construction of a vision system based on the representations given. In particular we have not discussed the analytic power necessary to decide whether constraint sets are satisfiable. We believe that quite weak analytic methods can lead to powerful interpretive capabilities even though they fail to detect large classes of inconsistencies. Nor have we described in detail the methods to carry out the necessary geometric reasoning. These have been discussed in E7J which includes explicit rules for symbolic geometric reasoning in states of uncertain knowledge. We have provided a representational scheme which facilitates the computaeionai processes necessary for interpretation. The scheme uses restriction graphs to provide specializations of models, to enable case analysis in reasoning about incompletely specified situations, and to manage multiple hypothesized instantiations of modeled objects in a single image. 26 ACKNOWLEDGEMENTS This research was sponsored by ARPA contract MDA-903-76-C-0206 and NSF contract DAR-78 15914. Support was also provided by the ALCOA foundation. REFERENCES 111 Arnold, R. David and Thomas 0. Binford, “Geometric Constraints in Stereo Vision,” Proc. SPIE Meeting, San Diego, July 1980. C21 Baker, H. Harlyn, “Edge Based Stereo Correlation,” Proc. ARPA Image Understanding Workshop, Baltimore, Apr. 1980, 168- 175. 131 Binford, Thomas O., “Visual Invited paper at IEEE Systems .Con.erence, Miami, Dec. 1971. Perception by Computer,” Science and Cybernetics 141 Einford, Thomas O., Proc. NSF Grantees Conference, Cornell Univ., Sep. 1979. 151 Bobrow, Daniel G. and Terry Winograd, “An Overview of KRL, a Knowledge Representation Language,” Cognitive Science 1, 1977, 3-46. 161 Brooks, Rodney A., “Goal-Directed Edge Linking and Ribbon Finding,” PYOC. ARPA tmage Understanding Worhshop, Palo Alto, Apr. 1979, 72-78. 171 Brooks, Rodney A. and Thomas 0. Binford, “Representing and Reasoning About Partiatiy Specified Scenes,” Proc. ARPA Image Understanding WorRshop, Baltimore, Apr. 1980,95-103. I81 Brooks, Rodney A., Russell Greiner and Thomas 0. Binford, “The ACRONYM Model-Based VisionSystem,” Proc. of IJCAI-79, Tokyo, Aug. 1979, 105-l 13. 191 Lowe, David, “Solving for the Paramters of Object Models from Image Descriptions,” Proc. ARPA Image Understanding Worksliop, Baltimore, Apr. 1980, 121-127. Cl01 Soroka, Barry I., “Debugging Manipulator Programs with a Simulator,” to be presented at CAD/CAM8, Anaheim, Nov. 1980. 27

1980

86

85

Sticks, Plates, and Blobs: A Three-Dimensional Object Representation for Scene Analysis Linda G. Shapiro John D. Moriarty Prasanna G. Mulgaonkar Robert M. Haralick Virginia Polytechnic Institute and State University Department of Computer Science ABSTRACT In this paper , we describe a relational model- ing technique which categorizes three-dimensional objects at a gross level. These models may then be used to classify and recognize two dimensional views of the object, in a scene analysis system. I. Introduction The recognition of three-dimensional objects from two-dimensional views is an important and still largely unsolved problem in scene analysis. This problem would be difficult even if the two- dimensional data were perfect, but the data can be noisy, distorted, occluded, shadowed and poorly segmented, making recognition much harder. Since the data is so rough it seems reasonable that very rough models of three-dimensional objects should be used in the process of trying to classify such data. In this paper we describe a relational model and discuss its use in a scene analysis sys- tem. There have been many approaches to modeling three-dimensional objects. For a comprehensive collection see the proceedings of the Workshop on Representation of Three-Dimensional Objects [13]. Also see Voelcker and Requicha [ll] and Brown [4] for mechanical design; York et.al. [14] for curved surface modeling using the Coons surface patch 151; Horn [6] and Waltz [12] for study of light and shadows; Badler et.al. [2] for study of human body modeling; and Pgin and Binford [1] and Neva- tia and Binford [7] for the generalized cylinder approach. The models we suggest are related to the generalized cylinder models, but are rougher descriptions that specify less detail about three-dimensional shape than do generalized cylin- ders. II Sticks, iptions Plates, and Blobs in Relational Dt?S- A relational description of an object consists of a set of parts of the object, the attributes of the parts, and a set of relations that describe This research was supported by the National Sci- ence Foundation under grants MCS-7923827 and MCS-7919741. how the parts fit together. Our models have three kinds of three-dimensional Parts: and blobs. Sticks are long, sticks, Plates. thin&-g onlv one significant dimension. Plates are flat- ish-wide parts consisting of two nearly flat sur- faces connected by a thin edge between them. Plates have two significant dimensions. Blobs are neither thin nor flat; they have three significant dimensions. All three kinds of parts are "near convex"; that is a stick cannot bend very much, the surfaces of a plate cannot fold too much, and a blob can be bumpy, but cannot have large concav- ities. Figure 1 shows several examples of sticks, plates, and blobs. sticks Figure 1 illust rates several examples sticks , plates and blobs. each of In describing an object, we must list the parts, their types (stick, plate, or blob), and their relative sizes; and we must specify how the parts fit together. For any two primitive parts that connect, we specify the type of connection and up to three angle constraints. The type of connection can be end-end, end-interior, end-cen- ter, end-edge, interior-center, or center-center where "end" refers to an end of a stick, "inte- rior" refers to the interior of a stick or surface of a plate or blob, "edge" refers to the edge of a plate, and "center" refers to the center of mass of any part. 28 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. For each type of pairwise connection, there are one, two, or three angles that, when specified as single values, completely describe the connec- tion. For example, for a stick and a plate in the end-edge type connection, two angles are required: the angle between the stick and its projection on the plane of the plate and the angle between that projection and the line from the connection point to the center of mass of the plate. Requiring exact angles is not in the spirit of our rough models. Instead we will specify permis- sible ranges for each required angle. In our relational model, binary connections are described in the CONNECTS/SUPPORTS relation which contains lo-tuples of the form (Partl, Part2, SUPPORTS, HOW, VLl, VHl, VL2, VH2, VL3, VH3) where Part1 connects to Part2, SUPPORTS is true if Part1 sup- ports Part2, HOW gives the connection type, VLi gives the low-value in the permissible range of angle i and VHi gives the high value in the per- missible range of angle i, i = 1, 2, 3. The CONNECTS/SUPPORTS relation is not suffi- cient to describe a three-dimensional object. One shortcoming is its failure to place any global constraints on the resulting object. We can make the model more powerful merely by considering tri- ples of parts (sl,s2,s3) where sl and s3 both touch s2 and describing the spatial relationship between sl and s3 with respect to s2, Such a des- ,ription appears in the TRIPLE CONSTRAINT relation and has two components: 1) a boolean which is true if sl and s3 meet s2 on the same end (or sur- face) and 2) a contraint on the angle subtended by the center of mass of sl and s3 at the center of mass of s2. The angle constraint is also in the form of a range. Our current relational description for an object consists of ten relations. The A/V rela- tion or attribute-value table contains global pro- perties of the object. Our A/V relations cur- rently contain the following attributes: 1) number of base supports, 2) type of topmost part, 3) num- ber of sticks, 4) number of plates, 5) number of blobs, 6) number of upright parts, 7) number of horizontal parts, 8) number of slanted parts. The A/V relation is a simple numeric vector, including none of the structural information in the other relations. It will be used as a screening rela- tion in matching; if two objects have very differ- ent A/V relations, there is no point in comparing the structure-describing relations. We are also using the A/V relations as feature vectors to input to a clustering algorithm. The resulting clusters represent groups of objects which are similar. Matching can then be performed on clus- ter centroids instead of on the entire database of models, Other relations include SIMPLE PARTS, PARALLEL PAIRS, PERPENDICULAR PAIRS, LENGTH CONST- RAINT, BINARY ANGLE CONSTRAINT, AREA CONSTRAINT, VOLUME CONSTRAINT, TRIPLE CONSTRAINT and CON- NECTS/SUPPORTS. III. Matching Relational matching of two-dimensional objects to two-dimensional models is a well-defined opera- tion. See Barrow, Ambler, and Burstall [3] for a discussion of exact relational matching, Shapiro [8] for relational shape matching, and Shapiro and Haralick [lo] for inexact matching. Our problem in scene analysis is to match two-dimensional per- spective projections of objects (as found in an image) to the three-dimensional models stored in the database. Our approach to this problem is to analyze a single two-dimensional view of an object, produce a two-dimensional structural shape description, use the two-dimensional description to infer as much as possible about the correspond- ing three-dimensional description, and then use inexact matching techniques in trying to match incomplete and possibly erroneous three-dimen- sional object descriptions to our stored three-di- mensional relational models. We decompose a two-dimensional view into sim- ple parts by a graph-theoretic clustering scheme as described in [9]. To match a two-dimensional object description to a three-dimensional model is to find a mapping from the tm-dimensional simple parts of the object to the sticks, plates and blobs of the model so that the relationships among the two-dimensional parts are not inconsistent with the relationships among the three-dimensional parts. For example, a binary CONNECTS relation can be constructed for the two-dimensional parts. For a pair (pl,p2) of three-dimensional model parts where (pl,p2,*,*,*,*,*,*,*,*,) is an element of the CONNECTS/SUPPORTS relation and a mapping h from three-dimensional model parts to two-dimen- sional object parts, if (h(pl),h(p2)) is not an element of the two-dimensional CONNECTS relation, then an error has occured. If a mapping accumu- lates too many errors from various n-tuples of various relations not being satisfied, that map- ping cannot be considered a match. As an example, suppose the three-dimensional model of a simple chair contains two plates (the back B and seat S) and four sticks (legs Ll, L2, L3, L4). The relation obtained from just the first two columns of the CCNNECTS supports rela- tion is {(S,B), (B,S) , U-J,S), (S,Ll) , (L2,S) r (S,L2), (L3,S) I (S,L3) (L4,S), (S,L4)]. Now con- sider the two-dimensional decomposition of Figure 2. We can construct the hypothetical connection relation C = {(sl,s2), (s2,sl), (s3,s2), (s2,s3), (s3,sl), (sl,s3) I (s4,s2), (s2,s4), (s4,sl), (sl,s4), (s5,s2), (s2,s5)]. Then the mapping f defined by {(S,s2), (Btsl) , (Ll,s3), (L2,s4), (L3,s5), (L4,s4)) accumulates no error while the mapping g defined by {(S,sl),(B,s2), (Ll,s3), (L2,s4), (L3,s5), (L4,s4)] accumulates error since (L3,S) is in the model, but (f(L3),f(S)) = (s5,sl) is not in C. Not all of the three-dimensional relations can be directly constructed from two-dimensional data. (If they could, the entire scene analysis problem would be much easier.) For example, only an esti- mate of whether one part supports another can be computed. Relations like PARRALLEL PAIRS and LENGTH CONSTRAINT can also be estimated. Rela- tions involving angles are probably the most dif- 29 m Figure 2 illustrates the decomposition of a two- dimensional chair by graph-theoretic clustering. ficult, since a perspective projection will change the angles between parts. Such .information should be left out of initial matching attempts and used later to try to validate a given match or to "hoose between several possible matches. The pre- cise definition of an inexact match from a two-di- mensional description to a three-dimensional des- cription is the subject of our current research. IV. Sumnary of Current and Future Research We have described a relational model for three-dimensional objects, in which the parts of an object are very grossly described as sticks, plates, or blobs. We are building a database of three-dimensional object models. The objects in the database are being clusterd into groups, using graph-theoretic clustering algorithm. Instead of comparing a two-dimensional view to every object in the database, it will be compared ini- tially only to a centroid objects in each group. Only in those groups where the unknown object is most highly related to the centroid will any full relational matching take place. Relational matching will be a form of the inexact matching we described in [lo]. The gen- eral method will be to obtain estimates of the three-dimensional relations from the two-dimen- sional shape and match these estimates against the three-dimensional models. Deriving the algorithms and heuristics for the matching is one of our most challenging tasks. References 2. Badler, NaI.I J. O'Rourke, and H. Toltzis, "A Spherical Representation of A Human Body For Visualizing Movement", Proceedings of - the IEEE, Oct. 79. 3. Barr==, A.P. Ambler, and R.M. Burstall, "Some Techniques for Recognizing Structure in Pictures", nition, S, Frontiers of Pattern Recog- Watanabe (ed),Academic Press, New York, 1972, pp. l-29. 4. Brown, C.M., A.A.G. Requicha and H.B. Voelcker, "Geometric Modeling Systems for Mechanical Design and Manufacturing", Proc. ACM 1978, Washington D.C., Dee 4:6, m -- l7l” . 5. Coons, S.A., "Surfaces for Computer-Aided Design of Space Forms," M.I.T. Project MAC, MAC-TR-41, June 1967 (AD 663504). 6. Horn, B., "Obtaining Shape From Shading Infor- mation", in The Psychology of Computer 7 Vision, (P.H. Winston, ed.), McGraw-Hill, New York, 1975, pp. 115-155. 7. Nevatia, R.K. and T.0, Binford, "Structured Descriptions of Complex Objects", Proc. Third International Joint Conference on Artificial Intelligence, Stanford, 1973.- 8. Shapiro, L.G., "A Structural Model of Shape", IEEE Transactions on Pattern Analysis and Machine Intelligence, -1-2, No. 2, -980. 9. Shapiro, L.G. and R.M. Haralick, "Decomposi- tion of Two-Dimensional Shapes by Graph- Theoretic Clustering," IEEE Transactions on Pattern Analysis and Machine Intelli- genZ&XX. PAMI- pp- lo-?Z&-ZZi.m 10. Shapiro, L.G. and R.M. Haralick, "Structural Descriptions and Inexact Matching", Tech- nical Report CS79011-R, Department of Com- puter Science, Virginia Polytechnic Insti- tute and State University, Blacksburg, VA 24061, Nov. 1979. 11. Voelcker, H.B. and A.A.G. Requicha, "Geometric Modeling of Mechanical Parts and Pro- cesses", Computer, Vol. 10, No. 12, Dec. 1977, pp.48-57. 12. Waltz, D., "Understandinq Line Drawinqs of Scenes with Shadows", in The Psychology of Computer Vision, (P.H. Winston, ed.x McGraw-Hi-w York, 1975. 13. Workshop on Representation of Three-Dimen- sional Objects, (R. W-Y, Director), University of Pennsylvania, May l-2, 1979. 14. York, B., A. Hanson, and E.M. Riseman, "Repre- sentation of Three-Dimensional Objects with Coons Patches and Cubic B-Splines", Department of Computer and Information Science, University of Massachusetts, Amherst, MA 01003. 1. Agin, G.J. and T.O. Binford, "Computer Des- criptions of Curved Objects", IEEE Tran- sactins on Computers, Vol. 25x03, -- April 1976. 30

1980

87

86

CONSTRAINT-BASED INFERENCE FROM IMAGE MOTION Daryl T. Lawton Computer and Information Science Department University of Massachusetts Amherst, Massachusetts 01003 ABSTRACT We deal with the inference of environmental information (position and velocity) from a sequence of images formed during relative motion of an observer and the environment. A simple method is used to transform relations between environmental points into equations expressed in terms of constants determined from the images and unknown depth values. This is used to develop equations for environmental inference from several cases of rigid body motion, some having direct solutions. Also considered are the problems of non-unique solutions and the necessity of decomposing the inferred motion into natural components. Inference from optic flow is based upon the analysis of the relative motions of points in images formed over time. Here we deal with environmental inferences from optic flow for several cases of rigid body motion and consider extentions to linked systems of rigid bodies. Since locality of processing is very important, we attempt to determine the smallest number of points necessary to infer environmental structure for different types of motion. I INTRODUCTION The processing of motion information from a sequence of images is of fundamental importance. It allows the inference of environmental information at a low level, using local, parallel computations across successive images. Our concern is with processing a particular type of image motion, termed optic flow, to yield environmental information. Optic flow [ll is the set of velocity vectors formed on an imaging surface by the moving projections of environmental points. It is important to note that there are several types of image transformations, caused by environmental motion, which are not optic flow. For example, image lightness changes due to motion relative to light sources, the motion of features produced by surface occlusion, moving shadows, and a host of transduction effects. The occurrance of these different types of image transformations requires explicit recognition so the appropriate inference technique can be applied for each. ---------------- This work was supported by NIH Grant NO. ROI NS14971-02 COM and ONR Grant No. NO001 4-75-C-0459. II CAMERA MODEL AND METHOD ---- The camera model is based upon a 3-D Cartesian coordinate system whose orgin is the focal point (refer to figure 1 throughout this section). The image plane (or retina) is positioned in the positive direction along, and perpendicular to, the Z-axis. The retinal coordinate axes are A and B. They are aligned with, and parallel to, the X and Y axes respectively. For simplicity and without loss of generality, the focal length is set to 1. A point indexed by the number i in the environment at time m is denoted by Pmi. The time index will generally correspond to a frame number from a sequence of images. The projection of an environmental point Pmi onto the retina is determined by the intersection of the retinal surface with the line containing the focal point and Pmi. The position of this intersection in the 3-D coordinate system is represented by the position vector Imi. In this paper, any subscripted I, A, or B, is a constant determined directly from an image. The significant relations concerning Pmi and Imi are I > ( X mi9 Yrni, Zmi J 2 > 3 > = mi 4 ) = i In the method used here, Equation 4 is used to transform expressed relations between environmental points into a set of equations in terms of image position vectors and unknown Z values. Solving these equations yields a set of Z values which provide a consistent interpretatiOn for the positions, over time, of the corresponding set of environmental points under the assumed relations. From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Retina Fig. 1 III INFERENCE FROM RIGID BODY MOTION ---- A. Arbitrary Motion of Rigid Bodies ---- The constraint equations developed for this case reflect the preservation of distances between pairs of points on a rigid body during motion. For two points i and j on a rigid body at times m and n, the preservation of distance yields 5 > - = - i i which expands into the image-based equation 6 > +2Zniznj(1nio Lj) - - 0 To determine a solution, we find the minimum number of points and frames for which the number of independent constraints (in the form of equation 6) generated equals or exceeds the number of unknown Z values. It is then necessary to solve the resulting set of simultaneous equations. Note that each such constraint is a second degree polynomial in 4 unknowns. We begin with the number of unknown Z values. For N (N>2) points in K (K>l) frames there are (NK)-1 unknown Z values. The minus 1 term reflects the degree of freedom due to the loss of absolute scale information. Thus, one of the Z-values can be set to an arbitrary value. The number of rigidity constraints generated by a set of N (N>2) points in K (K>l) frames is the product of 3*(N-2) and (K-1). The first term is the minimum number of unique distances which must be specified between pairs of points, in a body of N points, to assure its rigidity. Thus 4 points require 6 pairwise distances (all that are possible). For configurations of more than 4 points, its is necessary to specify the distance of each additional point to only 3 other points to assure rigidity. The second term is the number of interframe intervals. Each distance specified must be maintained over each interframe interval. The number of constraints is greater or equal to the number of unknowns when Thus minimal solutions (But not unique! necessarily see below) can be found when (N=5,K=2,number of constraint equationszg) or (N=4,K=3,number of constraint equations=l2), in agreement with [2]. The rigidity equations can be simplified by adding restrictions on allowable motions of environmental points. In the following sections we investigate two such restrictions. B. Motion Parallel to the XZ Plane -P-P Here the Y component of an environmental point is assumed to remain constant over time. Otherwise its motion is unrestricted. This corresponds to an observer moving along an arbitrary path in a plane, maintaining his retina at an Orientation perpendicular to the plane, with the motion of objects also so restricted. For Point i at times m and n this is expressed as 9 > Ymi = Zmi Bmi= Zni Bni “Yni 10) Z Bmi ni=Zmi- Bni This allows a substitution, for points i and j, which simplifies the rigidity constraint to 32 - - where the bracketed expressions are constants determinable from an image. This case has a direct solution using 2 points in 2 frames. To see this, consider points 1 and 2 at times 1 and 2. This yields a system of 4 unknowns: 211,212,221,222. The substitution allowed by equation 10 reduces it to a system of 2 unknowns, Zll and 212. Zll can then be set to an arbitrary value 9 reflecting scale independence. 212 is then determined from a constraint of the form of equation 11 rel$ing 212 and the evaluated variable Zll . This is a quadratic equation of 212. C. Translations The constraint expressing the translation of points i and j on a rigid body at times m and n is 12 > I= j j 13 > - i- ii where the operation is vector sub traction. This reflects the preservation of 1 eng th and orientation under translation. Setting Zmi to a constant value C, to reflect scale independence, in equation 13 yields 3 simultaneous 1 inear equations in 3 unknowns CAmi = ZmjAmj + ZniAni - ZnjAnj CBd =ZmjBmj+ZniBni-ZnjBnj C = zlTlj +z,i -Znj Thus environmental inference from translation requires 2 points in 2 frames. A potential implication of this case is for interpreting arbitrary, and not necessarily rigid body, environmental motion. If the resolution of detail and the rate of image formation relative to environmental motion are both very high, then, in general, the motion of nearby points in images can be locally approximated as the result of translational motion in the environment. D. Solving the Constraints A- The rigidity constraints are easily differentiable and can be solved using conventional optimization methods (taking care to avoid the solution where all the Z-values equal zero>. There are, however, in the case of arbitrary rigid body motion , generally many solutions. Here we consider ways of dealing with this. One way utilizes feedforward. It is crucial to note that the rigidity equations needn’t be solved anew at each point in time. If the environmental structure has been determined at time t and an image is then formed at time t+l, half the unknowns in the system of constraint equations disappear. This greatly simplifies finding a solution. Additionally, the solution process can be further simplified by extrapolation of infered motion, if enough frames have been processed . But how can the positions of the environmental points be determined initially? I> . Prior knowledge of the environment could supply the initial estimates of the relative positions. 2). There may be a small number (perhaps less than 50) of generic patterns of image motion (which may be termed flow fields), each associated with a particular class of environmental motion. For example, translational motion is characterized by straight motion paths which radiate from or converge to a single point on the retina. Other flow fields we have analyzed also have such distinguishing characteristics. These characteristics would be used to recognize particular types of image motion, associated with particular types of environmental motion, to initialize and constrain the more detailed solution process based upon solving the constraints for arbitrary rigid body motion. 3). The observer could constrain his own motion for one sampling period to a case of motion for which environmental strut ture can be unambiguously determined. For example, by stabilizing the retina of a moving observer with respect to rotations relative to a stationary environment , all image motion could be interpreted as the result of translation. Another possibility is to use more than the minimum required number of points in the inference process to supply additional constraints. 33 IV APPROACHES TO OTHER CASES OF MOTION ----- A. Sub-Minimal Rigid Configurations A sub-minimal configuration is one consisting of I,2 or 3 points over an arbitrary number of frames or 4 points in 2 frames. Human subjects can get an impression of 3-D rigid motion from displays of such configurations [4], even though there are not sufficient generated constraints, by the above analysis, for a solution. How is this possible? Other assumptions must be used for the inference. A potential one reflects an assumption of smoothness in the 3-D motion. This can be had by minimizing an approximation of the acceleration of a given point. For a point i at times t, t+l, t+2 this can be expressed as 15) Ipti- pt+l,i I - Ipt+l,i- pt+Z,i Perhaps a sum of such expressions, formed using the substitution of equation 4, should be minimized for a set of points over several time periods along with the satisfaction of expressible rigidity constraints. B. Johansson Human Dot Figures --- Experiments initiated by Johansson have shown the ability of subjects to infer the structure and changing spatial disposition of humans performing various tasks with only joint positions displayed over time [53. Here we consider how such inference could be performed. First, it is necessary to determine which points are rigidly connected in the environment. Work by Raschid C63 as shown that this is possible on the basis of image properties only, using relative rates of motion between images. That is, without any inference of environmental structure. Given the determination of rigid linkages, it is necessary to find the relative spatial disposition of the limbs. An approach is to infer environmental position, for each limb, using the rigidity constraints and optimizing the smooth motion measure discussed above. If the figure is recognized as being human, several object specific constraints can also be used. These would involve such things as allowable angles of articulation between limbs, their relative lengths, and body ACKNOWLEDGEMENTS I greatly appreciate Ed Riseman, Al Hanson, and Michael Arbib for their support and the intellectual environment they have created. I would also like to acknowledge the last minute (micro-nano-second!) heroics of Tom Vaughan, Earl Billingsley, Maria de LaVega, Steve Epstein, and Gwyn Mitchell. REFERENCES Cl1 c21 [31 c41 151 C61 Gibson, J.J. The Perception of the Visual -- World. BostoG-- Houghton Mifflin and Co. 1950. Ullman, S. The Interpretation of Visual Motion. Cambridge, Massachusetts:‘ The MIT Press. 1979. Rogers, D.F. and Adams, J.A. Mathematical Elements for Computer Graphics. McGraw-Hill Book Company. 1976. Johansson, G., and Jansson, G. "Percieved Rotary Motion from Changes in a Straight Line." Perception and Psychophysics, 1968, Vol. 4 (3). Johansson, G, "Visual Perception of Biological Motion and a Model for its Analysis." Perception and Psychophysics 14:2 (1973) 201-211. - Rashid, R.F. "Towards a System for the Interpretation of Moving Light Displays", Technical Report 53, Department of Computer Science, University of Rochester, Rochester, New York 146727, May 1979. symmetries. 34

1980

88

87

STATIC ANALYSIS OF MOVING JOINTED OBJECTS Jon A. Webb Department of Computer Science University of Texas at Austin ABSTRACT The problem of interpreting images of moving jointed objects is considered. Assuming the existence of a connectedness model, an algorithn is presented for calculating the rigid part lengths and motion of the jointed objects just from the positions of the joints and some depth information. The algorithn is proved correct. I INTRODUCTION Vision research has only recently begun considering the three-dimensional motion of jointed objects, but progress has been relatively rapid. This paper presents a method for using a very general model to discover the motion and structure of a jointed object. The method is proved correct under reasonable conditions, which are stated precisely. These condit,ions are found to be satified in most normal observation of normal jointed object movement. Jointed objects are important because they include most of the significant moving objects in this world, e.g. rigid objects, hunans, and animals. The method to be described allows the recovery of a wealth of information by a single monocular observer of a moving jointed object. This information could aid recognition from a distance. This paper, like most other research in three-dimensional motion (C1-41) * adopts the feature point model. In this model only the positions of points rigidly attached to the object are recorded. This method makes the mathematical analysis more direct. Moreover, psychological research has shown that humans can readily interpret movies of people where only the joints can be seen 15-71. It is therefore reasonable to try to construct a program that could interpret such images. II THE MODEL -- A. Introduction This paper assumes the existence of a ---------------- * This research supported by the Air Force Office of Scientific Research under grant nLpnber AFOSR 77-3190. connectedness model. This model could be constructed by the methods of [31 or by other methods under development. The jointed object model for a jointed object consists of three parts: joints, rigid parts, and feature points. The - - feature points are fixed on the rigid parts, which are connected by the joints. In this paper, it will be assumed that the jointed object forms a tree (i.e., that is has no cycles) and that the feature points coincide with the joints. The rigid parts are not allowed to bend or stretch. The lengths of the rigid parts are unknown, but are calculated by the algorithm through observation of the jointed object. A connectedness model for a humanoid figure is shown in figure 1. Feature points are indicated by letters. m ?” /i ‘e d a f I. j r’“\ k 1 i Figure 1. B. Description Input The analysis proposed in this paper applies equally well whether the central projection or parallel projection model of vision is used, but central projection will be assuned as it most accurately describes the way caneras work. The camera will be assuned to be at the or igin, with the focal plane at (O,O,f>. Figure 2 shows this model. Figure 2. The correspondence between model and image feature points must be established. The correspondence problem for moving objects has been considered in c2-41. These correspondence algorithms are based on nearest neighbor, and work 35 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. well ([31 reports 98% accuracy) for fraes with small time intervals between them. The algorithm to be described requires a z coordinate for some feature point in every frame. This point will be called the reference point. For simplicity, it will be assuned that the reference point is the same in every frame. The z coordinate of the reference point can be obtained by several means, including the support assumption (used in Cl1 for this purpose and proposed for psychological reasons in [91> but no method is entirely satisfactory. This will be discussed briefly in section IV. III THE ALGORITHM A. Introduction The algorithm treats the model the root being the reference point. this tree for the humanoid model. as a tree with Figure 3 shows The starting _ . point of a rigid part is its joint nearest the reference point (in this tree); its ending point is the joint farthest from the reference point. A first rigid part is said to be above a second if it lies on a path frcm the second to the reference point. Similarly, the second is said to be below the first. a d f h k i 1 Figure 3. The algorithn works by calculating the lengths and the positions of the ending points of the topmost rigid parts (these ending points are m, c, e, and a in figure 4). Next, rigid part lengths and ending point positions immediately below these rigid parts are calculated. The process continues until the positions of all the joints and the lengths of all the rigid parts have been calculated. The calculation of the lengths of rigid parts is done using known lower bounds on their lengths. These lower bounds are obtained from previous frames. (In the first frame a lower bound of zero is used). If the lower bound is too small to account for the observed positions of the joints, the smallest rigid part length that will work is calculated and a new lower bound is established. B. Formal Statement of the Algorithm -- For each frame, do the following for each rigid part in the tree, going from top to bottom: 1. Let the position of the starting point of this rigid part be (x,y,z), the observed coordinates of the ending point be (u,v>, and the lower bound on the rigid part length be r. If the rigid part length is exactly rr then the ending point lies on a sphere of radius r with center at (x,y,z>. At the sane time, the ending point lies on a line through the origin and (u,v,f), where f is the focal length. This situation is shown in figure 4. The coordinates of the ending point under these assumptions can easily be calculated using the quadratic formula. Figure 4. This method gives two values for the position of the end of the rigid part. These two values represent two reflections of the rigid part that could account for the observed position of the ending point. For the algorithm to work, and calculate the correct rigid part lengths, the correct reflection must be chosen. It is assumed that the correct reflection is always chosen by some process. While deciding which of the reflections is correct might be a hard problem (see section IV), once the correct reflection is chosen it can be tracked fairly easily since the two reflections normally differ greatly in the z coordinate, and in the angle they make at the starting point. 2. If the quadratic formula yields no value for the position of the end of the rigid part this means that the rigid part length must be longer than r. Calculate a new lower bound on the rigid part length by the formula (1) r = SQRT[(x-p~~)~+(y-pv)~+(z-pf)~l where (2) p= (ux+vy+fz) u2+v2+f2 The coordinates of the ending point are (Pu*Pv,Pf) l The situation giving rise to this formula is shown in figure 5. Figure 5. Whenever a rigid part length is changed, the previously calculated lower bounds on rigid part lengths below the changed rigid part become invalid, so they must be set to zero. This action introduces an order dependence into the algorithm; for the algorithm to work correctly, the proper view of a rigid part must be seen after the proper views of rigid parts above it are seen. This restriction will be discussed in greater detail later. 36 C. Experimental Results these restrictions can be removed. An experiment was run using three hand-drawn hunanoid figures and the algorithm given above. The figures were drawn with specific rigid part lengths in mind. The rigid part lengths were recovered by the algorithm to within an average relative error of about lo-15%. D. Proof of the Algorithm --- It will now be shown that the algorithm will eventually calculate the correct rigid part lengths and three-dimensional joint positions. In order to show this, these assunptions are necessary: 1. The correc known. t reflec tions of the joints must be 2. Each rigid part must be seen at some time in a position that satisfies figure 6. That is, the angle between the origin, the endpoint, and the starting point of the rigid part must be a right angle. 3. If rigid part A is above rigid part B, condition 2 must be satisfied for B after it is satisfied for A. Theorem. Under the above conditions, the givenwthm will correctly calculate the length and endpoint position for every rigid part. Proof. Let R be a rigid part. The proof will be by induction by the nLanber of rigid parts above R. If there are no rigid parts above R then R is attached to the reference point. As soon as condition (2) is satisfied for R formula (1) will correctly calculate R's length and R's endpoint will be correct. If there are any rigid parts above R then their correct lengths and endpoint positions will eventually be found. Once this has happened, conditions (3) guarantees that condition (2) will be satisfied for R, at which time formula (1) will be used to correctly calculate R's length. This completes the proof. IV EXTENSIONS TO THE ALGORITHM -- There are several restrictions placed on the data available to the system that are undesirable in the sense that hunans cannot make them in their observation of jointed objects. The most serious restrictions are the necessity of a connectedness model for the jointed object, needing a z-coordinate for the reference point in every frame, the necessity of knowing the correct reflections of the rigid parts, and the order dependence in rigid part views. These restrictions are necessary because the analysis of the moving object is only static, and does not take into account invariants in the object's motion. Dynamic analysis of the moving object is under active investigation and is yielding quite encouraging results that suggest that most, and perhaps all, of V SUMMARY A mathematical approach to the problem of jointed object observation has been presented. Given a connectedness model of the jointed object to be observed, the actual three-dimensional motion and rigid part lengths of the jointed object can be discovered by observation of the jointed object. This is done by constantly making minimizing assumptionqabout the object. Further research must take into account the actual motion of the object in a more sophisticated my. In order to overcome the deficiencies of the currently proposed method it is necessary to have a more complete understanding of how objects can be expected to move. ACKNOWLEDGEMENTS Fruitful discussions with J. K. Agtw~l. Larry Davis, Worthy Martin, and John Roach are gratefully acknowledged. REFERENCES 1. 2. 3. 4. 5. 6. 7. 8. 9. Roberts, "Machine perception L G, of three-dimensional solids.11 in Optical and electro-optical information processing, Jx Tippett, et al., Eds., 159-197. 1965. Ullman, S, The interpretation of visual motion. The MIT PreK Cambridge, MA. 19r Rashid. R F. "Lights: A study in motion.11 In Proc of the ARPA image understanding workshop, ---- Los Angeles, CA. 57-68. November 1979. Roach, J W, Determining the three-dimensional motion and model of objez from a sequence of --- -- images. Ph.D. dissertation, University of Texas at Austin, Department of Computer Science. May 1980. Johansson, G, "Visual perception of biological motion and a model for its analysis." Perception and Psychophysics, 2, (21, 201-211. 1973. - Kozlowski 9 L T and J E Cutting, "Recogni zing the sex of a wal ker from dynam ic point-l ight displays .I1 Perception and (61, 575-580. 1977. - Psychophysics, 21, Johan sson, G, "Spatio-temporal differenti ation and integration in visual mot ion percept ion." Psychological Research, 38, 379-383. 1976. Johansson, G and G Jansson 9 "Perceived motion from changes in a straight rotary line." Perception and Psychophysics, 2, (3). 165-170. 1968. - Gibson, J J, The perception of the visual world. Houghtofiifflin Co., E!ozn,950. 37

1980

89

88

A TECWIQUE FOR ESTABLISHING COMPLETENESS RESULTS IN THEOREM PROVING WITH EQUALITY Gerald E. Peterson Department of Mathematical Sciences University of Missouri at St. Louis St. Louis, MO 63121 ABSTRACT This is a summary of the methods and results of a longer paper of the same name which will appear elsewhere. The main result is that an automatic theorem proving system consisting of resolution, paramodul- ation, factoring, equality reversal, simplification and subsumption removal is complete in first-order logic with equality. When restricted to sets of equality units, the resulting system is very much like the Knuth-Bendix procedure. The completeness of resolution and paramodulation without the func- tionally reflexive axioms is a corollary. The methods used are based upon the familiar ideas of reduction and semantic trees, and should be help- ful in showing that other theorem proving systems with equality are complete. I INTRODUCTION A. Paramodulation Attempts to incorporate equality into automa- tic theorem provers began about 1969 when Robinson and Wos [6] introduced paramodulation and proved that if the functionally reflexive axioms were added to the set of clauses, then resolution and paramodulation constituted a complete set of in- ference rules. In 1975 Brand [l] showed that re- solution and paramodulation are complete even with out the functionally reflexive axioms,, Unfortunat ly, the usefulness of these results is limited because unrestricted paramodulation is a weak in- ference rule which rapidly produces mountains of irrelevant clauses. B. The Knuth-Bendix Procedure e- In 1970 Knuth and Bendix [2], working indepen- dently of Robinson and Wos, created a very effec- tive procedure for deriving useful consequences from equality units. Their process used paramodula- tion, but since it also used simplification and subsumption removal, most of the derived equalities were discarded and the search space remained small. The main defects of this procedure are that each equality must be construed as a reduction, so the commutative law is excluded, and the process works only on equality units, so most mathematical theo- ries, including field theory, are excluded. C. The Goal -- Since resolution and paramodulation constitute a complete set of inference rules, their use will provide a proof of any valid theorem if given suf- ficient (usually very large) time and space. On the other hand, the Knuth-Bendix process is effec- tive (usually small time and space) on a small class of theorems. We need to combine these two approaches and produce, if possible, an effective, complete prover. Some progress toward this goal has been re- ported. For example, the commutative law can be incorporated into the Knuth-Bendix procedure by using associative-commutative unification [4,5]. Also, restricted completeness results (i.e. the set of clauses must have a certain form) have been ob- tained for systems which appear to be more effec- tive than resolution and paramodulation [3]. D. Contributions of this Paper -- An impediment to progress toward the goal has been the lack of an easily used technique for ob- taining completeness results. We show here how the use of semantic trees can be generalized to provide completeness proofs for systems involving equality. We use this technique to obtain unrestricted com- pleteness results for a system which is thought to be fairly effective. The verification of effective- ness will require experiments which have not yet been performed. A. Semantic Trees II METHODS AND RESULTS One approach to obtaining completeness theorems is the use of semantic trees. To obtain a semantic tree T(S) for a set S of clauses, we first order the atoms of the Herbrand base I3, 87 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. say 13 = 1~1,~~'". 1. Then we build the binary tree T by giving each node at level k-l two sons labelled Bk and sBk, respectively. There will then be a one- to-one correspondence between the branches of T and the Herbrand interpretations. If the set S is unsatisfiable, then it will be falsified by every branch of T and as we move down a branch b of T we will come to a node n b at which it first becomes clear that b does not satisfy S. The node nb is called a failure node of T. The -- portion of T(S) from the root and extending up to and including the failure nodes is called the closed semantic tree for S, 'c(s). An inference node ofs a node whose children are both failure nodes. Every failure node nb has an associated clause Cb in S which caused the failure. That is, there is a ground instance CbO of C b such that if L is a literal of CbB, then QL occurs on b at or above nb with one such QL occurring at nbO It can be shown that the two clauses associated with the children of an inference node will resolve to produce a new clause C which causes failure at or above the inference node and therefore r(S U C) is smaller than -r(s). By performing a sequence of such resolutions we can eventually get the closed semantic tree to shrink to a single node and this will imply that the empty clause has been inferred. B. Incorporating Equality Problems arise when we attempt to use this pro- cess to obtain completeness results for systems in- volving equality. If S is E-unsatisfiable, then S is falsified by every E-interpretation but not nec- essarily by every interpretation. Thus it will only be on branches which are E-interpretations that failure nodes will exist in the usual sense. The other branches must be handled in some other manner. 1. Altering Interpretations The approach we use is to alter an arbi- trary interpretation I in a way such that the re- sulting interpretation I* is an E-interpretation. If I is itself an E-interpretation, then no altera- tion is needed, I* = I. The alternation is made as follows. First order 8 in a way such that each equality atom occurs before any atom which contains either side of the equality as a subterm. (Other restrictions are also needed on this order.) For an arbitrary interpre- tation I, define a partial order +(I) on 8 such that A-+B means essentially that B has been obtained from A by replacing a subterm s of A by a term t and I(s=t) = T. Now define I*(A) as I(A) if A is irre- ducible anfd as I*(B) if A-tB. 2. Substitutions For ground substitutions 0, 0' (0 = l$l+-tl,...,vk+tk)), we write 843' if 8' is identical to 8 except that one term t. of 8 has been replaced J by t! and t.+t!. J J J We say that 8 is irreducible if every term ti of 8 is irreducible. Suppose CBl and (Xl2 are ground instances of a clause C. If 81+e2 then I*(cel) = I*(Cf3,). 3. Failure Nodes Let Ib be the interpretation associated with a branch b of T(S). Then 1;; will be an E-in- terpretation and will, therefore, be falsified by some clause in S. That is, there will be a ground instance CB of a clause C in S such that I;l(Ce) is false and 8 is irreducible (Ib). (If 8 were re- ducible we could, by the previous paragraph, reduce it to an irreducible 8' such that I;(Cel) = F.) Every literal L of C8 will be falsified by 1; and there will exist a failure node nb such that %L occurs on b at or above n b with one such SL occur- ring at n b' These failure nodes can be split into two categories as follows. An R failure node, nb, -- is one such that the associated clause C is irre- ducible (Ib) (thus Ib(C) = F) and a P failure node --- is any failure node which is not an R failure node. 4. Inference Nodes The two categories of failure nodes lead to two categories of inference nodes. A resolution inference node is a node with two R failure node children ams essentially the same thing as an inference node in a semantic tree for a set without equality. A paramodulation inference node is a P failure node nb such that every equalityode ances- tor of nb has a brother which is an R failure node. 5. Summary of the Completeness Proof It is easy to show that if S has no E- model, then r(s), the closed semantic tree for S, has either a resolution or paramodulation inference node. If 'c(s) has a resolution inference node, then there will be a resolvent C of two clauses of S such that T(S U C) is smaller than T(S). If -c(s) has a paramodulation inference node n b' then there is a clause C28 such that I;(C,e) = F and C2B isreducible (Ib), say C28+E. Now C28 reduces to E using some equality s=t such that I,(s=t) = T. Since s=t occurs in the ordering 88 REFERENCES of 8 before the atom in C28 to which it applies, a node labelled s=t occurs on b above n b' This node has a brother which is an R failure node and hence there is a clause Cl0 such that s=t is a literal of Cl0 and if L is any other literal of Cl@, then Ib(L) = F. It follows that Cl6 and C28 have a para- modulant C'. This ground paramodulation can be lifted to the general level since 8 is irreducible and therefore s must start somewhere in C2. (The lifting lemma for paramodulation holds only in this case.) Thus there is a clause C which is obtained by paramodulating Cl into C2 and which has a ground instance C1 which is more reduced than C28. This greater reduction can be the basis for an ordering of the closed semantic trees involved and in the sense of this order, the tree for S U C will be smaller than the tree for S. C. Deletion of Unnecessary Clauses - 1. Subsumption clauses 2. Completeness is not lost if subsumed are deleted from S as the proof proceeds. Simplification If Cl = (s=t), C2 contains an instance so of s as a subterm, and s0e > toe for all ground stitutions 8, then the clause C = C,[to] is a sub- simplification of C2 using Cl. If a clause C has been simplified, then C may be deleted. (Our proof of this fails when the atom simplified is an equality of a certain form, but there are other reasons for believing it is still valid in this case.) D. The Final Result --- A complete system for first-order logic with equality may consist of resolution, paramodulation, factoring, equality reversal, simplification, and subsumption removal with the following restrictions. 1. Simplification and subsumption removal are given priority since they do not increase the size of s. 2. No paramodulation into variables. 3. All paramodulat ions replace s by t where for at least one ground substitution 8, se > te. 111 [21 [31 I141 151 161 D. Brand, "Proving Theorems with the Modifi- cation Method." SIAM J. Comput. 4 (1975) -- 412-430. D. E. Knuth Problems in and P. B. Universal Bendix, "Simple Word Algebras." in Leech. J. (ed.), Computational Problems in Abstract Algebras, Pergammon Press, 1970, 263-297. D. S. Lankford and A. M. Ballantyne, "The Refutation Completeness of Blocked Permutative Narrowing and Resolution." presented at the fourth workshop on automated deduction, February 1-3, 1979. D. S. Lankford and A. M. Ballantyne, "Decision Procedures for Simple Equational Theories with Commutative-associative Axioms: Com- plete Sets of Commutative-associative Reduc- tions." Technical Report, Mathematics Department, University of Texas at Austin, August 1977. G. E. Peterson and M. E. Stickel, "Complete Sets of Reductions for Some Equational Theories." to appear in JACM. G. A. Robinson and L. Wo s, "Paramodulation and Theorem Proving in F irst Order Theories with Equality." Machine Intelligence 4, American Elseviermork, 1969, 135-150. 4. If s > t then no reversal of the equality (s=t) will be necessary, and if Cl is obtained from C by reversing (t=s) then C may be deleted.

1980

9

89

BOOTSTRAP STEREO Marsha Jo Hannah Lockheed Palo Alto Research Laboratory Department 52-53, Building 204 3251 Hanover Street, Palo Alto, CA 94304 ABSTRACT Lockheed has been working on techniques for navigation of an autonomous aerial vehicle using passively sensed images. One technique which shows promise is bootstrap stereo, in which the vehicle's position is determined from the per- ceived locations of known ground control points, then two known vehicle camera positions are used to locate corresponding image points on the ground, creating new control points. This paper describes the components of bootstrap stereo. I INTRODUCTION Before the advent of sophisticated navigation aids such as radio beacons, barnstorming pilots relied primarily on visual navigation. A pilot would lookart the window of his airplane, see landmarks below him, and know where he was. He would watch the ground passing beneath him and estimate how fast and in what direction he was moving. Today, there exist applications for which a computer implementation of this simple, visually oriented form of navigation would be useful. One scenario hypothesizes a small, unmanned vehicle which must fly accurately from its launch point to its target under possibly hostile circum- stances. I I 1 2 - L TERRAIN PROFILE KNOWN FEATURES -DETERMINE VEHICLE LOCATION ---DETERMINE LOCATION OF UNKNOWN TERRAIN FEATURES TIME POSITION OF CRAFT DETERMINE DERIVED FROM POSITION OF I : I 0 0, b I - 1 1 0, b c, d t 2 I c. d I e.f Figure 1 Navigation Using Bootstrap Stereo. Cur overall approach to the problem involves providing the vehicle with a Navigation Expert having approximately the sophistication of an early barnstorming pilot. This expert will navi- gate partly by its simple instruments (altimeter, airspeed indicator, and attitude gyros), but mostly by what it sees of the terrain below it. This paper covers one aspect of the Navigation Expert, a technique which we call bootstrap stereo. II THE BOOTSTRAP STEREO CONCEPT Given a set of ground control points with known real-world positions, and given the 1ocatioIs of the projections of these points onto the image plane, it is possible to determine the position and orientation of the camera which collected the image. Conversely, given the positions and or- ientations of two cameras and the locations of corresponding point-pairs in the two image planes, the real-world locations of the viewed ground points can be determined [l]. Combining these two techniques iteratively produces the basis for bootstrap stereo. Figure 1 shows an Autonomous Aerial Vehicle (AAV) which has obtained images at three points in its trajectory. The bootstrap stereo process begins with a set of landmark points, simplified here to two points a and b, whose real-world coordinates are known. From these, the camera position and orientation are determined for the image frame taken at Time 0. Standard image- matching correlation techniques [23 are then used to locate these same points in the second, overlapping frame taken at Time 1. This permits the second camera position and orientation to be determined. Because the aircraft will soon be out of sight of the known landmarks, new landmark points must be established whenever possible. For this purpose, "interesting points" -- points with a high likelihood of being matched [3] -- are se- lected in the first image and matched inthe second image. Successfully matched points have their red- world locations calculated from the camera posi- tion and orientation data, then join the landmarks list. In Figure 1, landmarks c and d are located in this manner at Time 1; these new points are later used to position the aircraft at Time 2. Similarly, at Time 2, new landmarks e and f join the list; old landmarks a and b, which are 38 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. no longer in the field of view, are dropped the landmarks list. Once initialized from a set of known land- marks, bootstrap stereo has four components -- camera calibration, new landmark selection, point matching, and control point positioning. Because camera calibration and control point positioning have been well covered in the photogrammetric and imaging literatures (e.g., [1], [4], [S], [63), we will discuss only landmark selection and point matching in the following sections. III NEW LANDMARK SELECTION Because the aircraft rapidly moves beyond the known landmarks, new landmark points must constant- ly be established. For this purpose, "interesting points" -- points with a high likelihood of being matched [3] -- are selected in the old image of each pair, then matched with their corresponding points in the new image and located on the terrain. Matching is done on the basis of the normal- ized cross-correlation between small windows of data (typically 11 x 11) around the two points in question. Matching has trouble in areas that con- tain little information or whose only information results from a strong linear edge, therefore such areas make poor candidate landmarks. To avoid mismatches from attempting to use such aras, various measures on the information in the window have been used, including the simple statistical variance of the image intensities over the window [2] and the minimum of the directed variances over the window [3]. We have combined these into another interest measure which we call edged variance, which appears to perform better than either of its components [7]. We have defined our interesting points to be those which are local peaks in our interest measure, with a lower bound established to reject undesirable areas. Figure 2 includes some examples of the application of this interest measure. IV POINT MATCHING The actual matching of points in an image pair is done by maximizing normalized cross- correlation over small windows surrounding the points. Given an approximation to the displace- ment which describes the match, a simple spiral- ing grid search is a fairly efficient way to refine the precise match [2J. To provide that initial approximation, we have employed a form of reduction matching [3]. We first create a hierarchy of N-ary reduc- tion images. Each NxN square of pixels in an image is averaged to form a single pixel at the next level. This reduction process is re- peated at each level, stopping when the image becomes approximately the size of the correlation windows being used. Matching then begins at the smallest images, with the center point of the first image being matched via a spiral search. There- after, each matched point spawns four points around itself, offset by half a window radius along the diagonals of the window. These are mapped down to the next level of images, carrying their parent's displacement (suitably magnified) as their suggested match approximation. These matches are refined by a spiraling search before spawning new points. This process continues until the largest images are reached, effectively setting up a grid of matched points. In our implementation of bootstrap stereo, reduction matching is used to determine approx- imate registration of the images and to initial- ize the second-order match prediction polynomials. Matching of old landmarks and of interesting points to create new landmarks uses these polynomials to predict an approximate match, which is then re- fined by a local search. Autocorrelation threshold- ing is used to test the reliability of the match, then points are located more closely than the image grid permits by parabolic interpolation of the X- and Y-slices of the correlation values. V ANEXAMPLE In Figure 2, we present an example of the control-point handling portion of bootstrap stereo. The original data set, a sequence of 3 images from a video tape taken over the Night Vision Laboratory terrain model, is shown in Figure 2a. Figure 2b shows the interesting points in the first image, indicated by + overlays. If these were the control points from a landmark processor, we would use them to locate the first camera. These landmark points are next matched with their corresponding points in the second image; Figure 2c shows the successful matches overlaid on the first and second images. From the image plane positions of these points, the position and orientation of the second camera are determined. Next, the areas of the second image which were not covered by matches are blocked out and interesting points are found in the uncovered areas, as seen in Figure 2d. The old landmark points and the interesting points are then matched in the third image, as shown in Figure 2e. The old control points from the second image are used to calibrate the third camera; the camera calibrations are then used to locate the matched interesting points on the ground, forming new control points. These two steps are then repeated for subsequent pairs of images in longer sequences. VI CONCLUSICNS When an autonomous aerial vehicle must navigate without using external signals or rad- iating energy, a visual navigator is an enticing possibility. We have proposed a Navigation Expert capable of emulating the behavior of an early barn- storming pilot in using terrain imagery. One tool 39 such a Navigation Expert could use is bootstrap stereo. This is a technique by which the vehicle's position is determined from the perceived positions of known landmarks, then uses two known camera positions to locate real-world points which serve as new landmarks. The components of bootstrap stereo are well established in the photogrammetry and image process- ing literature. We have combined these, with improvement, into a workable system. We are work- ing on an error simulation, to determine how the errors propagate and accumulate. VII REFERENCES Cl3 C23 Thompson, M. M., Manual of Photogrammetry, American Society of Photogrammetry, Falls Church, Virginia, 1944. Hannah, M. J., Computer Matching of Areas in Stereo Image=, PhD,Thesis, AIM#239, Computer Science Department, Stanford University, California, 1974. Moravec, H. P., "Visual Mapping by a Robot Rover", Proceedings of the 6th IJCAI, Tokyo, Japan, 1979. Duda, R. 0. and P. E. Hart, Pattern Classi- fication and Scene Analysis, John Wiley and Sons, New York, New York, 1973. Fischler, M. A. and R. C. Bolles, "Random Sampling Consensus", Proceedings: Image Understanding Workshoe, College Park, Mary- land, April 30, 1980. Gennery, D. B., "A Stereo Vision System for an Autonomous Vehicle", Proceedings of the 5th IJCAI, Cambridge, Massachusetts, 1977. Hannah, M. J., "Bootstrap Stereo", Proceedings: Image Understanding Workshop, College Park, Maryland, April 30, 1980. Figure 2 An Example of the Control-Point Handling for Bootstrap Stereo a> The original sequence of 3 images. b) The interesting points in Image 1. c) The matched points between Images 1 and 2. d) e) The areas of Image 2 covered by matches, The control points in Image 2 matched to with interesting points found Image 3. in the uncovered areas. 40

1980

90

LOCATING PARTIALLY VISIBLE OBJECTS: THE LOCAL FEATURE FOCUS METHOD SRI Internati onal, Menlo Park, California 94 025 ABSTRACT Robert C. process is robust, because it bases its decisions on grou s of mutual1 is rela P- ively fast, % consistent features, and it ecause it concentrates on key features that are automatically selected on the basis of a detailed analysis of CAD type of models of the objects. @I. INTRODUCTION There are several tasks that involve locating partial1 relative y easy tasks, such as 4 visible objects. The 9 range from ocating a single two-dimensional object, to the extremely difficult task of locating and identifying thrF;-t;TEnsional objects jumbled together in a bin. we describe a technique to locate and identi y . gap-3 overlapping two-dimensional objects on the basis of two-dimensional models. Sequential 1,2,3] approaches have and parallel [4,51roblem een taken to solve this F ' In the sequential approach, one feature a ter another is located and as much information as possible is derived from the position and orientation of each feature. This approach is fast because it locates the minimum number of features; however, if the ob'ects are complicated, determining the or a er of the features to be located may be difvficult. Development of the location strategy becomes even more difficult when mistakes are taken into account. In the parallel approach, all the features in an image are located, and then large grou s of r-;,g;ed to recogn'ze relaxation 5,7 t 7 8 :j",;tgisto;;$- niques can be used to determine the eature'groups. This ap roach is robust because it bases its decisions on a E 1 the available information, and the location strategy is straightforward because all the features are used. For even moderately complex objects, however, the quantit approac h of data to be processed makes use of this impractical on current computers. Described here is a method called the Local Feature FOCUS (LFF), that combines the advanta es of the sequential and parallel approaches, whi 3 e avoidin ifI some of their disadvantages. This is achieve by careful analysis of the object models and selection of the best features. Belles II. LOCAL FEATURE FOCUS METHOD The basic F rinciple of the LFF method is to locate one rela ively reliable feature and use it to partially define a coordinate system within which a f roup of other key features is located. Enough o the secondary features are located to uniquely identify the focus feature and determine the position and orientation of the object of which it is a part. Robustness is achieved by using a s arallel matching scheme to make the final ecisions, and speed is achieved by carefully selecting information-rich features. The idea of concentrating on one feature is not new; it has been use P urpose vision programs he ability to generate their secondary features automatically from objet models. This automatic feature selection, when perfected, will significantly reduce the need for peo !c le wil to program recognition procedures and thus make possible quick and inexpensive application of the LFF method to new objects. OBJECT MODELS FOCUS FEATURES AND THEIR SECONDARY FEATURES I # , I OBJECT IMAGE OF EXECUTION-TIME OBJECTS- PROCESSING IDENTITIES -AND POSITIONS Figure 1 THE TOP-LEVEL BLOCK DIAGRAM As Figure 1 shows, the analysis of ob'ect,, models is performed once during training c ime and the results of the analysis are used repeatedly during nexecution time, making this a preach particularly attractive when large t num ers of objects are to be processed. In the rest of this paper, we concentrate on the training-time analysis. III. ANALYSIS The goal of the analysis is to examine a model of an object (or objects) such as the one in Figure 2, and generate a fist of focus features and 41 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Figure 2 AN OBJECT MODEL Figure 3 AN EXAMPLE TO BE PROCESSED their associated secondary features. Given this information and a picture such as the one in Figure 3, the execution-time system tries to locate occurrences of the objects. In the current implementation of the system, ob'ects are modeled as structures of regions, each o I which is bounded TK b a'sequence of line segments and arcs of circles. e execution-time system uses a maximal-cli ue ?! P ra h-matching method to locate the groups o B ea ures objects. that correspond to occurrences of the Therefore, the analysis is tailored to produce the information required by the maximal- clique matching system. In particular, the description of each secondary feature includes the feature-type, its distance from the focus feature, a,E;t;r;ist of the possible identities for the The analysis to produce this information is perfirmed in five steps: (1) (2) (3) Location of interesting features Grouping of similar features Rotational symmetry analysis of each object (4) Selection of secondary features (5) Ranking of focus features. The purpose of the first step is to generate the set of all features of the objects that could be located at execution time. Typical features include holes, corners, protrusions, and intrusions. For the model in Fi ure 2, the set of features contains all 14 interna !? holes. In the second step, the set of features is P artitioned into subsets of "similar" features. eatures are defined to be similar if they are like1 For to be indistinguishable at execution time. t e model in Figure 2, feature detectors can a distinguish at most,,three "small holes," and large ty es !? ho es. $f holes: "slots," Therefore, the set of interesting features is partitioned into three subsets, each defining a possible focus feature. In the third step, T$ s mmetry anal I sis a complete rotational e rotationa of each object is performed [12]. symmetry is used to determine the number of structurally different occurrences of each feature. Because the model in Figure 2 is twofold rotationally symmetric, the features occur in pairs, the members of which are indistinguishable on the basis of the relative positions of other features of the object. Instead of four types of small holes, there are only two, one on the axis between between the slots and one off that axis. Figure 4 SECONDARY FEATURES FOR SMALL HOLES The fourth step in the analysis is the most corn licated. P The oal is to select secondary fea ures for each ? ecus feature. The secondary features must distinguish between the structurally different occurrences of the focus feature and determine the position and orientation of the object. In Figure 2, for example, given an occurrence of a small hole, what nearby features could be used to determine whether it is one of the holes on the axis or off of it? There are two slots close to the small hole on the axis and on1 one near the off-axis occurrence. In addition, g t e slots are at different distances from the holes. Let Dl be the distance between the on-axis small hole and its slots and let D2 be the distance from the off-axis small hole to the nearest slot. Figure 4 shows circles of radii Dl and D2 centered on the two different types of small holes. Tabulated below are the feature occurrences that are sufficient to determine the type of the small hole and compute the position and orientation of the object. ON-AXIS SMALL HOLE -- Two slots at Dl No slots at D2 OFF-AXIS SMALL HOLE -- No slots at Dl One slot at D2 The analysis in step 4 locates secondary features in two substeps. First it performs a rotational symmetry analysis centered on each structural1 different occurrence of a focus feature. T% is analysis builds a descri R tion of the object in terms of of features t at are similar and equidis ?! roups ant from the focus feature. Figure 5 shows the groups of features produced by the current system when focusing on one of the small holes. In the second substep, the anal sis iteratively selects groups of features from t ese K descriptions to be included in the set of secondary features associated with tne focus feature. Groups 42 REFERENCES are selected for their contribution to identifying an occurrence of the focus feature or determining the position and orientation of the object. Figure 5 FEATURE GROUPS ABOUT A SMALL HOLE The fifth and final step in the training-time analysis is the rankin 8 oal is to determine t e a of the focus features. The order in which the focus eatures should be checked at execution time. The current system simply ranks them according to the number of secondary features required at execution time. IV. DISCUSSION The LFF method is a simple combination of the se uential and parallel approaches. 9. It offers the re lability of a arallel approach and most of the speed of a sequen 7. la1 approach. The s eed is achieved by usin to define a coor 5 the location of the !f ecus feature inate system within which the other features are located. Quickly establishing a coordinate system significantly reduces the time required to find secondary features. The utilit reliabilitv of 4 of the LFF method depends on the ocatinn focus features and the number of structurally-different occurrences of these features in the objects. Fortunately, most industrial features. !i arts have good candidates for focus. time so they he problem is to find them at tra;;ing can be used at execution time. fact, the more information gathered at training IX-X-,, the more efficient the system at execution Also, as the training-time analysis is made more'automatic correspondingly less time is required of a human programmer. The current implementation of the training- time analysis forms the basis for a coS@n;;ly automatic feature selection system. extensions are ! ossible. could select ex ra features to guarantee tha For example the s si;z % execution-time system would function proper1 even if a prespecified number of mistakes were ma % the feature detectors. e by The system could use the orientation of a focus feature, if it exists, to determine the orientation of the feature-centered coordinate system. The system could also select two or.more groups of features at one time, which is necessary for some more difficult tasks such as distinguishing an object from its mirror image. Finally, the system could incorporate the cost and reliability of locating a feature in the evaluation of the feature. In conclusion, the LFF method is a combination of the sequential and parallel approaches that B rovides s eed and reliability for many two- imensiona P location tasks. The automatic selection of features makes it particularly attractive for industries such as the aircraft industry that have hundreds of thousands of different parts and cannot afford a special-purpose program for each one. hl r4 L-31 t41 II51 b1 [71 b1 191 r.103 Cl11 IAl S. Tsuji and A. Nakamura, "Recognition,,of an Ob'ect in a Stack of Industrial Parts, IJSAI-75 Tbilisi, Proc. (August i 975). Georgia, USSR, pp. 811-818 S. W. Holland, "A Programmable Compute: Vision System Based on Spatial Relationshi General Motors Research Pub. s, GMR-20 8 (February 1976). 1; W. A. Perkins, "A Model-Based Vision System for Industrial Parts," IEEE Transactions on Computers, Vol. C-27, pTp';;f26-143 (Yebruam . A. P. Ambler et al., "A Verzatile Computer- Controlled Assembly System Proc. IJCAI-73, Stanford, California, pp. 29m(August 1973). S. W. Zucker and R. A. Hummel. "Toward a Low- Level Description of Dot Clusters: Labeling Edge, Interior, and Noise Points," Computer Graphics and Image Proc ssing Vol. y, No. 5, R. C. Bolles, "Rgbust Feature Matching Through Maximal Cliques, Proc. SPIE's Technical Symposium on Imagi~plications Ior AUtOmatea Xidustrial inspection an?f-Kssembly, Wasnlngton, D.C. (April 19'19) . S. T. Barnard and W. B. Thomnson. llDi.SDaratV Analysis of Images )( To appear in IEEE- " Transactions h n Pattern Analysis amachine lntellieence ulv 1Ytw). -- R. 0. Duda and P. E. Hart, "Use of the Hough Transform To Detect Lines and Curves in Pictures," CACM, Vol. 15, No. 1, pp. 11-15 (January 1977)T S. Tsuji and F. Matsumoto, "Detection of Ellipses by a Modified Hou h Transformation," IEEE Transactions on Compu ers, ? -Pp. '/'('I-'(81 (Au~St 1 Y'(U) Vol. C-27, No. . J. T. Olsztyn_and L. Rossol, "An A gl,;;;:on o~~~o;~;x~ Vision to a Simulated Ap T IJCPR-73, Washington, D.C. Oct6bee3y D. F.,,McGhie, "Programmable Part Presenta- ~~~~iligence Research A plied to Industrial in the SRI Ninth Report on Machine Automation, pp. 39-44 &gust 1979). R. C. Bolles, "Symmetry Analysis of Two- Dimensional Patterns for Computer Vision," Proc. IJCAI-79, Tokyo, Japan, pp. 70-72 mst 43

1980

91

INTERFERENCE DETECTION AND COLLISION AVOIDANCE AMONG THREE DIMENSIONAL OBJECTS* N. Ahuja, R. T. Chien, R. Yen, and N. Bridwell Coordinated Science Laboratory University of Illinois at Urbana-Champaign Urbana, Illinois 61801 ABSTRACT Two methods for detecting intersections among three dimensional objects are described. The first method involves detecting overlap among the projections of the objects on a given set of planes. The second method uses a three dimension- al octree representation of the objects. Intersec- tions are detected by traversing the trees for the obstacles and the moving objects. Application of the methods to collision avoidance is discussed. I INTRODUCTION An important problem in the robotics manipula- tion of its environment is that of following colli- sion free trajectories when objects have to be moved. A representation or model of the spatial configuration of objects is necessary to obtain safe solutions, i.e. those which avoid collisions. In addition, a capability?0 plan an efficient trajectory, say, one following the shortest avail- able path between the starting and the goal con- figurations, is desirable. In either case, a procedure is required to decide if an arbitrary configuration involves occupancy of a given region of space by multiple objects. The problem of interference detection usually involves static objects whereas collision detec- tion refers to a situation where at least one ob- ject is in motion. Collision detection may be viewed as a sequence of intersection checks among appropriately defined static objects, and at appropriate time intervals. Thus the basic prob- lem appears to be that of detecting intersection among a set of objects. Given a configuration of objects, it is not hard to develop an algorithm to check any inter- sections among the objects. However, in a reason- ably complex environment and for reasonable speeds of object manipulation, e.g. translation, rotation, etc., the availability of a limited computational power may demand efficient procedures. At the heart of the design of such procedures lies the need for a good representation of the objects. Algorithms must be developed that take advantage of the properties of chosen representation to efficiently track the dynamic environment. solid in terms of a set of relatively simpler, planar patches used to approximate its surface. For the case of convex objects, Comba [21 obtains a pseudo-characteristic functioninterms of expres- sions for the planar patches. The function as- sumes nonpositive values in regions that approxi- mate the parts of the space where all the objects intersect. Maruyama 151 compares the minimal boxes containing the objects. Boyse 111 considers three different types of objects: solids, surfac- es and containers. Intersections are detected by checking if an edge of one object intersects a face of another. Collisions are detected by checking interference among the obstacles and the surfaces traced by the edges and faces of the mov- ing objects. Udupa cl11 uses two different representations for the manipulator and the environment. The ob- stacles are modelled by polyhedra (not necessarily convex) in Cartesian space. The manipulator links of the Scheinman arm are approximated by the mini- mum bounding cylinders. Abstractions are intro- duced by replacing the links by their axes and enlarging the obstacles proportionately, by retain- ing only the major link, etc. Lozano-Perez and Wesley [41 also grow the obstacles and shrink the moving parts such that collisions in the modified representation occur if and only if they occur in the original space. It is clear that the representation of the objects plays a major role in determining the feasibility and performance of any intersection or collision detection method using that representa- tion. This paper discusses two methods of repre- sentation with the aim of making the resulting interference checking operations more efficient. Section II examines a method based upon a set of two dimensional projections of the objects. It uses a conservative criterion to detect the occur- rence of interference. Section III discusses a representation of the objects by regions of a fix- ed shape, but varying sizes determined by a recur- sive decomposition of the space into successively smaller regions. Section IV outlines the use of the representations described in Sections II and III to collision avoidance. Section V presents some concluding remarks. In the past, mos,t of the related work has used polyhedral approximation of objects [1,2,4, 5,111. This facilitates a treatment of a complex From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. II PLANAR PROJECTIONS A planar projection of a three dimensional configuration of objects will always show an over- lap between the projections of any objects that intersect. However the reverse is not necessarily true, i.e., overlapping projections do not neces- sarily correspond to intersecting objects. Such false alarms may sometimes be decreased by consid- ering projections on a collection of planes. A configuration then is adjudged to be safe if each pair of objects has nonoverlapping projections on at lease one plane. However, for any given number and choice of planes, spatial arrangements of non- interfering objects may be devised whose project- ions overlap. An increase in the number of planes may only decrease, but not eliminate, the probabil- ity of erroneous decisions. The error may also be lower for objects with simpler shapes, e.g., convex objects. A. Basic Method ~- We will make the usual and reasonable assump- tion that objects may be represented by polyhedra. Each polyhedron is uniquely described by the coor- dinates of, and the adjacency relationships among its vertices. The projection of a polyhedron on any plane is determined by the projections of the vertices, using the original adjacency relation- ships. The set of vectors OV! for various i defines the vertices of the projecsion. To determine the shape of the polygonal projection, we must deter- mine its border edges. This is done by obtaining the convex hull of the projections of the vertices. For convex polyhedra, the edges of the convex hull will be a subset of the set of straight lines Vl V! where V. and V. are adjacent vertices. On the J other hind, noaconvex polyhedra may give rise to nonconvexpolygonalprojections. The convex hull of the corresponding projected vertices will include edges V! V! where Vi and Vj are not adjacent. The actual or er of the projection can be found by i: a replacing all such edges VI Vi by a sequence of edges V!- V! , V! VI-, ..t ,JV!- Jl J? J2 J3 VI-, where k > 2, j, = 1, j, = J, and Vj Jk-l Jk and V. r Jr+l are adja- cent, 1 5 r I k-l (fig. 1). Obtaining the convex hull of N projected vertices takes O(N log N) time ClOl. This also determines the overall complexity of obtaining a polygonal projection, since the com- putations required to obtain the vertices Vi are O(N). Thus, assuming that the fraction of edges of a polygonal projection that form concave cavities is low, the entire set of operations takes O(N log N) time. Given a set of M planes, projections of all the objects are obtained on each of the planes. The polygons corresponding to a given pair of ob- jects to be examined for intersection are then checked for overlap in the different planes. If a plane is found in which the two objects do not overlap, noninterference between them is guaran- teed. Otherwise the objects may or may not inter- sect. Shamos LlOl gives an O(m+n) algorithm to find the intersection of a convex m-gon with a convex n-gon. Shamos [lOI also describes a simpler algo- rithm suggested by Hoey. We use the latter. We are not interested in obtaining the exact region of intersection, but only in knowing if two given polygons intersect. For an m-gon and an n-gon, this needs only O(m+n> computations compared to the O(m log m) and O(n log n) computations required to obtain the individual polygons. When the projection of an object is nonconvex, we must extract a description in terms of convex polygons in order to make use of the above algo- rithm. One obvious way is to use the convex hull of each nonconvex m-gon. However this introduces an additional source of false alarm since the cav- ities in an object, appearing as concavities in its projection, are treated as solids. An alternative way is to decompose a noncon- vex polygon into a set of convex polygons (fig. 2). Each of these convex polygons must then be con- sidered in each of the intersection tests where , 45 Figure 1 A nonconvex polyhedron and its nonconvex poly- gonal projection. To obtain the actual pro- jection from its convex hull, the dotted line must be replaced by the remaining two sides of the triangle. A /e WY : ..A.*.. . . . . . . / Figure 2 Decomposition of a nonconvex polygonal projection into con- vex polygons. the parent polygon is involved. Pavlidis [71, and Feng and Pavlidis [31 give algorithms for con- vex decomposition. Schachter [91 constructs a partial Delaunay triangulation of an m-gon to obtain an O(rm> algorithm, where r is the number of concave vertices. B. Improving Efficiency The complexity of intersection detection in- creases linearly with the number of projection planes used, since each plane must be examined for overlap until a zero overlap is found, or the planes are exhausted. To avoid applying the intersection algorithm to pairs of polygons that are far apart, coarse tests of intersection on the envelopes of the polygons may be used. For exam- ple, rectangular envelopes may be obtained by identifying the points having minimum and maximum x or y coordinate values. Similarly, a circular envelope may be obtained in terms of the diameter of the polygon. An overlap between two envelopes can be detected in constant time. However, ob- taining either of the two envelopes itself takes O(m) time for an m-gon, and hence such a coarse test may be of marginal utility since the inter- section algorithm for polygons is already linear in the total number of vertices. Suppose we are considering a pair of objects that do not intersect. Also suppose that their projections on at least one of the planes do not overlap. Then we would like to discover such a plane at the earliest, It may be useful to order the planes for examination according to some mea- sure of the likelihood that a given plane demon- strates the desired relationships. An example of such a measure is the total black (polygonal) area. It takes O(m) time to compute the area of an m-gon. Thus the planes may be ordered for examination by performing a computation on each plane which is linear in the number of vertices in the plane. The choice of the appropriate planes depends upon the given configuration. In general, a mini- mum of three planes would appear to be desirable to work with three dimensional objects. The com- putation of projections is trivial when the three orthogonal planes are used. The three projections are obtained by successively replacing one of its three coordinates by zero. Other planes with convenient orientations may also be used. III THREE DIMENSIONAL REPRESENTATIONS Extensions of the methods for representing two dimensional images C8lmay be used for the rep- resentation of three dimensional objects. Thus MAT (medial axis transform), generalized cylinders and recursive subdivision of space may all be used among others. In this paper, we will be concerned with the third of these methods. A. Octrees Just as the plane is recursively divided into squares in the quadtree representation of the images [12, 131, the three dimensional space may be subdivided into octantsL143. We start with the entire space as one block. If a block is com- pletely contained within the object whose repre- sentation is sought, it is left alone. Otherwise, it is divided into eight octants (fig. 3a) each of which is treated similarly. The splitting contin- ues until all the blocks are either completely within or completely outside the object (fig. 3a), or a block of minimum allowed size is reached, reflecting the limits on resulution. A method of subdivision of space was also employed by Maruyama c61. However, he used rectanguloids. A rectangu- lar block was divided into two, along one of the three axes as determined by the region of inter- section of the block with the object. The recursive subdivision allows a tree des- cription of the occupancy of the space (fig. 3b). Each block corresponds to a node in the tree. Let us label a node black or white if it corresponds to a block which is completely contained within the (black) object or the (white) free space, respectively. Otherwise the node is labelled gray. The gray nodes have children unless they are of the minimum allowed size, in which case they are relabelled black. The free space covered by such nodes is thus treated as part of the object in order to be conservative in detecting interference. The final tree has only black or white leaves. Figure 3 (a) An object and its representation by recursive subdivision of space into octants. (b) Octree for the object in (a). north-east, The north-west, south-west and the south-east oc- tants in the upper layer correspond to the children having labels 1, 2, 3 and 4, respect- ively. The nodes corresponding to the octants in the lower layer are labelled 5, 6, 7 and 8. Dark (white) circles indicate black(white) leaves. The children nodes are arranged in increasing order of label values from left to right. 46 B. Interference Detection Suppose we are given the octree representa- tions of a pair of objects. Clearly, the objects intersect if there exists at least one pair of corresponding nodes in the two trees such that one of them is black, and the other is black or gray. Thus the existence of interference can be deter- mined by traversing the two trees in parallel. Let A and B denote a pair of corresponding nodes at any time during the traversal. If either A or B is white, we do not traverse their children, and continue the traversal of the remaining nodes. The depth of the traversal along each path down from the root is determined by the shallower of the two paths terminating in a white leaf. The time required is proportional to the number of nodes in the subset of the nodes traversed. To detect interference among n objects, n> 2, n trees must be traversed. The traversal below a certain node in a tree stops only if either the node is white, or the corresponding nodes in the remaining n-l trees are all white. The time re- quired depends upon the actual number of nodes visited. IV COLLISION AVOIDANCE The use of interference detection in robotics manipulators lies in planning trajectories of mov- ing objects in a given environment. The occupancy of any part of the space by more than one object can be used as indicative of a collision. To avoid a collision its imminence must be foreseen. Therefore, intersections should be detected in a modified space. Suppose that it requires at least a distance d to brake the speed of a moving object or to change its course to avoid a collision. The value of d may depend upon different factors in- cluding the speed. Then we must detect any obsta- cles within a distance d along the path of a mov- ing object. For the obstacles off the path of the object, the required safe distance can be easily computed in terms of the relative location of the obstacles. Similarly, for two moving objects the required safe distance is 2d if they are approach- ing each other head on, and larger otherwise, given by the sizes of the objects, their speeds and directions of motion. To use intersection detection for collision avoidance, the moving objects are grown Cl,111 by a thickness d. Any intersection among the modi- fied set of objects produces a timely warning of a possible collision. The representations of the (static) obstacles are computed only once. Thus, we have a fixed set of projections of the obstacles on the planes being used, and a single octree representing the space occupied by the obstacles. Each moving ob- ject is represented separately. Every time an interference check has to be made, the representa- tions of the moving objects are obtained. Thus, in case of the planar projections, the current state of each of the moving objects is projected on each of the planes, which already have the appropriate projections of the obstacles. The usual procedure to check intersection is then car- ried out. In case of the octree representation, a new tree is generated for each of the moving ob- jects. A parallel traversal of these trees and the obstacle tree detects any interference present. The above checks for intersection are applied to a snapshot of the potentially continuously varying spatial configuration of objects. Hence care must be taken to ensure that no collisions will occur between any two successive executions of the algorithm. This requires that not only should we detect any existing intersections, but also note if some objects are too close, and might collide before the instant when the algorithm is applied again. The safe distance D between any two objects is proportional to the speeds and relative loca- tions of the objects involved. The interval T between two successive applications of the algo- rithm is controlled such that the objects do not come closer than D during the time interval T. This is accomplished by growing the objects further by a distance D. Then the problem of collision avoidance only involves periodic intersection de- tection among the modified objects. V CONCLUDING REMARKS We have discussed two approaches to detecting interference and collision among three dimensional objects. The first method involves detecting over- laps among projections of the objects on a given set of planes. It may thus be viewed as a two and a half dimensional approach. The criterion used for detecting interference is a conservative one, since for any given set of planes there exist many spatial configurations of noninterfering objects whose projections overlap on each of the planes. The second method uses a three dimensional octree representation of the objects, and does not suffer an information loss from the reduction in dimen- sionality like the first method. Here, the inter- ference is detected by a parallel traversal of the octrees for the obstacles and for each of the moving objects. Computer controlled manipulation often invol- ves the use of a manipulator such as the Scheinman arm. Its degrees of freedom are given by the boom length and the joint angles. These parameters thus define a natural representation of the arm. In particular, the updating of the representation as the arm moves becomes trivial. However, for the other objects in the environment, the above representation is not very useful. In addition, the origin of the cylindrical coordinate system above moves with the manipulator. Therefore, the representation of the whole environment must be modified for each new position of the manipulator as noted by Udupa [ill. The use of the Cartesian space for the representation of all objects re- quires regeneration of the octree for the moving objects only. Use of hardware such as proximity sensors may significantly improve the efficiency of the proce- dures. For example, a "cluttered zone" signal from a certain sensor may channel the available computational power to a detailed investigation of the desired region of space, leaving the relative- ly safe areas unexamined. This essentially amounts to being a hardware implementation of comparing coarse envelopes of objects in parallel. The computation of projections of the poly- hedra in Section II may also be replaced by cameras in the appropriate planes. This will provide real time projection generation. The projections are then treated as a collection of two dimensional binary images. The projections of the moving ob- jects are tracked and their positions checked against the projections of the obstacles. We have not addressed the problem of finding good trajectories for moving an object from a giv- en source position to a given goal position. The representations discussed here, in conjunction with the planning strategies discussed in [4,11], may be used to develop the desired trajectories. Ex- periments employing the methods described in this paper, and two Scheinman arms are being carried out currently. ACKNOWLEDGEMENTS This work was supported in part by the United States Department of Transportation and Federal Aviation Administration under Contract DOT FA79- WA-4360 and the Joint Services Electronics Program (U. S. Army, U. S. Navy and TJ. S. Air Force) under Contract N00014-79-C-0424. Cl1 c21 c31 c41 c51 [61 REFERENCES J. W. Boyse, "Interference Detection among Solids and Surfaces," Comm. ACM 22, January 1979, pp. 3-9. P. G. Comba, "A Procedure for Detecting Inter- sections of Three-Dimensional Objects," Jnl. ACM, 15, July 1968, pp. 354-366. H. Feng and T. Pavlidis, "Decomposition of Polygons into Simpler Components," IEEE Trans. Comp. C-14, June 1975, pp. 636-650. T. Lozano-Perez and M. A. Wesley, "An Algo- rithm for Planning Collision-free Paths among Polyhedral Obstacles," Comm. ACM 22, October 1979, pp. 560-570. K. Maruyama, "A Procedure to Determine Inter- sections between Polyhedral Objects," Int. Jnl. Comp. Inf. Sci. 1, 3, 1972, pp. 255-266. K. Maruyama, "A Procedure for Detecting Inter- sections and its Application," University of Illinois Computer Science Technical Report No. 449, May 1971. c71 [81 c91 Cl01 Cl11 Cl21 Cl31 Cl41 T. Pavlidis, "Analysis Pattern Recognition 1, of Set Patterns," 1968, pp. 165-178. A. Rosenfeld and A. C. Kak, Digital Picture Processing, Academic Press, New York, 1976. B. J. Schachter, "Decomposition of Polygons into Convex Sets," IEEE Trans. Comp. C-27 November 1978, pp. 1078-1082. , M. I. Shamos, Computational Geometry, Springer-Verlag, New York, 1977. S. Udupa, "Collision Detection and Avoidance in Computer Controlled Manipulators," Proc. 5th Int. Joint Conf. Art. Intel., Cambridge, Massachusetts, 1977, pp. 737-748. G. M. Hunter and K. Steiglitz, "Onerations on Images Using Quadtrees,"-IEEE-Trans. Pattern Analysis Mach. Int. 1, April 1979, pp. 145- 153. A. Klinger and C. R. Dyer, "Experiments in Picture Representation using Regular Decom- position," Computer Graphics and Image Pro- cess% 5, 1975, pp. 68-105. C. L. Jackins and S. L. Tanimoto, "Ott-trees and their Use in Representing Three- dimensional Objects," University of Washing- ton Computer Science Technical Report, January 1979. 48

1980

92

AUTOMATED INSPECTION USING GRAY-SCALE STATISTICS Stephen T. Barnard SRI International, Menlo Park, California ABSTRACT A method for using gray-scale statistics for the inspection of assemblies is described. A test image of an assembly under inspection is registered with a model image of a nondefective assembly and the two images are corn ared on the basis of two statistical tests: a X 3 test of the two marginal gray-level distributions and the correlation coefficient of the joint distribution. These tests are made in local subareas that correspond to important structure, such as parts and subassemblies. The subareas are compiled in an off-line training phase. TheX* measure is most sensitive to missi?ng or damaged parts, whereas the correlation coefficient is most sensitive to mispositioned parts. It is also possible to detect overall lighting changes and misregistration with these measures. Two examples are presented that show how the tests detect two types of defects. I INTRODUCTION Binary machine-vision techniques have received a great deal of attention for industrial inspection [1,2,3,41. High-contrast lighting and thresholding may be used to obtain an accurate silhouette that can be processed at video rates to yield useful features, such as area, perimeter, centroid, and higher moments. In addition, structural information is available in the geometric relationships between the local features of the outline (holes, corners, and so on). This kind of information is sometimes sufficient for some industrial automation (IA) tasks, such as part identification and acquisition. Other tasks, however, are not so easily approached. Although many simple parts can be adequately represented by a mere outline, most assemblies cannot because they are typically composites of several overlapping parts or subassemblies. Binary techniques will not be effective in such cases because thresholding will not, in general, separate the important components. Thorough binary inspection of even simple parts may not be feasible if one wishes to find defects in surface finish or other types of defects not limited to the outline. Gray-scale techniques have lately received more attention [5,6,7]. Compared to binary methods, there is a great variety of ways to use gray-scale information. This paper describes an approach for exploiting gray-scale information for inspection in a very basic form. Statistical tests of marginal and joint intensity distributions are used to compare test assemblies with an ideal model assembly. These tests are very efficiently computed and they are sensitive to defects and discrepancies that are not easily handled in the binary domain. II REPRESENTATION We use a representation that directly specifies the expected appearance of the assembly. Important structures of the assembly are represented as subareas that are tagged for special consideration. Using this representation, an image of a test assembly is more-or-less directly compared to a model with a minimum of preprocessing. It is necessary to do some sort of photometric and geometric normalization to the test image to bring it into correspondence with the model. At the lowest level, an assembly is represented by a gray-scale image. At the highest level, an assembly is represented by a list of named subareas, each of which corresponds to a particularly meaningful segment. Attached to each of these subareas are properties that specify the important characteristics; these characteristics identify the corresponding segment as "good" or "defective." Ideally, these subareas could have arbitrary shapes and sizes, but, for now, think of them as rectangular windows. Each has a specific location and size in the normal reference frame. The inspection system begins with a representation of an ideal assembly, called the model. This may be constructed interactively in a training phase. Given a low-level representation of a test case (i.e., an image), the system proceeds to build a high-level representation by comparing segments of the test image to segments of the model. The first step is to bring the test image into geometric registration with the model image. This is not strictly necessary. We could directly compare regions in the normal reference frame 49 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. (i.e., in the model image) with regions in the translated and shifted reference frame. Nevertheless, geometric registration simplifies further processing by establishing a one-to-one correspondence between model pixels and test pixels. We have assumed that the positional variation of the test assemblies is restricted to translation and rotation. We must therefore determine three parameters--Ax, Ay, and 6~ . There are several ways to approach this problem. Binary techniques may be adequate to determine the normalization parameters to subpixel accuracy with a system such as the SRI vision module [3]. In the gray-scale domain, one may search for maximal cross- correlation, although this will probably be very time-consuming without special hardware. A potentially more efficient method is to find distinguishing local features in the test image and then match them to their counterparts in the model b1- Once the translation and rotation has been determined it is easy to register the test image using a linear interpolation [g]. III STATISTICAL COMPARISON Two statistical measures that are useful for comparing model and test subareas are theX* test and the correlationP*coefficient. They are both extremely efficient and simple to implement, and they are sufficiently different to distinguish two broad classes of defects. A. The X2 Test - -- TheX2 test measures the difference between two frequency distributions. Let hm(k) be the frequency distribution of gray-scale intensities in a model window. Let ht(k) be the frequency distribution of a test window. We can consider h to be a hypothetical ideal distribution. The X 2m test gives a measure of how far h+, deviates from the hypothetical distribution h,. The significance of the test depends on the number of samples. x2 z-y& ((h,(k)-h+) I2 --------------- . k ht(k) This yields a measure of difference, but, to be consistent with what follows, we want a measure of similarity. Let f = e XL/C where c is some positive constant. 7 is a measure of the similarity of two distributions (in the X2 sense). If the distributions are identical, then 7 will be unity; if they are very different, t will be close to zero. B. The Correlation Coefficient Let hmt be the joint frequency distribution of the model and test windows. That is, hmt(u,v) is the frequency with which a model pixel has gray value u and its corresponding test pixel has gray value v. Let ml be the mean of h, and m2 be the mean of ht. Let 0, be the standard deviation f12 be the standard deviation of ht. of h,, and The central moments of the joint distribution h mt are p(i,j> = 1 c ( (Xm(k)-ml)i * (Xt(k)-m2)j) --- n k where X,(k) and Xt(k) are gray values of the kth pixel in the model image and and the test image, respectively. The correlation coefficient, p , is cc OJ) P ------- 7 O2 p is in the interval [-1,1]. If it has the value +l or -1 the total "mass" in the joint frequency distribution must lie on a straight line. This will be the case when the test image and the model image are identical, and P will be +l. In general, if there is a linear functional dependence between the test and model windows, P will be +l (or, in the extremely unlikely case that one window is a llnegativetl of the other, P will be -1). If the windows are independent distributions, however, P will be 0. We can reasonably expect that intermediate values will measure the degree of dependence between the two windows. c. r=p 7 is not sensitive to the location of pixels. It simply measures the degree of similarity between two marginal distributions. P , on the other hand, measures the degree to which the model pixels agree with their corresponding test pixels; therefore, it is sensitive to location. This implies that T is a good test for missing and severely damaged parts (for they are likely to change the distribution of the test pixels compared to the distribution of the model pixels), while P is a good test for the proper location of parts (for misalignment will change the joint distribution). A systematic change in lighting can also be detected. T would be small because the lighting change would change the intensity distributions in the test windows, but P would be large because the test and model distributions would still be well- correlated. If this observation were made for only one window, it would not be very meaningful. However, if we make the reasonable assumption that most of the windows locate regions that are not defective, this leads to the observation that a systematic pattern of small 7 and large P indicates a lighting change. Small detectable mi sregistration errors are also Small misregistration would produce large 7 because the margi .nal d istributions of the test windows would not be much different from the IV EXPERIMENTS model windows. On the other hand, P would be smaller than if there were poor registration because the windows would not correlate as well. The same result for a single window would be caused by a misplaced part, but, again using the assumption that most of the windows locate non- defective regions, a systematic pattern of large r and small P over many windows would indicate an overall misregistration. These relationships are summarized in Table 1. Table 1 Defect Pattern vs. 7 and P * OK Missing Misplaced Lighting Registration **++**++~+*****~**+***~**~~~~**~****~~*~~*~~~~**~ * * r + LARGE SMALL LARGE SMALL LARGE * * (SYSTEMATIC) * P" LARGE SMALL SMALL LARGE SMALL * * )E D. A Two-Stage System - The gray-scale statistics discussed above provide a basis for an integrated minimal system for inspection that is composed of two modules--a training module that is u&d off-line and that allows a human operator to build a high-level description of the assembly, and an inspection module that matches images of test assemblies to the model built in the training phase. In the training phase the operator works with an image of an ideal assembly, identifying the important parts that are likely to be defective and describing the modes of defects that may be relevant. For example, the location of a particular part may not be precisely fixed, but rather permitted to range over a rather large area. In this case the operator might indicate that 7 (the location insensitive measure) is a relevant test, but not P. In another case there may exist areas that have extremely variable appearance, perhaps because of individual part identification markings, and these areas might be excluded from testing altogether. In the normal case, however, a part will be fixed in one location and orientation, perhaps with some tolerance, and the operator will merely specify allowable limits for 7 and P. The on-line inspection phase is entirely automatic. The knowledge about the assembly collected in the training phase is applied to specific test assemblies and a judgment is made as to how well what is known fits what is seen. The 7 and P measures are computed for each area and are used to derive probability estimates of the various types of defects. We have tried the f and p tests on two assemblies. Figure 1 and Table 2 show the results for a water pump. The upper left portion of Figure 1 is the "model" image of a nondefective pump in a normal position and orientation. Several windows representing important parts of the assembly are also shown. The upper right portion of Figure 1 is a "test" image of a defective assembly. dark pulley in the center is missing. The round, In the lower left part of Figure 1 the test image has been registered with the model. The lower right image is the difference between the model image and the registered test image, and has been included to _ indicate how close the registration is. Table 2 shows the t windows. Note that 7 and P are both very small and p values for the various for the (missing) pulley compared to the other (nondefective) parts, just as predicted. 7 is also small for the window representing the total assembly because this includes the defective pulley. Figure 1. Table 2 Pump Statistics * T (c=800) * o * *++**++*+**JHC~******************~*** Total 8 0492 Y .801 Pulley + .236 8 .354 * <= Defect Link * .981 * .824 * spout * -919 * -904 * Clip1 * .862 9 0879 * Clip2 + 0978 * .780 * Clip3 * ,949 * .898 * ++*~**~**+*+***~***+~~~~~~*~*~~~* 51 Figure 2 and Table 3 show the results for a hardcopy computer terminal. In this case the write head is not positioned in the correct place. Note that the window for the "head" includes the entire area where it might be located. As predicted, the 7 value is high, while the P value is small. In practice this might not be considered a defect because it may be permissible for the head to be located anywhere along the track. If this were the case, the window could be tagged to indicate that P is not a relevant test. Figure 2. Terminal Table 3 Terminal Statistics Jc 7 (c=800) * p * *+++****~++++******8~*~~~**~*~~~~~ Total * 0746 * .g10 * Platen * 0674 * .868 * Head * .890 * 0458 * <= Defect Keys * 0740 * 0923 * *~~*8~*~~~~~****~~8**~~~~*~~~~~ REFERENCES 1. M. Ejiri, T. Uno, M. Mese, and S. Ikeda, "A Process for Detecting Defects in Complicated Patterns,n Computer Graphics and Image Processing, Vol. 2, pp. 326-339 (1973). 4. 5. 6. 7. 8. 9. G. J. Gleason and G. J. Agin, "A Modular Vision System for Sensor Controlled Manipulation and Inspection," A.I. Center Technical Note 178, SRI International, Menlo Park, California (March 1979). G. J. Gleason, A. E. Brain, and D. McGhie, "An Optical Method for Measuring Fastener Edge Distances," A.I. Center Technical Note 190, SRI International, Menlo Park, California (July 1979). M. L. Baird, "SIGHT-I: A Computer Vision System for Automated IC Chip Manufacture," IEEE Trans. on Syst., Man, Cybern., Vol. SMC- 8, Nomebruary 197'c W. Perkins, "A Model Based Vision System for Industrial Parts," IEEE Trans. Comput., Vol. C-27, pp. 126-143 (1978). R. J. Woodham, "Reflectance Map Techniques for Analyzing Surface Defects in Metal Castings," Ph.D. Dissertation, Massachusetts Institute of Technology, Cambridge, Massachusetts (September 1977). S. T. Barnard, "The Image Correspondence Problem," Ph.D. Dissertation, University of Minnesota, Minneapolis, Minnesota (December 1979). A. Rosenfeld and A. Kak, Digital Picture Processing (New York: Acamrn76). 2. W .C. Lin and C. Chan, "Feasibility Study of Automatic Assembly and Inspection of Light Bulb Filaments," Proc. 1445 (October 19?r IEEE, Vo1.63, pp. 1437- 52

1980

93

HUMAN MOVEMENT UNDERSTANDING: A VARIETY OF PERSPECTIVES Norman I. Badler Joseph O'Rourke Stephen Platt Mary Ann Morris Department of Computer and Information Science Moore School D2 University of Pennsylvania Philadelphia, PA 19104 ABSTRACT Our laboratory is examining human movement from a variety of perspectives: synthesis of anima- ted movements and analysis of moving images. Both gestural (limb and hand) and facial movements are being explored. Key Words and Phrases: Human movement, move- ment representation, motion analysis, computer vision, facial expression, facial analysis, con- straint network. I HUMAN MOVEMENT SYNTHESIS Our laboratory is examining human movement representation from a variety of perspectives, in- cluding synthesis of three-dimensional animated movements and analysis of moving images. These broad areas are further refined into gestural (limb and hand) and facial movements since very different sorts of actions take place in jointed skeletal movement and "rubber-sheet" facial distortions. Our human body model r5](Fig. 1) and hand model [2](Fig. 2) are based on spherical decompositions of three-dimensional objects. Our goals have been to develop representations for human movements which permit both analysis and synthesis, are "com- plete" in terms of the range of human movements, and yet provide a "high-level" interface suitable for specifying or describing movement. Many of these issues are addressed in a recent survey by Badler and Smoliar [6], so we shall emphasize only the more recent work in human motion analysis and facial expression synthesis. II HUMAN MOVEMENT ANALYSIS There have been a rather small number of at- tempts to analyze complex movement presented in time-varying images. Rashid [13] uses moving light displays to track body joints and infer three- dimensional connections. Tsotsos [14] describes non-articulated shape changes. Badler [l] attempts conceptual descriptions of rigid jointed movements. Recently O'Rourke [9,10] describes a computer sys- tem which accepts a sequence of two-dimensional images of a solid, three-dimensional body [5] per- forming some motion sequence (Fig. 3). The output of the system is a description of the motion as coordinate-versus-time trackings of all body joints and as movement instructions suitable for controll- ing the simulation of a human body model [3,4,15]. The simulation includes a detailed model of a human body which incorporates the structural relationships between the body parts and the physical limitations to relative movement between the parts, This model is used to guide the image analysis through a pre- diction/verification control cycle. Predictions are made at a high level on the basis of previous analysis and the properties of the human model. The low level image analysis then verifies the pre- diction and the model is adjusted according to any discrepancies found or new knowledge acquired. This cycle is repeated for each input frame. The information extracted from the image is in- tegrated into the current position of the model through a constraint network operating on real- valued coordinate regions in three-dimensional space. Possible positions of body features are represented as unions of orthogonally-oriented rectangular boxes. Relationships among body parts-for example, distance constraints imposted by the body skeleton- are enforced in the network. As new joint positio- nal information is extracted from the image it is added to the network and its geometrical conse- quences immediately propagated throughout the net- work, Only head, hands, and feet are located in the image space , yet all remaining body joints may be tracked by the geometric inference process in the network. Figure 4 shows the constraint boxes for each joint of the body given the moving images of Fig. 3. III FACIAL ANIMATION We are also investigating the representation and simulation of the actions performable on the human face, The face presents an interesting prob- lem for simulation, as it is composed of mostly in- dependent sections of non-rigid masses of skin and muscle, such as the areas around the eyes, mouth, and cheeks, This type of action is basically dif- ferent from gross body movement in that a facial action will affect the visible results of other actions performed in the same area of the face. 53 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. Our internal representation of the face is based on FACS, the Facial Action Coding System [7]. The system categorizes basic actions performable and recognizable (by a human notator) on the face. It is also easily translated into a state-descrip- tion of the face, in terms of muscle contractions. A complete recognition and simulation system for the face would consist of a camera, computer processing to obtain an idealized internal representation of the action, and a simulation of the action perform- ed on a graphic output device. Once the camera image is obtained, analysis is performed to produce the AU (FACS Action Unit) state of the face. This analysis is relatively simple, as it consists of identifying the presence/absence of "features" such as wrinkles and bulges on the face. (Note that this analysis does not require "recognition" of a particular face, just good comparison between suc- cessive images of the same face.) The current tech- nique under investigation uses an improved method of "rubber-sheet" matching [8]. Each AU effects only a small set of muscles; their union gives the muscle-status of the face. The specified muscle contractions are then simulated on the face. The face is represented by a network of points and interconnecting arcs (Figs. 5 and 6) 1121. It also has a higher level of organization which par- titions the network into the skin surface and spe- cific muscles. (It is this muscle organization which distinguishes our work from that of Parke [XL].) The skin is a "warped" two-dimensional surface of points and arcs. The points represent the basic surface, while the arcs contain informa- tion specific to their locale, such as the elasti- city ("stretchiness") of the skin between the arc's points. Stretching the skin (by contracting a mus- cle) causes first local motion, followed by propa- gation of the skin distortion. Muscles are also points, connected to but beneath the skin. They are distinguished by being the initiation of the distortion of the skin surface. An AU is thus mere- ly a set of muscles, with appropriate magnitudes of initial force of contraction. IV FUTURES Our research into human movement understanding has the joint goals of achieving conceptual des- criptions of human activities and producing effect- ive animations of three-dimensional human movement from formal specification or notation systems 161. One application of this effort is in the synthesis and analysis of American Sign Language. A proto- type synthesizer for ASL is being designed to faci- litate experimentation with human dynamics and movement variations which have linguistic import for facial and manual communication. ACKNOWLEDGEMENTS The support of NSF Grants MCS76-19464, MCS78- 07466, and O'Rourke's IBM Fellowship are gratefully acknowledged. lx1 I: 23 13J PJ [5J 171 b31 El ho3 IKl l-12J 1133 t-143 REFERENCES Badler, N.I. Temporal scene analysis: Concep- tual descriptions of object movements. Tech. Rep. 80, Dept. of Computer Science, Univ. of Toronto, Feb. 1975. Badler, N.1, and J. O'Rourke. Representation of articulable, quasi-rigid, three-dimensional objects, Proc. Workshop on Representation of Three-Dimensional Objects, Philadelphia, PA, May 1979. Badler, N.I., J. O'Rourke, and B. Kaufman. Special problems in human movement simulation, Computer Graphics Summer 1980. Badler, N.I., J, O'Rourke, S.W. Smoliar, and L, Weber, The simulation of human movement by computer, Tech. Rep. Dept. of Computer and Information Science, Univ. of Pennsylvania, July 1978. Badler, N.I., J. O'Rourke, and H. Toltzis. A spherical representation of a human body for visualizing movement. IEEE Proceedings 67: 10 (1979) pp. 1397-1403. Badler, N.I. and Smoliar, S.W. Digital repre- sentations of human movement. Computing Sur- veys 11: 1 (1979) pp. 19-38. Ekman, P. and W. Friesen. Facial ActionCoding System. Consulting Psychologists Press,Palo Alto, CA. 1978. Fischler,M,A. and R.A. Erschlager. The repre- sentation and matching of pictoral structures. IEEE Tr. on Computers C-22: 1 (1973). O'Rourke, J. Image analysis of human motion. Ph.D. Diss. Dept. of Computer and Information Science, Univ. of Pennsylvania, 1980. O'Rourke, 3, and N.I. Badler. Human motion analysis using constraint propagation. To appear IEEE-PAM1 November 1980. Parke, F.I. Animation of faces. Proc. ACM Annual Conf. 1972, pp. 451-457. Platt, S. Animating facial expressions. MSE Thesis, Dept. of Computer and Information Science, Univ. of Pennsylvania, 1980. Rashid, R. Lights: A system for the inter- pretation of moving light displays. To appear IEEE-PAMI, Nov. 1980. Tsotsos, J. A framework for visual motion analysis. Ph.D. Diss. Dept. of Computer Science, Univ. of Toronto, 1980. [15] Weber, L., S.W. Smoliar, and N.I. Badler. An architecture for the simulation of human move- ment, Proc. ACM Annual Conf. 1978, pp. 737- 745, 54 Simulated movement Figure 4. Constraint boxes for mo vements in Fig. 3. showing some of th butline and muscles 55 Figure 7. Upper porti muscles; t has been pu right brow.

1980

94

AN OPTIMIZATION APPROACH FOR USING CONTEXTUAL INFORMATION IN COMPUTER VISION Olivier D. Faugeras Image Processing Institute University of Southern California Los Angeles, California 90007, U.S.A. ABSTRACT Local parallel processes are a very efficient way of using contextual information in a very large class of problems commonly encountered in Computer Vision. An approach to the design and analysis of such processes based on the minimiza- tion of a global criterion by local computation is presented. INTRODUCTION The problem of assigning names or labels to a set of units/objects is central to the fields of Pattern Recognition, Scene Analysis and Artificial Intelligence. Of course, not all possible names are possible for every unit and constraints exist that limit the number of valid assignments. These constraints may be thought of as contextual information that is brought to bear on the particu- lar problem, or more boldly as a world model to help us decide whether any particular assignment of names to units makes sense or not. Depending upon the type of world model that we are using, the problem can be attacked by discrete methods (search and discrete relaxation) or continuous methods (continuous relaxation). In the first case our contextual information consists of a description of consistent/compatible labels for some pairs, or more generally n-tuples of units. In the second case the description includes a numerical measure of their compatibility that may or may not be stated in a probabilistic framework. Initial estimates of likelihoods of name assignments can be obtained from measurements performed on the data to be analyzed. Usually, because of noise, these initial estimates are ambiguous and inconsistent with the world model. Continuous relaxation (also sometimes called probabilistic relaxation or stochastic labeling) is thus concerned with the design and study of algorithms that will update the original estimates in such a way that ambiguity is decreased and consistency (in terms of the world model) is increased. This work was supported in part by the Defense ARPA contract F-33615-76-C-1203. The author is with the Image Processing Institute and Department of Electrical Engineering, USC, Los Angeles, California 90007. More precisely, let us denote by "Lithe finite set of N units and by ztthe finite set of M possible labels. In the discrete case, the world model consists of an n-ary relation RC("L(x,qn. The fact that the n-typle {(ul,~,),...,(un,~n)} belongs to R means that it is valid to assign name ai to unit Ui for i=l,...,n. In the continuous case, the world model consists of a function c of WQP into a closed interval [a,b] of the real line: c: (QxZ~)~ + [a,bl In most applications [a,bl=[O,l] or [-l,l] and n=2. The numbers c(ul,~l,...,un,Rn) measure the compa- tibility of assigning label Ri to unit ui for i=l ,***, n. Good compatibility is reflected by large values of c, incompatibility by small values. We will present in this paper two ways of measuring the inadequacy of a given labeling of units with respect to a world model and show that these measures can be minimized using only local cooperative computation. We will compare this approach with the original probabilistic relaxation scheme proposed by Rosenfeld, Hummel and Zucker [3] and a matching scheme proposed by Ullman [6]. To conclude the section, we will discuss the possibility of using Decentralization and Decomposition techniques to alleviate the curse of dimensionality and show how the Optimization approach can be extended very easily to the analysis of multilevel, possibly hierarchical, systems. We will not discuss in this paper any specific application. For an early application to scene analysis and discussion of some of the issues addressed in this paper, see [2]. For recent surveys, see [l] and [4]. For an applica- tion to graph matching, see [18]. I. Basic Optimization Based Probabilistic Relaxation Scheme We assume that attached to every unit ui are measures of certainty pi(a), for kin gthat can be thought of loosely as probabilities c Pi(R) = ' (1) R in ;;e 56 From: AAAI-80 Proceedings. Copyright © 1980, AAAI (www.aaai.org). All rights reserved. The wor_ld model is embedded in a function c mapping m .aL into [O,l], Again, c(ul,R,u2,m) measures the compatibility of calling unit ul,~ and unit u2,m. This function also allows us to define a topological structure on the set of units by assigning to every unit Ui and label R in ZT!Z. a set Vi(R) of related units uj for which these exists at least one label m in 2 such that c(ui,R,uj,m) is defined. A compatibility vector di is then computed for every unit ui that measures the compatibility in each label R in Xwith the current labeling at related units in Vi(a). The simplest way of defining Qi(a) is [l]: c Qi(a) = +I uj in v (a) Qij(R) (2) i where /Vi(g) 1 is the number of units in the set Vi(a) and Qij(") is given by: Qij (‘) = c C(Ui,R,U j ,m)Pj (ml (3) m in 2 Loosely speaking, Qi(a) will be large if for many units U* in Vi(a), the compatible labels (that is the labgls m such that c(ui,R,uj,m) is close to 1) have high probabilisties, and low otherwise. In some cases the compatibility coefficients may be given a probabilistic interpretation, that is c(ui,R,uj,m) is the conditional probability pij(R/m) that unit ui is labeled R given that unit uj is labeled m. The next step in designing the Relaxation scheme is to specify a way of combining the two sources of information thatwe can use, i.e. the initial probabilities and the contextual informa- tion, to update the label probabilities. This updating should result in a less ambiguous and more compatible overall labeling in a sense that will remain vague until later on. Rosenfeld et al. [3] proposed the following iterative algorithm: for every unit ui and every label R in di set p(n+l) (a) = i (4) The denominator of the right hand side is simply a normalizing factor to ensure that numbers p("+l)(~) still add up to one. Intuitively, the libels R whose compatibility Qi(a) is larger than others will see their probability increase whereas the labels with smaller compatibility will see their probability decrease. One criticism with this approach is that it does not take explicitly into account measures of the two most important characteristics of a labeling of units, namely its consistency and its ambiguity. Faugeras and Berthod [5,7,81 have proposed several such measures and turned the labeling task into a well defined optimization problem which can be solved by local computations in a network of processors. We saw before that we can asso$iate with every unit Ui a probakility vector pi and a compatibility vector Qi whose components are given by equation (2). In general, the vectors di are not probability vectors in that their components do not sum to 1. This can be easily changed by normalizationand we can define: qi(a> = Qi (‘) c Q; (4 m in i I The vectors G, are now probability measure of coi&sistency for unit u; can be defined as the vector norm1 (5) vectors and a (local measure) (6) where ll*Ilcan be any norm (in practice the Euclidean normb Similarly a local measure of ambiguity can be defined as Hi = R in ~pi(")(l-pi(")) = '-ll~ill~ (7) where ~~*~~2 is th e Euclidean norm. Combining the two measures yie Ids a local criterion Ji = clCi + BHi (8) where c1 and 8 weight the relative importance we attribute to ambiguity versus consistency. A global measure can then be defined over the whole set of units by averaging the local measures. Using the arithmetic average for example, we define J= Ji (9) all units ui The labeling problem can then be stated as follows: given an initial labeing {pi '('))of the set of units Q, find the local minimum of the function J closest to the initial conditions, subject to the constraints that the vectors p. are probability vectors. More precisely, &is implies that c pi(R)=1 and pi(~)>0 for all units - R in 2 U. 1 (gal Aside from the fact that we are now confronted to a well-defined mathematical problem which can be tackled using Optimization techniques, we are also sure that some weighted measure of inconsistency and ambiguity is going to decrease. 57 As pointed out in [8], one minor drawback with the definition (6) of consistency is that its n$nimiz%tion implicitly implies the minimization of qi and pi and therefore the maximization of the entropy term H. (equation (7)). Thus there is an inherent problhm with the definition (8) in the sense thai consistency and ambiguity tend opposite directions. One very simple way resolving that contradiction is to define measure of both ambiguity and consistency J! = 1 -Gi l ;;i where 0 denotes the vector inner product, definition of a global criterion proceeds before: J' = c J! all units u. 1 1 and the labeling problem can be stated as to go in of a local as (10) The now as (111 (&I, replacing J with J'. This is similar to the minimal mapping theory developed by Ullman [6] for motion correspondence. Given two image frames with elements ui in the first one (our units) and element k (our names) in the second one, he studied the problem of selecting the most plausible correspondence between the two frames. Defining the cost qi(k) of pairing element ui with element k and the variables pi(k) equal to 1 if Ui is paired with k and 0 otherwisephe rephrased the motion correspondence probl em as a linear programming problem by defining the cost function (LPI 513, = c c all element u. all element k Pi(k)qi(k) 1 (12) which is precisely equation (11). The important difference between criteria J' and JA is that the costs qi(k) in (12) are not functions of the variables pi(k) whereas in (11) they are. In particular,minimizing J' is not an LP problem, in general. Nevertheless, the parallel between the two approaches is interesting and confirms that what we called the compatibility coefficients qi(k) defined in Eq. (5) are also a measure of the satisfaction/profit implied in assigning name k to unit u.. 1 II. Computational Requirements: Locality, Parallelism, Convergence As described in [7,9], imagine that we attach to the set Q of units and the sets Vi= U R in & ‘i CR) a simple network, that isapair <G,R> where G is a connected graph and R a set of processors, one for each node in the graph. There is a one to one correspondence between the units and the nodes of the graph on one hand, and the nodes of the graph and the processors on the other hand. This in turn implies that there is a one to one correspon- dence between the neighbors of the ith processor ri, i.e., the processors that are connected by arcs of G to r., and units in V.. 1 1 As shown in [7,8], the minimization of criteria J or J' can be achieved by using only local computation. More precisely, denoting by &7 (a function of all the vectors $i) either criterion J or J', we can attach to every unit ui a local gradient vector &-k ='Fi(cj) a Si (13) where Fi is a function of the vectors sj of units u. in the set Vi of neighbors previously defined. Eiplicit formula for the functions F. can be found in [4,5,7,8]. The iterative scheme is then defined as $n+ll +(n> i = Pi + p P. n l (14) where p is a positive stepsize and Pi a linear project?on operator de$e;T$ped by the constraints imposed on the vector pi [5,7], (for example that it is a probability vector). The main point is that both functions Fi and operator Pi can be computed by processor ri by communicating only with neighboring processors (local computation) while guaranteeing that the cost function 8 will decrease globally. It was stated before that a large amount of parallelism can be introduced in the process of minimizing criteria J and J'. This is achieved by attaching to every unit ui a processor ri connected only to processors r. attached to units u. related t0 Uim The global csiterion can then be Minimized by having processors ri perform simple operations mostly in parallel while a simple sequential communication process allows them to work toward the final goal in a coordinated fashion. If nonetheless our supply of processors is limited, we may want to split our original problem into several pieces and assign sequentially our pool of processors to the different pieces. The net result has of course to be the minimization of the original global criterion and some coordination must therefore take place. Solutions to this problem can be found in the so-called Decomposition and Decentralization techniques which have been developed to solve similar problems in Economics, Numerical Analysis, Systems Theory and Optical Control [12,13,14,15]. Decomposition techniques proceed from an algorithm standpoint: we are confronted with a problem of large dimensionality and try to substitute for it a sequence of problems of smaller dimensionalities. Decentralization techniques take a different viewpoint: we are confronted with a global problem and have at our disposal P decision centers. The question is whether it is possible to solve the global problem while letting the decision centers solve only local problems. The structure of criteria J and J' as sums of local measures allows us to develop both types of techniques [121. The key idea is to partition the set of units. For detailed algorithms, see [16]. III. Extension to Hierarchical Systems, Conclusions The optimization approach presented in Section I can be extended to the case where several labeling problems are present and embedded in a pyramid or cone structure with, for example, L levels. The different levels can be the same picture at different spatial resolutions as in [17] or represent different states of abstraction. For example the lowest level could be the edge element level, then the link level [lo], then the level dealing with elementary shapes like straight lines, ellipses, cubits, etc... These different levels form a multilevel system, each level having to solve a stochastic labeling problem. command vector for level i, that is Gi Let Vi be the is a NiMi dimensional vector, if there are Ni units and Mi possible classes, obtained in concatenating the probability vectors $j, j=l,...,$i. At leyel i we have to minimize a criterion J;(v,.$,.....v,). The fact that criterion Ji depends&up& the c&&and vectors at other levels accounts for the inter- action between the levels. A natural, but not always rational, way of solving this multilevel problem is to assume that every level i (i=l,... ,L) considers as given the command vectors of the other levels and minimizes its own criterion. The result is a nongcooperative equilibrium [12l or Nash point (ul,...,uL) verifying: Ji(;l,...,zi l,;i,:i+l,...,:L) < J.(;l,...& 1, -1 + -f vi>"i+19"'> L :I for all i and c.. This notion can certainly be criticized beca?ise by cooperating each of the L levels can, in general, improve its situation compared with the non-cooperative case. In other words, (61 t$e following situation is possible: if ,..+.,uL) set (Ui,.. i-y a Nash point, there exists another .,uL) of command vectors such that Ji(Zi ,...,$) < Ji(;l ,...,:L, for all i. This introduces the notion of Pareto points which, intuitively, are optimal in the sense that it is impossible to find another set of L command vectors that will decrease all criteria. It is possible to show that under very general conditions [12], Pareto points can be obtained by minimizing only one criterion! In other words if ?i=(?il,...,zL) is a Pareto point, then there exists L positive number Al ,...,AL such that 3 is a minimum of criterion J$ ,...,-;,) = L XiJi(;l,...,GL) i=l the ‘i’s can therefore be interpreted as weigh fat tors the different levels have agreed upon. .ting Another interesting possibility is to assume a hierarchical structure within the L levels, level 1 being the lowest and level L the highest. We then have a cascade of optimization problems similar to what happens in the price decentraliza- tion technique men$ioned i+n section II, where level 1 considers v2,... ,vL as given and computes + u1 = mJn Jl(Gl,c2,... ,:,I 3 This defines zl as a function of !2,... ,; . Then levzl $ solve5 th$ probl$m of minimizing kri$erion .JJ~ul(v2,...,vL),v2,...,vL) with respect to v2, . . . Even though the theory of hierarchical multi- level systems is still in its infancy it has been recognized for some time now [ll] that it carries the possibility of solving many difficult problems in Economics, Physiology, Biology [13,14,15], Numerical Analysis and Systems Theory [12], Optimal Control. It is clear that this theory is relevant to Image Analysis. In conclusion, we think that probabilistic relaxation techniques will play a growing role in the near future as building blocks of more and more complex vision systems. The need to quantify the behavior of these relaxation processes will become more and more pressing as the complexity of the tasks at hand rapidly increases and the global optimization framework offers a solid basis for this analysis. REFERENCES El1 PI [31 [41 [51 [61 [71 L.S. Davis and A. Rosenfeld,%ooperating processes for low-level vision: a survey," TR-123, Department of Computer Sciences, University of Texas, Austin, January 1980. H.G. Barrow and J.M. Tenenbaum, "MSYS: A System for Reasoning About Scenes," Tech. Note 121, AIC-SRI Int., Menlo Park, Ca., 1976. A. Rosenfeld, R.A. Hummel and S.W. Zucker, "Scene Labeling by Relaxation Operations," IEEE Trans. on Syst., Man, and Cybern., SMC-6, No. 6, pp. 420-453, June 1976. 0-D. Faugeras, "An Overview of Probabilistic Relaxation Theory and Applications," Proceedings of the NATO ASI, Bonas, France, June-July 1980, D. Reidel Publishing Co. O.D. Faugeras and M. Berthod, "Scene Labeling: An Optimization Approach," Proc. of 1979 PRIP Conference, pp. 318-326. S. Ullman, The Interpretation of Visual Motion, MIT Press, 1979. O.D. Faugeras and M. Berthod,"Improving Consistency and Reducing Ambiguity in Stochastic Labeling: An Optimization Approach," to appear in the IEEE Trans. on Pattern Analysis and Machine Intelligence, 1980. 59 [8 1 DOI Ull WI D31 1141 WI 1161 El71 W3 1 M. Berthod and O.D. Faugeras, "Using Context in the global recognition of a set of objects: an optimization approach," 8th World Computer Congress (IFIP 80). S. Ullman, "Relaxation and Constrained Optimization by Local Processes," Computer Graphics and Image Processing, 10, pp. 115- 125, 1979. S.W. Zucker and J.L. Mohammed, "A Hierarchi- cal System for Line Labeling and Grouping," Proc. of the 1978 IEEE Computer Society Conference on Pattern Recognition and Image Processing, pp. 410-415, Chicago, 1978. M.D. Mesarovic, D. Macho and Y. Takahara, Theory of Hierarchical Multilevel Systems, Academic Press, 1970. J.L. Lions and G.I. Marchuk, Sur Les Methodes Numeriques En Sciences Physiques Et Econo- Collect miques, ion Methodes Mathematiques de L'Informatique, Dunod, 1976. Goldstein,"Levels and Ontogeny,"Am. Scientist 50, 1, 1962. M. Polanyi, "Life's Irreducible Structures,11 Science 160, 3884, 1969. D.F. Bradley, "Multilevel Systems and Biology - View of a Submolecular Biologist,11 in Systems Theory and Biology (M.D. Messarovic, ed.), Springer 168. O.D. Faugeras, "Decomposition and Decentrali- zation Techniques in Relaxation Labeling," to appear in Computer Graphics and Image Processing, 1980. A.R. Hanson and E.M. Riseman, of Natural Scenes, in A. Hanson and Segmentation E. Riseman, eds., Computer Vision Systems, Academic Press, NY, 1978, 129-163. O.D. Faugeras and K. Price, "Semantic Description of Aerial Images Using Stochastic Labeling," submitted to the 5th ICPR. 60

1980

95

SWIRL: AN OBJECT-ORIENTED AIR BATTLE SIMULATOR Philip Klahr, David McArthur and Sanjai Narain;: The Rand Corporation 1700 Main Street Santa Monica, California 90406 ABSTRACT ROSS, an object-oriented language that has evolved over the last two years as part of the knowledge-based simulation research at Rand [1,3,6,7,8,91. We describe a program called SWIRL designed for simulating military air battles between offensive and defensive forces. SWIRL is written in an object-oriented language (ROSS) where the knowledge base consists of a set of objects and their In the following sections we discuss the goal of SWIRL, outline the main objects in the air-battle domain, and note how those objects and their behaviors map onto the ROSS objects and ROSS behaviors that constitute the SWIRL program. We discuss some of the problems encountered in designing an object-oriented simulation, and present our solutions to these problems. associated behaviors. problems we encountered We discuss some of the and in designing SWIRL present our approaches to them. I INTRODUCTION Object-oriented programming languages such as SMALLTALK [2], PLASMA [4], and DIRECTOR [5], as well as ROSS [S], enforce a 'message-passing' style of programming. A program in these languages II THE GOAL OF SWIRL - ---~ The goal of SWIRL is to provide a prototype of a design tool for military strategists in the domain of air battles. SWIRL embeds knowledge about offensive and defensive battle strategies and tactics. SWIRL accepts from the user a simulation environment representing offensive and defensive forces, and uses the specifications in its knowledge base to produce a simulation of an air battle. SWIRL also enables the user to observe, by means of a graphical interface, the progress of the air battle in time. Finally, SWIRL provides some limited user aids. Chief among these are an interactive browsing and documentation facility, written in ROSS, for reading and understanding SWIRL code, and an interactive history recording facility for analyzing simulation runs. This, coupled with ROSS's ability to easily modify simulation objects and their behaviors, encourages the user to explore a wide variety of alternatives in the space of offensive and defensive strategies and to discover increasingly effective options in that space. consists of a set of objects called actors that interact with one another via the transmission of messages. set of message Each actor has a set of attributes and a templates. Associated with each message template is a behavior that is invoked when the actor receives a message that matches that template. A behavior is itself a set of message transmissions to other actors. Computation is the selective invocation of actor behaviors via message pasSing. This style of computation is especially suited simulation in domains that may be thought of as to consisting of autonomous interacting components. In such domains one can discern a natural mapping of their constituent components onto actors and of their Indeed, interactions experts in many onto message transmissions. may find the domains object-oriented metaphor a natural one around which to organize and express their addition, object-oriented simulations can achieve knowledge In high intelligibility, modifiability and credibility III SWIRL'S DOMAIN - ~~ [1,6,91. However, while these languages provide a potentially powerful simulation environment, they can easily be misused, since good programming style in object-oriented languages is not as well-defined as in more standard procedural languages. In this paper we describe a program called SWIRL, designed for simulations in the domain of air battles, and use SWIRL to demonstrate effective simulation programming techniques in an object-oriented language. SWIRL is written in 5: Views expressed in this paper are the authors' own and are not necessarily shared by Rand or its research sponsors. In our air-battle domain, penetrators enter an airspace with a pre-planned route and bombing mission. The goal of the defensive forces is to eliminate those penetrators. Below we list the objects that comprise this domain and briefly outline their behaviors. 1. Penetrators. These are the primary offensive objects. They are assumed to enter the defensive air space with a mission plan and route. 2. GCIs. "Ground control intercept" radars detect incoming penetrators and guide fighters to 331 From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. intercept penetrators. 3. AWACS. These are airborne radars that also detect and guide. 4. SAMs. Surface-to-air missile installations have radar capabilities and fire missiles at invading penetrators. 5. Missiles. These are objects fired by SAMs. 6. Filter Centers. They serve to integrate and interpret radar reports; they send their conclusions to command centers. 7. Fighter Bases. Bases are alerted by filter centers and send fighters out to intercept penetrators when requested to by command centers. 8. Fighters. Fighters receive messages from their base about their target penetrator. They are guided to the penetrator by a radar that is tracking the penetrator. 9. Command Centers. These represent the top level in the command-and-control hierarchy. Command centers receive processed input about penetrators from filter centers and make decisions about which resource (fighter base) should be allocated to deal with a penetrator. 10. Target. Targets are the objects penetrators intend to bomb. Figure 1 shows an example snapshot of an air-battle simulation. A complete description of the SWIRL domain can be found in [7]. IV THE DESIGN OF SWIRL - -~- In this section we outline how the above flow of command and control among the different kinds of objects is modeled in ROSS. The first step in modeling in an object-oriented language such as ROSS is to decide upon the generic actors and their behaviors. A generic object or actor in ROSS represents an object type or class and includes the attributes and behaviors of all instances of that class. For example, the generic object FIGHTER represents each of the individual fighters that may be present in any simulation environment. Second, one may need to design a set of auxiliary actors to take care of modeling any important phenomena that are unaccounted for by the generic objects. A The Basic Objects 2 -~ We begin by defining one ROSS generic object for each of the kinds of real-world objects mentioned in the previous section. We call these objects basic objects. Each of these has several different attributes representing the structural knowledge associated with that type of object. For example, to express our structural knowledge of penetrators in ROSS, we create a generic object Figure 1. Graphical Snapshot of SWIRL Simulation. 332 called PENETRATOR and define its the following ROSS command: attributes using times. (The entire hierarchical SWIRL is given in [7].) organization for (ask MOVING-OBJECT create generic PENETRATOR with Each object type in the class hierarchy can be position 1 a position' construed as a description or view of the objects max-speed 'a maximum speed' below it. One object (AWACS) happens to inherit speed 'current speed' its behaviors along more than one branch of the bombs 'current number of bombs' hierarchy (via RADAR and MOVING-OBJECT). Such status 'a status' 'multiple views' Or 'multiple-inheritance' is flight-plan 'a flight plan') possible in ROSS but not in most other object-oriented programming environments. where phrases in single-quotes represent variables. C. Modeling Non-Intentional Events To capture the behaviors of each kind of - real-world object, we begin by asking what The basic objects and their behaviors have a different kinds of input messages each of these clear correspondence to real-world objects and real-world objects could receive. For example, a their responses to the deliberate actions of fighter can receive a message (a) from its fighter others. These actions comprise most of the base telling it to chase a penetrator under significant events that we wish to simulate. guidance from a radar, (b) from that radar telling However, there are several important kinds of it to vector to a projected intercept point with events that represent side effects of deliberate the penetrator, or (c) an 'in-range' message actions (e.g., a penetrator appearing as a blip on informing it that the penetrator is in its radar a radar screen is a side effect of the penetrator range. Each of these messages then becomes the flying its course and entering a radar range). basis for a fighter behavior written in ROSS. To Such events are important since they may trigger determine the structure of each of these behaviors other actions (e.g., a radar detecting a penetrator we ask what messages the object transmits in and notifying a filter center). However, these response to each of its inputs. For example, in non-intentional events do not correspond to response to a 'chase' message from its fighter real-world message transmissions (e-s., base, a fighter will send a message to itself to penetrator does not notify a radar that it ha: take off, and then send a message to the specified entered the radar's range). An important issue in radar requesting guidance to the penetrator. The the development of SWIRL has been how to capture following ROSS command captures this behavior: these non-intentional events in an object-oriented framework (i.e., via message transmissions). (ask FIGHTER when receiving (chase 'penetrator guided by >gci) One method of capturing non-intentional events (-you unplan all (land)) could be to refine the grain of simulation. The (-you set your status to scrambled) grain of a simulation is determined by the kind of (if (-you are on the ground) real-world object one chooses to represent as a then (-you take off)) ROSS object. A division of the air-battle domain (-requiring (-your guide-time) tell -the gci into objects like penetrators and radars is guide -yourself to -the penetrator)). relatively coarse grained; a finer grain is possible. In particular, one could choose to (The '-'s signal abbreviations. The ROSS create objects that represent small parts of the abbreviations package t81 enables the user to introduce English-like airspace through which penetrators fly. Then, as expressions into his programs and to tailor the expressions to his own penetrators move they would send messages to those preference. This approach towards readability is sectors that they were entering or leaving (just as 'send particularly flexible, since the user is objects moving through real space impact or not restricted to system-defined messages' to that space). Sectors could be any English associated with radars whose interface.) ranges they define, and act as intermediary objects to notify radars B. Organizing Objects Into a Hierarchy when penetrators enter their ranges. Essentially - - - this solution proposes modeling the situation at an almost 'molecule-strikes-molecule' level of detail The behaviors of basic objects often have many that are revealed in the process of since, by adopting this level, one can achieve a commonalities strict mechanical cause-and-effect chain that is defining their 'behaviors. For example, GCIs, AWACS and SAMs all share the ability to detect, and their simple to model via message transmissions between real objects. detection behaviors are identical. We can take advantage of ROSS's inheritance hierarchy (see [8]) However, although this method solves one to reorganize object behaviors in a way that both it causes two others that make it emphasizes these conceptual similarities and modeling problem, intractable. First, the method entails a eliminates redundant code. For example, for prohibitive amount of computation. Second, in most objects that have the behaviors of detection in cases, the extra detail would make abstract generic object modeling very common, we define a more called RADAR to store these common behaviors. We awkward and unnatural (at least for our purposes in building and using SWIRL). The natural level of then place it above GCI, AWACS and SAM in the decomposition is that of 'coarse objects' such as hierarchy, so that they automatically inherit the penetrator and fighter. To the extent we stray behavior for detection whenever necessary. Hence from this, we make the simulation writer's job more we avoid writing these behaviors separately three difficult since he can no longer conceive of the 333 task in the way that comes simplest or most y PERFORMANCE FIGURES The ROSS interpreter is written in MACLISP and runs under the TOPS-20 operating system on a DEC20 (KLlO). The space requirement for this interpreter is about 112K 36-bit words. SWIRL currently contains the basic and auxiliary objects mentioned above, along with approximately 175 behaviors. Compiled SWIRL code uses about 48K words. A typical SWIRL simulation environment contains well over 100 objects and the file defining these objects uses about 3K words. Total CPU usage for the simulation of an air-battle about three hours long is about 95 seconds. This includes the time needed to drive the graphics interface. VI CONCLUSIONS - We have found the object-oriented environment afforded by ROSS to be a rich and powerful medium in which to develop a non-trivial simulation program in the domain of air-battles. The program adheres to the criteria of intelligibility, modifiability and credibility laid out in [1,6,9]. Liberal use of the abbreviations package has led to highly English-like code, almost self-documenting in nature. BY adhering to several stylistic principles for object-oriented programming, such as the Appropriate Knowledge Principle and the Appropriate Decomposition Principle, we have been able to further enhance SWIRL's modifiability. This has enabled us to easily experiment with a wide range of air-battle strategies. Coupled with a graphics interface which allows us to quickly trace a simulation run and test the credibility of its behaviors, SWIRL provides a powerful environment in which to develop and debug military strategies and tactics. naturally to him: In summary, we reject this technique because it violates the following principle that we have found object-oriented simulators: useful in designing THE APPROPRIATE DECOMPOSITION PRINCIPLE: Select a level of object decomposition that is 'natural' and at a level of detail commensurate with the goals and purposes of the model. A second solution for modeling non-intentional events would be to allow the basic objects themselves to transmit the appropriate messages. For example, if we allow a penetrator (with a fixed route) to see the position and ranges of all radars, it could compute when it would enter those ranges and send the appropriate 'in-range' messages to the radars. This solution is computationally tractable. However, it has the important drawback that it allows the penetrator to access pieces of knowledge that, in reality, it cannot access. Penetrators in the real world know their routes but they may not know the location of all enemy radars. Even if they did, they do not send messages to radars telling the radars about themselves: In short, we reiect this technique because it violates another useful principle that can be formulated as follows: THE APPROPRIATE KNOWLEDGE PRINCIPLE: Try to embed in your objects only legitimate knowledge, i.e., knowledge that can be directly accessed by being modeled. the real-world objects that are D L Auxiliary Objects We feel that the above principles should be considered by anyone attempting to develop an object-oriented simulation, as they are critical to insure readable and conceptually clear code. The solution we offer in SWIRL represents one technique that adheres to both principles. After we decompose the domain into a set of basic objects, we create auxiliary objects to handle non-intentional events. Auxiliary objects are full objects in the ROSS sense. However, they do not have real-world correlates. Nevertheless, such objects provide a useful device for handling certain computations that cannot be naturally delegated to real-world objects. We have included two auxiliary objects in SWIRL, the SCHEDULER and the PHYSICIST. The SCHEDULER represents an omniscient, god-like being which, given current information, anticipates non-intentional events in the future and informs the appropriate objects as to their occurrence. The PHYSICIST models non-intentional events involving physical phenomena such as bomb explosions and ecm (electronic counter measures). Although we have now introduced objects which have no real-world correlates, the objects that do have real-world correlates adhere to the above principles. Hence, code for the basic objects remains realistic and transparent. REFERENCES Faught, W. S., Klahr, P. and Martins, G. R. 'An Artificial Intelligence Approach To Large- Scale Simulation." In Proc. 1980 Summer ~ - Computer Simulation Conference, Seattle, 1980, 231-235. Goldberg, A. and Kay, A. "Smalltalk- Instruction Manual.' SSL 76-6, Xerox PARC, Palo Alto, 1976. Goldin, S. E. and Klahr, P. "Learning and Abstraction in Simulation." In Proc. IJCAI-81, Vancouver, 1981, 212-214. Hewitt, C. "Viewing Control Structures as Patterns of Message Passing." Artificial Intelligence 8 (1977), 323-364. Kahn, K. M. "Director Guide." AI Memo 482B, MIT, 1979. Klahr, P. and Faught, W. S. "Knowledge-Based Simulation." Proc. AAAI-80, Palo Alto, 1980, 181-183. Klahr, P., McArthur, D., Narain, S. and Best, E. "SWIRL: Simulating Warfare in the ROSS Language." The Rand Corporation, 1982. McArthur, D. and Klahr, P. "The ROSS Language Manual." N-1854-AF, The Rand Corporation, Santa Monica, 1982. McArthur, D. and Sowizral, H. "An Object- Oriented Language for Constructing Simulations." Proc. IJCAI-81, Vancouver: 1981, 809-814. [II [21 [31 [41 [51 [61 [71 [81 [91 334

1982

1

96

SPEX: A Second-Gcncration Espcrimcnt Iksifg System Yumi Iwasaki and Pctcr Friedland IIcuristic Programming Project, Computer Science Department Stanford University, Stanford, Ca. 94305 Abstract ‘I’hc design of laboratory cxperimcnts is a complex and important scientific task. The MOI,GEN project has been dcvcloping computer jystems for automating Ihc design process in the domain of molecular biology. SPEX is a scLonl-.gcncration syslcm which synlhcsizcs the best ideas of two previous MOI,CEN hierarchical planning systems: stcpwisc rcfincmcnt of skeletal plans ant1 a layered cotlirot structure. It has been tcstcd successfully on several problems in the task domain and promises to serve as a testbcd for future work in explanation, exp;rimeat debugging, and ctnpirical evaluation of different basic design strategies. 1. !nt reduction Expcrimcnt design is the process of choosing an ordered set of laboratory operations to accomplish some given analytical or synthetic goal. This process is one of the fundamental tasks of cxpcrimcntal scientists; it involves large amounts of domain expcrtisc and specialized design heuristics. ‘I’hc design of such expcrimcnt plans has been one of the fundamental research efforts of the MOIGEN project at Stanford. SIJi<X (Skclctal Planner of EXpcrimcnts) is a second-generation experiment design system. It is a synthesis of the best ideas of two previous planning systems, and will serve as a “laboratory” for the empirical testing of design strategies at many levels. SPEX will also be used for MOI-GEN work on experiment verification, optimijlation, and debugging. This paper is a report of the work in progress on SPEX. 1.1. Previous MOLGEN Planning Systems Fricdland developed an experiment planning system using the methodology of hierarchical planning by progressive refmemcnt of skeletal plans [l] [2]. A skeletal plan is a linear scqlrcnce of several abstract steps; actual plans arc gcncratcd by refining each of the abstract steps to use specific tcchniqucs and objects by going down a general-to-specific hierarchy of laboratory operations stored within a knowlcdgc base built by cxpcrimcnt molecular biologists. F’ricdland’s experiment planner chooses a skeletal-plan suitable for the given goal of an cxpcriment and rcfincs each step by choosing the best spccialilation of the laboratory method at each lcvcl of abstraction. Stcfik developed another cxpcrimeht design system [3]. His hierarchical planner first constructs an abstract plan by simple clifl‘crcnce reduction, and then gcncratcs a specific plan from that abstract plan by propagating constraints. It has a multi-layerec! control structure to separate out diffcrcnt lcvcls of decisions to bc made by the planner [4]. The two systems were complcmcntary. Fricdland’s system made cfficicnt USC of large amounts of domain knowledge to produce practical, but not necessarily optimal cxpcrimcnt designs for a large subset ofanalytical tasks in molecular biology. The assumption of near- indcpcndcnce of abstract plan-steps worked well in the great majority of casts. Stcfik’s system took much longer to plan reasonable experiments, but worked bcttcr when plan-steps were highly dcpcndcnt and kept a much richer description of the planning process, this bccausc of the well-designed control structure. Sl+TX was dcvclopcd to synthcsi;lc two fundamental ideas from these planners, namely Fricdtand’s skeletal-plan rcfinctncnt and Stcfik’s multi- layered control structure, in the hope of making further progress in the construction of a design system that would be used by experts. In addition, SPl’X has a grsatly cnhanccd capacity to simulate the changing world st;ltc during an cxpcriment. ‘I‘hc rctnaindcr of this paper describes the laycrcd control structure and the simulation mechanism used by SPEX. I,ikc Friedland’s and Stefik’s systems, the knowledge base ofSPl<X is constructed using the Unit System [5], a frame-based knowledge rcprcscntation systctn dcvelopcd in the MOi,Gl1N project. In SEX, the Unit System is also used to rcprcscnt a tract of the planning process and the changing states of objects in the world. 2. Method 2.1. Layers of Control In order to leave a trace of a planning process, it is necessary to identify the different kinds of operations the planner is expected to perform and to represent the entire process as a scqucncc of such opcrntions and their conscqucnccs. The notion of a multi-layered control structure was introduced and operations at three diffcrcnt tcvcts in lhc planning process were idcntificd and rcprcscntcd within SPBX. The bottom level, called the Iloruain Space by Stcfik, consists of the objects and operators in the task domain, termed lob-s!eps. They are experiment goals, skctctal-plans, and laboratory tcchniqucs. On top of the IJotnain Space exists the De5ig~2 Space, which consists of the vzt-ious planning tasks pcrfortncd by SPEX, for example, the tasks of finding an appropriate skclctal-plan, expanding a skeletal-plan, or refining a technique. Thcsc arc termed plan-sfeps. When such tasks arc cxccutcd, they crcatc or dctete lab-steps in the Domain Space, or create new tasks in the I)csign Space. I:inally, the third layer. the S/rcrfeu Space, consists of scvcral different strntcgics to control the execution of tasks in the Design Space. Diffcrcnt types of decisions are made in the three different spaces. In the Domain Space, decisions are biology-oricntcd. The two major types of decisions are environmental, i.c. whcthcr environmental conditions and structural propcrtics allow a given laboratory technique to be used, and detailed selection, i.e. tl:c process of deciding on the From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. basis of selection heuristics among several tcchniqucs all of which will carry out the specified experimental goal. In the I>csign Space, decisions are more goal-oricntcd. l’hesc decisions rclatc the specific goal of a step in the skeletal plan to the potential utility of a laboratory technique for satisfying that goal. Finally, in the Strategy Space, choices arc made among various design strntcgics, whcthcr to rcfinc in a breadth-first, depth-first, or heuristic manner, for cxamplc. I,ab-steps in the Domain Space arc also rcprcsentcd by units. I.ah- step units initially record the 11orrtain Space results of decisions made in the I>csign Space, but may also record later decisions made in the Domain space. Some of the slots in the prototypical lab-step unit arc shown in l;igurc 3-2. Unit: LAD-STEP 2.2. Modeling World Objects FOLLOWING-STEP: <UNIT> A pointer IO the next iubstep in the plan. ----- Ncithcr previous MOI,GEN design system did a thorough job of monitoring the state of the laboratory cnvironmcnt during the simulated execution of an cxperimcnt plan. nut, the laboratory environment, i.e. the physical conditions and the molecular structures involved in the cxpcrimcnt, undergoes changes in the COUISC of carrying out the experiment. Therefore, for the planning of an experiment design to bc consistently successful, it is csscntial to predict the changes caused by carrying out a step in the plan. SEX simulates those changes, using the part of its knowlcdgc base that contzirls the yotcntial effects of carrying out an laboratory tcchniquc. ‘I’hc prcdictcd states of the world at certain points in the cxpcrimcnt arc used as part of the selection criteria in choosing the appropriate technique for the next step in the plan. ‘I’hcy arc also used to make sure that the preconditions for a chosen technique arc met before the clpplication of the tcchniquc. If any of the preconditions arc not satisfied, a sub-plan to modify the world state is produced and inscrtcd in the original plan. PRECEDING-STEP: <UNIT> A pointer IO the prcccd~ng lab-step in the plan. ----- CREATED-BY: <UNIT> A pointer to 11x plan-step which created thiy lab-step. ----- STATUS : <Sl RING> When a lubstep has just been created, it b given the status ‘jurt-created”, I ater when it is ndcd in cr out, the status is chnngcd to “ruled-In” or “ndcd-out”. 14gurc 3-2: Slots in a lab-step SPEX uses an agenda mechanism to keep track of all the pending tasks. The Strategy Space of SPEX c:msists of simple strntcgics about how to choose the task from the agenda. Currently, a slratcgy is chosen at the beginning of a session by the user and it is used throughout the session. ‘I’here arc three strategies currently available in the system. With Strategy 1. the agenda is used like a qucuc and tasks arc fctchcd in first-in, firsL-out manner. With Strategy 2, the agenda is n~mnally used like a stack and tasks arc fctchcd in first-in, List-out fashion. With 3. Implementation Strategy 3, the plan-steps arc given priorities according to their types. SPEX is implcmcnted in Interlisp, making extensive USC of the IJnit System in order to rcprescnt each operation in diffcrcnt planning spaces. In the Design Space, thcrc arc at present scvcn types of plan- Tasks of a higher priority are executed before any tasks of lower priorities. We are currently cxpcrimenting with various prioritizing schcmcs for the different plan-step types. steps: obtaining the goal of an cxpcrimcnt, choosing a skclctnl plan, expanding a skclctal plan, refining a laboratory tcchniquc, ruling in or out a laboratory tcchniquc, comparing altcrnativc laboratory techniques, and checking the world state to verify that the preconditions for application of a tcchniquc arc satisfied. When a new task needs to hc gencratcd, a plan-step unit is created to represent the task and to 1c;rvc a record of SPEX’s pcrformancc. A plan-step unit has slots containing such information as when the task was created, when it was pcrformcd and what the consequence was. Some of the slots in the prototypical plan-step unit arc shown in Figure 3-l. this plan-step NtXT-PLAN-STEP: <UNIT> 7%u next plan-step J&hed ----- LAST-PLAN-STEP: <UNIT> The previous plan-step fetched 3.1. Simulating the World State At the bcgilllling of ;I planning scss~on, the user is asked to provide a description of the world ol!jccls. ‘l‘his inclirclcs the current physical cnvironmcnt (tcinpcraturc, 1~1-1, etc.) of the cxpcrimcnt and what tic knows about rhc d&led molecular structure of his cxpcrimcntal objects, normally nucleic acid scqucnces. A unit is created to rcprcscnt the initial dcscriplion for each HY~II:! object. When a skclctJ-plan is cxpnndcd to individual steps, units arc crcatcd to rcprcscnt rhc simulated state of each world objjcct bcforc and Acr ,Ipplication of each step. This is done by ulili& (1 simulation information stored in the laboratory-tcchniqucs hierarchy of the knowlcdgc base. l;igurc 3-3 sl~ows some of the slots in the prototypical unit fc)r describing Iluclcic acid structures. ‘I’his unit is called l)N/\-S’l‘liUC’I‘U121~. Its composiiion h,js cvolvcd over scccrnl years of colLrborari\c wwk by scvcral of tlw molecular biologists associalcd with the IL101 .CEN project. It rcprcscnts all of the potential information a scicnlist might wish to supply about a particul‘lr nucleic acid structure. In an actual cxpcrimcnt, a scicnlist would bc able to fill only a few of the olcr 50 slots in the prototypical unit. I:igurc 3-4 shows rlic actual values gi\,cn to the sample slots during the plannin g of an cxperimcnt design by swx. STATUS: <STRING> The .status of this plan-step; either Succeeded’: “postponed, ” or ‘created”. Figure 3-1: Slots in a plan-step 342 Unit: DNA-STRUCTURES ----_ STRANDLDNESS: <STRING> One of: ["HYBRID" "SS" "DS"] ----_ LENGTH: <INTEGFR> [l lOM] MEASUREMCNT UNIT EtASE-PAIRS ----- #-TERMINI: <INTEGER> co 41 ----- #-FORKS: <INTLCER> co 21 ----_ 10POLOGY: <STRING> One of: ["DELIA-FORM" "TtIEIA-I-ORM" "EYC-FORM" "Y-FORM" "LINEAR" "CIRCULAR"] ----- TYI'E: <STRING> One of: ["HYBRID" "RNA" "DNA" ] Figure 3-3: Slots in DNA-STRUCTURE Unit: STRUC-1 STRANF'CDNESS: <STRING> "SS" LENGTH: <INfFGCR> 300 MFASUREMENT UNIT UASE-PAIRS #-TERMINI: <INFEGCR> 2 ----- #-FORKS <INTEGER> 0 ----- TOPOLOGY <STRING> "LINEAR" ----- TYPE: <STRING> "DNA" Figure 3-4: Slots in an instance of DNA-S’I‘RUC’I’URC~ Sl’l+X ha:; been successfully tostctf 01i scv2r::l of 1112 same problems uccd 2s test ca\cs in l*‘ricdlnnd’s thesis [l]. Combining the iclcas 9f skclctnl-plan refinement and multi-la: cl-cd control ~,iruclurc proved uscti~l in keeping the l)omain Space oi iicicncy of I~ri~:d!:~nd’s system, while introducing thL Planning and Stratcu Space flexibility of Stcfik’? syTtcm. ‘l’hc grenlly improved simulillion mcclianism incrcascd the reliability of the decisions made and allowed SPlIX to sl~ggcst, in detail, ways of correcting low-level incompatibilites bctwccn chosen laboratory tccllIliqlles. SI’I<X keeps a clear trace of all the decisions rnaclc during tilt cxpcrimcnt design process. This tract will now bc used for a variety of’ purposes in ongoing MOI~GIIN research. One goal of this rcscarch to add a dctailcd cxplnn&ion capability to SPEX. Iiarlier MOI.GEN cxpclimcnt design bqstcms only provided explanation by listing the domain-specific rules used to make dcc-isions at the Icvcl of laboratory tcchlliyucs. The planning trace and the modular IKILLIK of the diffcrcnt planning spaces will allow this explanation Elcility to bc cxpandcd. ‘l‘hc user will bc able to direct his questioning to domain. design, or strategic motivations. WC envision an initial cxplanntion facility silnilar to that used by ME’CIN [(,I. ‘1’11~ user asks questions using a szt of key words i\S “Why”, “WilCn”, “how”, CIC. WC believe that beside<; promoting the USC of SPEX among expert users for whom full explanations arc a nccersity, the explanation facility will also greatly Lrciiitatc debugging of the knowI~Jge base. A necessary cumplcrncnt to the task of cxpcrimcct &sign is the task of cxpcrimcnt debugging. Initial cspcrimcnt designs prod~~cccl by even the very best of scientists rarely work perfectly the first time. WC arc certain that the same will be bc true of SPEX-produced cxpcrimcnt designs. Given the scqucncc of tcchniqucs employotl in an cxpcrimcnt, a dcbuggcr will compnrc the prcdictcd states of the world bcforc and after each step and the actual world states during the cxpcrimcnt (sL:pplicd by the niolccular biologist ulcr). ‘l’hen, it will point out the steps which might possibly have gone wrong and suggest solutions: altcmstivc techniques or a rcmcdial proccdurc to be licrformcd. If the “buggy” cxpcriment &sign had come from a human scictltist, then the debugging information will enable him to correct his personal “knowledge bitsc.” In a similar manner, wc bclicvt: that the Lomparison of actual laboratory results to the previously predicted results should allow the automatic improvcmcnt of the knowlcdgc base in the cast of a Sl’I<X-gcncratcd cxpcrimcnt design. WC do not illltiCipXC trujor difficulties in building the cxpcrimcnl dcbuggcr gikcn the existing mcchanicms of SPIIX. In summary, SI’l;X rcprcscnts a synthesis of scvcr31 mcthodologics for dcqign, nnmcly skclctnl-plan rcfincmcnt and the multi-layered control structure. It is a framework for a gcn~‘~,al-pul’posc design testing and tlcbugging system, which cm bc easily tailored to do planning in any specific npplication domain. ‘I’hcrc arc no nlolcculi~r-biology specific mcch,lnisms inhcrcnt in SI’IIX; nil of th: domain-spcci~ic kn()\vlcL]gc in ill tllc :lssocinled knowlcdgc b:jsc. SI’liX C;\Il ah) h ~lsCd to test clir[cFctlt l>itSiC &sign ~ir,ltCgiC!: by LhC i~~!~)~ciIlCtltiltioll Oi‘ In~llly fiddilioll;ll stra[c$ics iti its Str,ltcsy Sp:lcC. WC bclic\c tI13t IllC lKSt way to dctcrminc the efficacy of tlic Illittly diffcl cnt potential ~li~dtcgics is cmpiric;J, and Sl+tX will bc useful as a hhratory for thCsC cxperimcntS. This work is p:lrt of the MOI~CiFN project, n joint rc$car.ch effort among Ihc I):partmcnts o!’ Computer Scicncc, Medicine, and iGochcniistry al Stanford Uuivcrsity. '1'1~ mearch has been \upported ~ntlcr NSl: gr#\rlt MC%&16247. Cornl~L1ti~tiollal rcsourccs hn~c IXCII provided by the SUMEX-AIM Nation,\] lGomcdica1 Rcscarch I<csourcc, NI I I grant RR-00785-08, and by the Dcpal tlncnt of Computer Science. WC wish to th,lnk our many cnthusinstic MO1 6 l:N collA)orlltors for their Ast,\ncc in tllis work. WC arc cspeciallq grateful to I<cnL Ihch and I-arty Kcdcs for pro\ iding tilC ~?lOlKtlli~~ biology c\pcrLisc ncccbsary to test Sl’I:X and tu f2ruic l~uch;tlli111 and Mike Gcnzscrcth for advice on the artificial intclligcllcc methotiolo~ics cmpioycd. 343 References 1. Fricdiand, P.E., Ktrowledge- Rased l:icperhen~ Desigtl In Mol~~~~!ur Gerre[ics, Phl1 dissertation, Stanford University, OctobCr 1979. 2. Fricdland, P.E., “Knowlcdgc-1jascd Expcrimcnt Design in Molccularc Genetics ” LIC’AI-754 ‘The International Joint Confcrcnce on hrtif&al Intelligence, 1979, pp. 285-287. 3. Stcfik, M.J., PIunning wilh Conslrairtts, Phi> dissertation, Stanford University, January 1980. 4. Stcfik, M.J., “Planning and MC&Planning,” HPP-mcmo HPP- 80-13, Stanford University Heuristic Programming Project, 1380. 5. Smith, KG., Friedland, P.R., “Unit Package User’s Guide,” HPP-memo :wP-80-28, Stzmford University Heuristic Programming Project, 1980. 6. Scott, AC., Clanccy, W.J., Davis, R., Shortliffc, E.H., “Explanation Capahilitics of I’roduction-Rascd Consultation System,” Anwican fournal of ~orqwilaliot~al Linguislics, 1979, .

1982

10

97

CIRCUMSCRIPTION IMPLIES PREDICATE COMPLETION (SOMETIMES) Raymond Reiter Department of Computer Science Rutgers University New Brunswick, N. J. 08903 ABSTRACT Predicate completion is an approach to closed world reasoning which assumes that the given sufficient conditions on a predicate are also necessary. Circumscription is a formal device characterizing minimal reasoning i.e. reasoning in minimal models, and is realized by an axiom schema. The basic result of this paper is that for first order theories which are Horn in a predicate P, the circumscription of P logically implies P's completion axiom. Predicate completion [Clark 1978, Kowalski 19781 is a device for "closing off" a first order representation. This concept stems from the observation that frequently a world descrip- tion provides sufficient, but not necessary, conditions on one or more of its predicates and hence is an incomplete description of that world. In reasoning about such worlds, one often appeals to a convention of common sense reasoning which sanctions the assumption - the so-called closed world assumption [Reiter 19781 - that the infor- mation given about a certain predicate is all and only the relevant information about that predicate. Clark interprets this assumption formally as the assumption that the sufficient conditions on the predicate, which are explicitly given by the world description, are also neces- sary. The idea is best illustrated by an example, so consider the following simple blocks world description: A and B are distinct blocks. A is on the table. B is on A. (1) These statements translate naturally into the following first order theory with equality, assuming the availability of general knowledge to the effect that blocks cannot be tables: BLOCK (A) BLOCK (B) 0~ (A,TABLE) ON (B,A) (2) A#B A# TABLE B # TABLE Notice that we cannot, from (2), prove that nothing is on B, i.e., (2) /+ (x> -ON(x,B), yet there is a common sense convention about the description (1) which should admit this conclu- sion. This convention holds that, roughly speaking, (1) is a description of all and only the relevant information about this world. To see how Clark understands this convention, consid- er the formulae (x) .x = A V x = B ZJ BLOCK (x) (3) (xy).x = A 6 y = TABLE V x * B d y = A 3 ON(x,y) which are equivalent, respectively, to the facts about the predicate BLOCK, and the predicate ON in (2). These can be read as "if halves", or sufficient conditions, of the predicates BLOCK and ON. Clark identifies the closed world as- sumption with the assumption that these sufficient conditions are also necessary. This assumption can be made explicit by augmenting the representa- tion (2) by the "only if halves" or necessary conditions, of BLOCK and ON: (x). BLOCK(x) 1 x=A V x=B (xy). ON (x,y) 3 x=A & y=TABLE V x=B & y=A Clark refers to these "only if" formulae as the completions of the predicates BLOCK and ON respectively. It now follows that the first order representation (1) under the closed world assumption is (x). BLOCK(x) : x=A V x=B (xy). ON(x,y) = x=A & y=TABLE V x=B & y=A AfB A # TABLE B # TABLE From this theory we can prove that nothing is on B - (x)-ON(x,B) - a fact which was not derivable from the original theory (2). Circumscription [McCarthy 19801 is a dif- ferent approach to the problem of "closing off" a first order representation. McCarthy's intu- itions about the closed world assumption are es- sentially semantic. For him, those statements derivable from a first order theory T under the closed world assumption about a predicate P are just the statements true in all models of T which are minimal with respect to P. Roughly speaking, these are models in which P's extension is minimal. McCarthy forces the consideration of only such models by augmenting T with the fol- lowing axiom schema, called the circumscription of P in T: T(t) & b> .4(x) = P(x)] = (xl. P(x) = @(xl From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. Here, if P is an n-ary predicate, then $ is an n-ary predicate parameter. T(4) is the conjunction of the formulae of T with each occur- rence of P replaced by 9. Reasoning about the theory T under the closed world assumption about P is formally identified with first order deduc- tions from the theory T together with this axiom schema. This enlarged theory, denoted by CLOSUREp(T), is called the closure of T with respect to P. Typically, the way this schema is used is to "guess" a suitable instance of 4, one which permits the derivation of something use- ful. a fact which is not derivable from the original theory T. Notice that in order to make this work, a judicious choice of the predicate parameter 4, namely (5), was required. Notice also that this choice of $ is precisely the antecedent of the "if half" (3) of ON and that, by (6), the "only if half" - the completion of ON - is derivable from the closure of T with respect to ON. For this example, circumscription is at least as powerful as predicate completion. To see how this all works in practice, con- sider the blocks world theory (2), which we shall denote by T. To close T with repect to ON, augment T with the circumscription schema X,Y> = ON(x,y)l ON(x,y) = $(x,Y) (4) Here 4 is a 2-place predicate parameter. Intuitively, this schema says that if 4 is a predicate satisfying the same axioms in T as does ON, and if 4's extension is a subset of ON's, then ON's extension is a subset of 4's, i.e., ON has the minimal extension of all predicates satisfying the same axioms as ON. In fact, this example is an instance of a large class of first order theories for which cir- cumscription implies predicate completion. Let T be a first order theory in clausal form (so that existential quantifiers have been eliminated in favour of Skolem functions, all variables are universally quantified, and each formula of T is a disjunct of literals). If P is a predicate symbol occurring in some clause of T, then T is said to be Horn in P iff every clause of T con- tains at most one positive literal in the pred- icate P. Notice that the definition allows any number of positive literals in the clauses of T so long as their predicates are distinct from P. Any such theory T may be partitioned into two disjoint sets To see how one might reason with the theory CLOSUREON( consider the following choice of the parameter + in the schema (4): Tp: those clauses of T containing exactly one positive literal in P, and T-Tp: those clauses of T containing no positive $(x,y) q x=A & y=TABLE V x=B 6 y=A (5) (but possibly negative) literals in P. Then T(4) is BLOCK(A) & BLOCK(B) & [A=A & TABLE=TABLE v A=B 6r TABLE=A] 6 CB=A 6 A=TABLE v B=B & A=AI & A# B & A# TABLE &B # TABLE so that, for this choice of $, CLOSUREON I-- T($) It is also easy to see that, for this choice of CLOSUREON j- (xy) .$(x,y> = ON(x,y) . Thus, the antecedent of (4) is provable, whence CLOSUREON t (XY> .ON(x,y) 3 $(x,y). 1. * e. Clark (1978) provides a simple effective procedure for transforming a set of clauses of the form Tp into a single, logically equivalent formula of the form (x). A(x) = P(x). The con- verse of this formula, namely (x). P(x) = A(x), Clark calls the completion axiom for the predicate P, and he argues that augmenting T with P's com- pletion axiom is the appropriate formalization of the notion of "closing off" a theory with re- spect to P. Our basic result relates this notion of closure with McCarthy's, as follows: Theorem Let T be a first order theory in clausal form, Horn in the predicate P.Let (x).P(x) 3 A(x) be P's completion axiom. Then CLOSUREp(T) I- (x). P(x) 1 A(x) CLOSURE (T) I- (xy).ON(x,y) ' x=A & y%BLE V x=B & y=A (6) i.e. the only instances of ON are (A,TABLE) and @,A ) l It is now a simple matter to show that nothing is on B, i.e. CLOSUREON t (x).-ON(x,B) i.e. P's completion axiom is derivable by cir- cumscription. Discussion Circumscription and predicate completion are two seemingly different approaches to the formal- ization of certain forms of common sense reason- ing, a problem which has recently become of major 419 concern in Artificial Intelligence (see e.g. [AI 19801). That circumscription subsumes predicate completion for a wide class of first order theories is thus of some theoretical interest. Moreover, circumscription is a new formalism, one whose properties are little understood. Predicate completion on the other hand, has a solid intuitive foundation, namely, assume that the given sufficient conditions on a predicate are also necessary. The fact that predicate com- pletion is at least sometimes implied by cir- cumscription lends support to the hypothesis that circumscription is an appropriate formalization of the notion of closing off a first order represen- tation. Finally, the theorem has computational im- port. Notice that in order to reason with McCarthy's circumscription schema it is first necessary to determine a suitable instance of the predicate parameter Cp. This is the central computational problem with circumscription. With- out a mechanism for determining "good $'s", one cannot feasibly use the circumscription schema in reasoning about the closure of a represen- tation. This problem of determining useful 4's is very like that of determining suitable pre- dicates on which to perform induction in, say, number theory. Number theory provides an induc- tion axiom schema, but no rules for instantiating this schema in order to derive interesting theorems. In this respect, the circumscription schema acts like an induction schema. Now the above theorem provides a useful heuristic for computing with the closure of first order Horn theories. For we know a priori, with- out having to guess a $ at all, at least one non trivial consequence of the circumscription schema, namely the completion axiom. Clearly, one should first try reasoning with this axiom before invok- ing the full power of circumscription by "gues- sing $‘s". REFERENCES AI (1980). Special issue on non-monotonic logic, Artificial Intelligence 13 (1,2), April. Clark, K. L. (1978). Negation as failure, in Logic and Data Bases, H. Gallaire and J. Minker (eds.), Plenum Press, NY, 293-322. Kowalski, R. (1978). Logic for data description, in Logic and Data Bases, H. Gallaire and J. Minker (eds.), Plenum Press, NY, 77-103. McCarthy, J. (1980). Circumscription - a form of non-monotonic reasoning, Artificial Intelligence 13, 27-39. Reiter, R. (1978). On closed world data bases, in Logic and Data Bases, H. Gallaire and J. Minker (eds.), Plenum Press, NY,55-76. 420

1982

100

98

Monitors as Responses to Questions: Determining Competence Eric Mays Department of Computer and Information Science Moore School of Electrical Engineering/D2 University of Pennsylvania Philadelphia, Pa. 19104 ABSTRACT This paper discusses the application of a propositional temporal logic to determining the competence of a monitor offer as an extended response by a question-answering system. Determining monitor competence involves reasoning about the possibility of some future state given a description of the current state and possible transitions. I INTRODUCTION The offer of a monitor as a response becomes possible when the system views the knowledge base (KB) as a dynamic rather than a static entity. That is, in addition to answering a user's question on the basis of the information the system currently contains, a system with a dynamic view can offer to monitor for additional relevant information which it will provide to the user if and when it learns of it. Such additional information could be about some possible future event or some previous event about which the system's knowledge is currently incomplete. In the following question-answer pairs, Q-Al illustrates a monitor for some possible future event. The pair Q-A2 is an example of a monitor for some additional information about a previous event. Responses such as Q-A2 require reasoning that some event, of which knowledge regarding its outcome would enable us to answer the question, has taken place. At some point in the future we will learn of its outcome, when we will then answer. Ql: Al: 42: A2: Has John registered for CSEllO? No, shall I let you know if he does? Did John pass CSEllO? I don't know yet. The semester has ended, but Prof. Tardy has not turned in his grades. Shall I let you know when I find out? In order to offer monitors as extended responses the system must be competent to offer to monitor for only those events which might possibly occur or, if the system has incomplete knowledge * This work i s partially supported by a grant from the Na tional Science foundation, NSF-MCS 81-O 7290. of some event that has occured, only that additional information it may learn of.* This requires some notion of what events are possible or what additional information may be acquired given the current state of the knowledge base. For example, ignorance of the stages through which undergraduates proceed in the university would leave a system attempting to offer monitors unable to discriminate between the following two cases. Ql: Is John a sophomore? Al: No, he's a freshman. Shall I let you know when he becomes a sophomore? Q2: Is Mary a sophomore? A2: No, she's a junior. Shall I let you know when she becomes a sophomore? The remainder of this paper is concerned with determining monitor competence with regard to possible future events. We leave open for now the question of competence for those monitors that require reasoning about incomplete knowledge of some previous event. II REPRESENTATION Temporal logic [3] is a modal logic for reasoning about the relative possibility of some state to some other state, where the relative possibility is with respect to time or sequences of events. (In contrast to, for example, relative possibility with respect to obligation or belief.) Although one might develop a suitable first order theory to deal with the problems discussed here, it seems worthwhile to study this problem within the framework of temporal logic for reasons of conceptual clarity. Restriction to the propositional setting enables us to concentrate on those issues involved with reasoning about possible change. We model the evolution of a KB in a propositional temporal logic. The future fragment is a unified branching temporal logic [l] which * In either case it must be able to identify those future conditions which are relevant to the user's intentions. The discussion here will be limited to determination of monitor competence. for a brief discussion on relevance. See [3] 421 From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. makes it possible to describe properties on some or all futures. By merging the existential operators with the universal operators, a linear temporal logic is formed for the past fragment (i.e. AXp <-> EXp). A. Syntax Formulas are composed from the symbols, - A set P of atomic propositions. - Boolean connectives: v, -. - Temporal operators: AX (every next), EX (some next), AG (every always), EG (some always), AF (every eventually), EF (some eventually), L (immediately past), P (sometime past), H (always past). using the rules, - If p e P, then p is a formula - If p and q are formulas, then (-p), (p v q) are formulas. - If m is a temporal operator and p is a formula, then (m)p is a formula. Parenthesis will occasionally be omitted, and &, ->, <-> used as abbreviations. B. Semantics A structure T is a triple (S, TT, R) where, - S is a set of states. - TT:(S -> 2-P) is an assignment of atomic propositi0t-Z to states. - R C (S x S) is an accessibility relation on s.- Each state is required to have at least one successor and exactly one predecessor, As (Et (sRt) & E!t (tRs)). Define an s-branch b = (..., s(-1), s=s(O), s(l), . ..) such that s(i)Rs(i+l). The satisfaction of a formula p at a node s in a structure T, <T, s> I= p, is defined as follows: (Note: "err denotes set inclusion.) <T, s> I= p <T, s> I= -p <T, s> I= p v <T, s> I= AGp (p is true at <T, s> I= AFp (p is true at <T, s> I= AXp (p is true at <T, s> I= EGp (P is true at <T, s> I= EFp (p is true at <T, s> I= EXp (p is true at <T, s> I= Hp (p is true at every time of the past) iff peTT(s), for p a proposition iff not <T,s>l=p q iff <T,s>l=p or <T,s>]=q iff Ab At ((teb & t>s) -> <T,t>l=pT every time of every future) iff Ab Et (teb & t>s & <T,t>l=p) some time of every future) iff At (sRt -> <T,t>l=p) every immediate future) iff Eb At ((teb & t>s) -> <T,t>l=p) every time of some future) iff Eb Et (teb & tks & <T,t>l=p) some time of some future) iff Et (sRt & <T,t>l=p) some immediate future) iff Ab At ((teb 6 t(s) -> <T,t>l=p) <T, s> I= Pp iff Ab Et (teb & t<s & <T,t>l=p) (p is true at some time of the past) <T, s> I= Lp iff Et (tRs & <T,t>l=p) (p is true at the immediate past) A formula p is valid if for every structure T and every node s in T, <T, s> I= p. C. Axioms In the following axioms Dl-3, Al-4, El-4 and rules Rl-3 form a complete deductive system for the future fragment [l]. Similarly, D4-5, Pl-4, Rl, R2, R4 are complete for the past fragment. The idea of Ul and U2 is that the relationship between the past and the future may be described locally. Using the induction axioms, we can derive the following theorems, which are a more conventional form (as in 131): EF(Hp) -> Hp P(AGp) -> AGp See [1] for a list of many other useful theorems for the future fragment. Dl) AFp <-> -EG-p D2) EFp <-> -AG-p D3) AXp <-> -EX-p D4) Pp <-> -H-p DS) Lp <-> -L-p Al) AG(p -> q) -> (AGp -> AGq) A2) AX(p -> q) -> (AXp -> AXq) A3) AGp -> p & AXp & AX(AGp) A4) AG(p -> AXp) -> (p -> AGp) El) AG(p -> q) -> (EGp -> EGq) E2) EGp -> p & EXp & EX(EGp) E3) AGp -> EGp E4) AG(p -> EXp) -> (p -> EGp) Pl) H(p -> q) -> (Hp -> Hq) P2) L(p -> q> -> (Lp -> Lq) P3) Hp -> p & Lp & L(Hp) P4) H(p -> Lp) -> (p -> Ilp) Ul) L(AXp) -> p U2) p -> WLP) Rl) If p is a tautology, then I- p. R2) If I- p and I- (p -> q), then I- q. R3) If I- p, then I- AGp. R4) If I- p, then I- Up III EXAMPLE Consider as an example representing that portion of a university KB dealing with students passing and registering for courses. Let the propositional variables Q and R mean "student has passed course" and "student is registered for course", respectively. One might have the following non-logical axioms: 1) (AG)(Q -> (AX)Q) - once course it remains so a student has passed a 422 2) (AG)((-Q & -R) -> (EX)R) - if a student has not passed a course and is not registered then it is next possible that s/he is registered 3) (AG)(R -> (EX)Q) - if a student is registered for a course then it is next possible that s/he has passed 4) (AG)(R -> (EX)(-Q & -R)) - if a student is registered for a course then it is next possible that s/he has not passed and is not registered 5) (AG)(Q -> -R) - if a student has passed a course s/he is not registered for it 6) (AG)(R -> -Q> - if a student is registered for a course s/he has not passed it (equivalent to 5) Given the following question, Is John registered for CSEllO?, there are three possibilities depending on the present state of the KB: 1) John is not registered (-R), but he has passed (Q>* If we consider John registering for CSEllO as a possible monitor, it would be ruled out on the basis that it is provable that John cannot register for CSEllO. Specifically, from Q and axioms 1 and 5, it is provable that -(EF)R. It would thereFore be incompetent to offer to monitor for that condition. 2) John is not registered (-R), but he has not passed (-Q). In this case we could offer to monitor for John registering for CSEllO, since (EF)R is derivable from axiom 2. 3) John is registered for (R), hence he has not passed (-Q). One could competently offer to monitor for any of the following: a) John no longer registered for CSEllO; (EF)-R b) John passed CSEllO; (EF)Q c) John registered for CSEllO again; (EF)(-R & (EX)R) This last case is interesting in that it can be viewed as a monitor for -R whose action is to set a monitor for R (whose aciton is to inform the user of R). Also, one may wish to include monitors that are responsible for determining whether or not some monitor that is set can still be satisfied. That is, just because something was possible once does not imply that it will always be possible. The user should probably be informed when a situation s/he may still be expecting (because a monitor was offered) can no longer occur. For example, if the system has offered to inform the user if John registers for CSEllO, then s/he should be informed if John receives advance placement credit and can no longer register. The following set of axioms illustrate the use of the past operators. Note that axiom 1 from the above set may be eliminated, due to the ability of the past operators to "look back". 1) (AG)((-(P)Q & -RI -> (EX)R) 2) (AG)(R -> OWQ) 3) (AG)(R -> (EX)(-Q 6( -R)) 4) (AG)((P)Q -> -RI 5) (AG)(R -> -(P)Q) A more important use of the past operators the ability to express conditions that depend on sequences of events. Thus, expressing the condition that in order for a student to registe for a course s/he must not have registered for i twice before (say, because s/he dropped out or failed), requires a formula of the following for (AG)(-(P)(R & (L)(P)R) & -(P)Q & -R -> (EX)R) is r t m: IV CONCLUSION A simple theorem prover based on the tableau method has been implemented for the propositional branching time temporal logic as described in [l]. Current investigations are aimed towards formulating a quantified temporal logic, as well as the complicated issues involved in increasing efficiency of making deductions. This effort is part of a larger, more general attempt to provide extended responses in question-answering systems (41. ,4 final comment as to general structure of this enterprise. One could conceivably develop a suitable first order theory to deal with the problems discussed here. It seems worthwhile, however, to study this problem within the framework of temporal logic for reasons of conceptual clarity. Restriction to the propositional setting enables us to deal strictly with those issues involved with reasoning about possible change. Also, we may be able to gain some insight into a reasonable decision procedure. ACKNOWLEDGEMENT Scott Weinstein, Aravind Joshi, Bonnie Webber, Sitaram Lanka, and Kathy McCoy have provided valuable discussion and/or comments. REFERENCES [l] M. Ben-Ari, Z. Manna, and A. Pneuli, "The Temporal Logic of Branching Time", Eighth Annual ACM Symposium on Principles of Programming Languages, Williamsburg, Va., January 1981. [2] E. Mays, A. Joshi, and B. Webber, "Taking the Initiative in Natural Language Data Base Interactions: Monitoring as Response", Proceedings of ECAI 82, Orsay, France, July 1982. [3] N. Rescher and A. Urquhart, Temporal Logic, Springer-Verlag, New York, 1971. [4] B. Webber, A. Joshi, E. Mays, and K. McKeown, "Extended Natural Language Data Base Interactions", to appear.

1982

101

99

A SYSTEMATIC APPROACH TO CONTINUOUS GRAPH LABELING WITH APPLICATION TO COMPUTER VISION* M. D. Diamond, N. Narasimhamurthi, and S. Ganapathy Department of Electrical and Computer Engineering University of Michigan Ann Arbor, MI 48109 ABSTRACT The discrete and continuous graph labeling problem are discussed. A basis for the continuous graph labeling problem is presented, in which an explicit connection between the discrete and continuous problems is made. The need for this basis is argued by noting conditions which must be satisfied before solutions can be pursued in a formal manner. Several cooperative solution algo- rithms based on the proposed formulation and results of the application of these algorithms to the problem of extracting line drawings are presented. I XHECIONTINlrnrlrGRAPHLAREXWGJzQRLEM A graph labeling problem is one in which a unique label, A from a set A of possible labels must be assigned to each vertex of a graph G = (V,E). The assignment must be performed given information about the relation- ship between labels on adjacent vertices and incomplete local information about the correct label at each vertex. In a discrete graph labeling problem [ 1,2,3], the local information consists of a subset, & s A, of the label set associated with vertex vi E V, from which the correct label for each vertex must be chosen. The contextual information consists of binary relations Ru s Axh, referred to as constraint relations, assigned to each edge vivj E E. The function of the constraint relations is to make explicit which labels can co-occur on adjacent vertices. The graph, label set, and constraint relations together form a constraint network [2,5]. An (unambi- guous) labeling is a mapping which assigns a unique label h E A to each vertex of the graph. A labeling is con- sistent if none of the constraint relations is violated, that is, if label h is assigned to vertex Vi and label h’ is assigned to vertex vj then the pair (h,X’) is in the con- straint relation Rii for the edge ViVj E E. Given initial labeling information, several search techniques have been developed which can be used to derive consistent labelings. The original backtracking search described by Waltz [ 11 was later implemented in parallel by Rosenfeld et al. [6], resulting in the discrete relaxation operator. At the same time a continuous analogue, the continuous graph labeling problem was * This work was supported in part by the Robotics Research Laboratory, and in part by the Ultrasonics Im- aging Laboratory both in the Department of Electrical and Computer Engineering, University of Michigan. proposed, as well as a continuous relaxation algorithm for its solution, and since then several other relaxation algorithms have been proposed [7,8]. In a continuous graph labeling problem, the initial information consists of strength measures or figures of merit, pi (Aj), given for each label Aj E A on each vertex Vi E I! The strength measures are assumed generated by feature detectors which are making observations in the presence of noise. They usually take onvalues in the range [O,l], a 0 indicating no response, and a 1 indicat- ing a strong response. The contextual information, which is represented in terms of constraint relations for the discrete graph labeling problem, are replace by measures of compatibility, usually taking values in the range [-1,1] or [O,l], which serve to indicate how likely the pairs of labels are to co-occur on adjacent vertices. Several problems have resulted in the extension of the graph labeling problem from the discrete to continu- ous case. In the discrete case the presence or absence of a pair in a constraint relation can be determined with certainty depending on what labelings are to be con- sidered consistent. In the continuous case, however, there is apparently no formal means to assign specific numeric values to the compatibility coefficients, partic- ularly for shades of compatibility between “impossible” and “very likely”, although several heuristic techniques have been proposed [7,9,10]. Furthermore, with respect to a constraint network, the concept of consistency is well defined, The objective the continuous relaxation labeling processes has often been stated to be that of improving consistency, however, the definition for con- sistency has not been given explicitly. This latter issue is circumvented in several of the optimization approaches which have been proposed [11,12,13], where the an objective function, defined in terms of the compa- tibility coefficients and the initial strength measures is given. However, because of the dependence of the objec- tive functions on the compatibility coefficients, and because no real understanding of the role which these coefficients play yet exists, it is often difficult to describe the significance of these approaches in terms of what is being achieved in solving the problem. In an alternate approach to the continuous graph labeling problem [14] an attempt has been made to maintain the characteristics of the original problem while allowing for more systematic approaches toward a 50 From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. solution. It is felt that solutions to the reformulated problem will be more useful because it will be easier to relate the results of the solution algorithm to what is being achieved in the problem domain. In order to develop this approach, we review the characteristics of the solutions to the graph labeling problem which have appeared so far (refer to Fig. 1). The inputs to the process are the initial strength measures ipi'( i=l,s..,W 0 j=l ,...,mj which can be represented by an 7~ x’I)?. dimensional vector: F = (p 10 (AI), p : @2), . . . , p:(&)) E Pm. Since the selection of a particular label at a given vertex is related to the label selections made at other (not necessarily adjacent) vertices, information about the label selection at that vertex is contained in the initial labeling values distributed over the extent of the net- work. The function of the global enhancement process, g, is to accumulate this evidence into the labeling values at the given vertex. The output vector: is used by a process of local maxima selection [ 151, s , to choose a labeling: r; = (A,, 43 . . . I &J, where & is the label assigned to vertex ui. Thus g is a function, g:Rnm + Rnm and s is a function is s:Rnm + C,,(A), where C,(A) is the set of possible label- ings. The hope is that labeling resulting from the pro- cess s (e @)) is an improvement over the labeling result- ing from direct local maxima selections @). If a numerical solution is to be sought for this prob- lem, then a formal definition must be given to the con- cept of an improved labeling. In previous work, particu- larly with respect to computer vision, improvements were rated subjectively, or in the case of an experiment global enhancement process, g Fig. 1: Function of the global enhancement process: X represents an improved labeling with respect to 2. local -+ g+ maxima I-! -x selection, S where the solution was known in advance, by the number of misclassified vertices. In our formulation this issue is resolved by assuming that the problem domain specifies an underlying constraint network, or can be modeled to do so. The objective is then to use the initial information to choose a labeling that is (a) consistent with respect to this constraint network, and (b) which optimizes a prespecified objective function. In this extension from the discrete to the continuous graph labeling problem, the constraint relations remain intact. We are currently investigating optimal solutions to this formulation of the graph labeling problem based on a maaGnu~-sum decision rule, that is, the rule is to choose a consistent labeling such that sum of the initial labeling values is maximal. A solution to this problem could be extended in a straightforward manner to cer- tain well established decision rules such as,is found, for example, in nearest neighbor classification. Though the decision rule serves to make explicit what is meant by an improved labeling, it is defined glo- bally. The problem remains to implement it in terms of a cooperative process. The concept of a cooperative process, although not well defined, can be characterized in terms certain general properties [U1179]. Our research is into algorithms which exhibit certain of these properties, such as locality, and simplicity. In an aptimul solution, the labeling algorithm must, further- more, perform the label selection in accordance with the given decision rule. Other important issues, such as speed of convergence are also being addressed. Two approaches which have some of these properties are demonstrated in the following section. The first is an heuristic approach based on dynamic programming [ 141 which converges very rapidly and with good results, but does not guarantee a consistent labeling. The second approach is based on linear programming. Details on the latter algorithm will be presented at a later date. III -- In this section, we demonstrate the application of the two approaches discussed above to the problem of extracting polygon approximations of the outlines of objects in a scene. The experiments described here are based on the reconstruction of simple closed curves (Fig. 2) when noise has been added to the initial labeling values. The graph used in this experiment is a 16.by 16 ras- ter. Each vertex is represented by a pixel, and is adja- cent to its eight immediate neighbors. The associated label set is shown in Fig. 3. A pair of labels on adjacent pixels are consistent if an outgoing line segment is not broken across a common border or corner, and incon- sistent otherwise. Examples of consistent pairs of labels are given in Fig. 4, and examples of inconsistent pairs of labels are given in Fig. 5. *In terms of decision theory, every consistent label- ing constitutes a class and the input vector p is a point in an 7~x9~~ dimensional feature space. 51 Fig. 2: Initial labeling lzl lzl til la vi III 6l El q El q Ic El q lil III Gl w El q q Fig. 4: Examples of locally consistent label pairs. Uniformly independently distributed noise was added to the labeling values at each pixel resulting in the labeling, by local maxima selection, shown in Fig. 6. The two cooperative algorithms were applied to the ini- tial labeling in attempt to reconstruct the original curves. The first is the dynamic programming approach with data moving along the eight major directions of the raster (two horizontal, two vertical, and four diagonal). The second is the algorithm based on a linear program- ming approach. The performance of these algorithms are presented in Fig. ‘7 and Fig. 6, which show the result- ing labeling (by choosing the label with greatest strength at each pixel) after 2 and 4 iterations. The dynamic pro- gramming approach reaches a fixed point after 2 itera- tions, however, the result is not a consistent labeling. The linear programming algorithm reconstructs the ori- ginal labeling after six iterations. IV IIMXzWB Our interest here has been to restate the continu- ous graph labeling problem in a manner which allows for a systematic approachs to a solution. The formulation which we have presented amounts to the classification of Fig. 3: Label set for line drawing description. Fig. 5: Examples of inconsistent label pairs. 52 1 I I VI IU I I I III nwl I DI I Fig. 6: Initial labeling plus noise. I i i ii L-44 i ii i Fig. 8a: Output of the linear programming algorithm after two iterations. Fig. 7: Output of the dynamic programming algorithm after two iterations. Note: the algorithm has reached a fixed point. Fig. 8b: Output of the linear programming algorithm after four iterations. 53 consistent labelings according to a prespecified decision rule. As with previous approaches, consistency is defined on a local basis to make sense with respect to a [101 particular problem. For example, if the objective is to extract continuous curves as in the experiment described above, consistency is maintained between pairs of labels when the scene events they represent do 1111 not allow for broken lines. The global nature of the deci- sion rule leads to a more intuitive description of what the techniques accomplishes with respect to the original [12 problem. However, as a consequence, the problem of implementing this rule on a local basis arises. Two approaches to the reformulated problem have been demonstrated above. Our present feeling is that a [ 13 linear programming approach should yield an optimal solution to the continuous graph labeling problem based on a maximum-sum decision rule. However, the restric- tion that the algorithm must be implemented in a local manner has led to some theoretical problems, such as resolving cycling under degeneracy which remain to be [l-+1 solved. Our investigation into these problems is continu- ing. Obviously, the value of this approach and any tech- niques which may be derived from it will depend on whether or not real world applications can be modeled in [151 such a manner so that the absolute consistency between pairs of labels is meaningful. We hope to demonstrate this in at least one problem, deriving line drawings from real world scenes, in forthcoming results. PI PI L31 WI bl PI PI ml PI WI REFERENCES D. L. Waltz, “Generating semantic descriptions from drawings of scenes with shadows,” Technical [171 Report A1271, M.I.T., 1972. U. Montanari, “Networks of constraints: funda- mental properties and application to picture pro- cessing,” Information Sciences, vol. 7, pp. 95- 132, 1974. R. M. Haralick and L. G. Shapiro, “The consistent labeling problem: part I,” IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-1, pp. 173-184, 1979. R. M. Haralick and L. G. Shapiro, The consistent labeling problem: part II,” IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-2, pp. 193-203, 1980. A. K. Mackworth, “Consistency in networks of relations,” Artificial Intelligence, vol. 8, pp. 99- 118, 1977. A. Rosenfeld, R. A. Hummel, and S. W. Zucker, “Scene labeling by relaxation operations,” IEEE D-an-s. Syst., Man, Cybern., vol. SMCQ, pp. 420- 433. S. Peleg, “A new probabilistic relaxation scheme ,” IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-2, pp. 362-369, 1980. R. L. Kirby, “A product rule relaxation method,” Comput. CA-aphics Image Processing, vol. 12, pp. 158-189, 1980. S. Peleg and A. Rosenfeld, “Determining compati- bility coefficients for curve enhancement relaxa- tion processes,” IEEE Trans. Syst., Man, Cybem., vol. SMC-8, pp 548-555, 1978. H. Yamamoto, “A method of deriving compatibil- ity coefficients for relaxation operators,” Com- put. Graphics Image Processing, vol. 10, pp. 256- 271, 1978. S. IJllman, “Relaxation and constrained optimiza- tion by local processes,” Comput. Graphics Image Processing, vol. 11, pp. 115-125, 1979. R. A. Hummel, and S. W. Zucker, “On the founda- tions of relaxation labeling processes,” 7+8@7, Dept. of Elect. Eng., McGill University, Montreal, Quebec, Canada. 0. D. P’augeras, and M. Berthod, “Improving con- sistency and reducing ambiguity in stochastic labeling: an optimization approach,” IEEE Trans. Pattern Anal. Machine Intell,, vol. PAMI-3, pp. 412-424, 1981. M. D. Diamond, and S. Ganapathy, “Cooperative solutions to the graph labeling problems,” Proc. PRIP 82 Conference on Patt. Recognition and Image Processings, June, 1982, to appear. S. W. Zucker, Y. G. Leclerc, and J. L. Mohammed, “Continuous relaxation and local maxima selec- tion: conditions for equivalence IEEE Trans. Pat- tern Anal. Machine Intell., vol. PAMI-3, pp. 1:7- 127, 1981. L. S. Davis, and A. Rosenfeld, “Cooperating processes for low-level vision: a survey,” TR-123, Dept. of Computer Science, University of Texas, Austin, 1980. E. C. Freuder, “Synthesizing constraint eupres- sions ,I’ Comm. ACM, vol. 21, pp. 958-966, 1978. 54

1982

102