A PROCESS MODEL OF CASED-BASED REASONING IN ... - CiteSeerX

A PROCESS MODEL OF CASED-BASED REASONING IN PROBLEM SOLVING Janet L. Kolodner Robert L. Simpson, J r . K a t i a Sycara-Cyranski School of I n f o r m a t i o n and Computer Science Georgia I n s t i t u t e of Technology A t l a n t a , Georgia 30332 ABSTRACT Much of the problem s o l v i n g done by both novices and experts uses "case-based" r e a s o n i n g , or reasoning by analogy to previous s i m i l a r cases. We e x p l o r e the ways in which case-based reasoning can help in problem s o l v i n g . According to our model, t r a n s f e r of knowledge between cases is guided lar­ gely by the problem s o l v i n g process i t s e l f . Our model shows the i n t e r a c t i o n s between problem s o l ­ v i n g processes and memory f o r e x p e r i e n c e . Our com­ puter program, c a l l e d the MEDIATOR, i l l u s t r a t e s case-based reasoning in i n t e r p r e t i n g and r e s o l v i n g common sense d i s p u t e s . 1.

USING PREVIOUS CASES IN PROBLEM SOLVING

Most approaches to problem s o l v i n g in Al t r e a t each problem as a unique case. One d e f i c i e n c y of such problem s o l v e r s is t h a t they cannot r e l y on previous e x p e r i e n c e . The same problem presented two d i f f e r e n t times r e q u i r e s the same (possibly long and complex) set of reasoning s t e p s . Nor can such systems learn from the mistakes they have made. Suppose, however, t h a t our systems could augment t h e i r problem s o l v i n g c a p a b i l i t i e s by making analogies to previous s i m i l a r cases. Reference to a previous case focuses reasoning on those p a r t s of a c u r r e n t problem which were im­ p o r t a n t in a n a l y s i s of the previous case, poten­ t i a l l y reducing the number of f e a t u r e s t h a t need to be c o n s i d e r e d . When the process of reaching a s o l ­ u t i o n to a problem i n v o l v e s many s t e p s , analogy to a previous case can o f t e n reduce the number of steps r e q u i r e d to reach a s o l u t i o n . Furthermore, if the set of cases used in problem s o l v i n g i n c l u d e those where e r r o r s had been d e t e c t e d and later f i x e d , then analogy to a previous case can not o n l y help i n r e c o v e r i n g from s i m i l a r e r r o r s , but a l s o i n a v o i d i n g s i m i l a r f a i l u r e s i n the f u t u r e . The problem s o l v i n g process which solves problems by analogy to previous or h y p o t h e t i c a l cases is a "case-based r e a s o n i n g " process. Ap­ p l i c a t i o n o f p r e v i o u s experiences i n problem s o l ­ v i n g is something people seem to do in areas as d i v e r s e as mathematics, p h y s i c s , m e d i c i n e , and law * This research has been supported in p a r t by NSF Grant Nos. IST-8116892 and IST-8317711 and in p a r t by the A i r Force I n s t i t u t e of Technology. The views expressed are s o l e l y those of the a u t h o r s .

( e . g . , Gick and Holyoak, 1980; Polya, 1945; Reed and Johnson, 1977; R i s s l a n d , 1982). It plays an e s p e c i a l l y important r o l e in l e a r n i n g new tasks ( e . g . , Anderson, et a l . 1984; Ross, 1982), and is an important p a r t of making p r e d i c t i o n s in unders­ tanding (Schank, 1982) . The example below i l l u s t r a t e s several places where analogy to a previous case can help in under­ standing and s o l v i n g a c u r r e n t one. A reasoner reads about the Sinai dispute (before the Camp David Accords). She is reminded of the Korean War since both are over land, both are c o m p e t i t i v e , in n e i t h e r can the c o n f l i c t be resolved completely f o r both s i d e s , and in b o t h , m i l i t a r y f o r c e had been used previous to n e g o t i a t i o n s . Based on t h i s remin­ ding, she p r e d i c t s t h a t I s r a e l and Egypt w i l l d i v i d e the Sinai e q u a l l y . She l a t e r reads t h a t t h i s advice was given and r e j e c t e d by both s i d e s . She is reminded of her daughters' q u a r r e l over an orange. She had suggested t h a t they d i v i d e it e q u a l l y , and they had r e j e c t e d t h a t , since one wanted to use the e n t i r e peel f o r a cake. R e a l i z i n g t h a t she h a d n ' t taken t h e i r r e a l goals i n t o account, she then suggested t h a t they d i v i d e it agreeably by p a r t s — one t a k i n g the p e e l , the other the fruit. This provides the suggestion t h a t f a i l ­ ures may occur because the goals of the d i s p u t a n t s are misunderstood. She t h e r e f o r e a t ­ tempts a r e i n t e r p r e t a t i o n of Israel's and Egypt's g o a l s , and decides t h a t I s r a e l wants the Sinai as a m i l i t a r y b u f f e r zone in support of n a t i o n a l s e c u r i t y , and Egypt wants the land back f o r n a t i o n a l i n t e g r i t y . Further reasoning aided by analogy to the orange d i s p u t e leads to the c o n c l u s i o n t h a t an "agreeable d i v i s i o n " based on the r e a l goals of the d i s p u t a n t s is appropriate. She is reminded of the Panama Canal dispute since the disputants, disputed object, p a r t i c i p a n t g o a l s , and s e l e c t e d r e s o l u t i o n plan are s i m i l a r t o those i n the c u r r e n t d i s p u t e . Analogy to t h a t i n c i d e n t guides refinement of the "agreeable d i v i s i o n " p l a n . Replacing the US by I s r a e l (the p a r t y c u r r e n t l y in c o n t r o l of the o b j e c t ) and Panama by Egypt (the p a r t y who used to own it and wants it b a c k ) , she p r e d i c t s t h a t Egypt w i l l get economic and p o l i t i c a l con­ t r o l o f the S i n a i , w h i l e i t s normal r i g h t of m i l i t a r y control w i l l be denied. As the example shows, case-based reasoning can a i d in i n i t i a l understanding of a problem; in genera­ t i o n of s o l u t i o n s ; and in cases of misunderstanding

J. Kolodneretal.

o r plan f a i l u r e , i n r e i n t e r p r e t a t i o n and s e l e c t i o n of a l t e r n a t e l i n e s of reasoning. A problem solver must have several c a p a b i l i t i e s to employ case-based reasoning. It must locate previous s i m i l a r cases in a p o t e n t i a l l y large long term memory of past problem s o l v i n g episodes. It must i n t e g r a t e new cases i n t o memory f o r f u t u r e use. It must determine whether or not a r e c a l l e d case is ap­ p l i c a b l e in s o l v i n g the new problem, and if many previous cases are a v a i l a b l e , which are p o t e n t i a l l y the most a p p l i c a b l e . Finally, i t must t r a n s f e r knowledge c o r r e c t l y from the previous case to the c u r r e n t one. Much work is being done c u r r e n t l y in the area of making analogies across domains ( e . g . , B u r s t e i n , 1983; Centner, 1982; Winston, 1980), mostly f o r making automatic knowledge a c q u i s i t i o n of a new domain e a s i e r . In g e n e r a l , the analogous concept is given and the computer's j o b is to do the t r a n ­ sfer c o r r e c t l y . Our goals are somewhat d i f f e r e n t . Though analogy to a previous s i m i l a r case r e s u l t s in a c q u i s i t i o n of new or more r e f i n e d knowledge in a domain the system already knows about, we focus on a n a l o g y ' s r o l e s d u r i n g problem s o l v i n g . We thus consider the requirements it places on the problem s o l v i n g system and the c o n s t r a i n t s put on the a n a l o g i c a l process by the task demands of the problem s o l v e r . Since we i n v e s t i g a t e a process t h a t is meant to be p a r t of a problem s o l v e r , we r e q u i r e t h a t the computer l o c a t e , choose, and f i g u r e out how to apply a p p r o p r i a t e analogies by itself. II THE CASE-BASED

REASONING

PROCESS

The case-based reasoning process works as f o l ­ lows: (1) Locate and r e t r i e v e p o t e n t i a l l y ap­ p l i c a b l e cases from long term memory. (2) Evaluate s e l e c t e d cases to determine the a p p l i c a b l e ones. (3) Transfer knowledge from the o l d case (s) to the c u r r e n t one. We begin by explaining memory s t r u c t u r e s and processes t h a t a l l o w previous cases to be remembered. We then consider the r o l e s those cases can play in problem s o l v i n g and how problem solver task domains guide t r a n s f e r of knowledge from one case to another. We consider the types of memory schemata necessary to support our model, and the choice of cases when several are a v a i l a b l e . F i n a l l y , we present our program, the MEDIATOR, which uses analogy to previous cases to resolve disputes. For more i n f o r m a t i o n about other aspects of the process, see (Kolodner & Simpson, 1984 or Simpson, 1985)• IIl.

LOCATING CASES IN MEMORY

In order to use previous experience, a problem s o l v e r must be able to i n t e r a c t w i t h a memory f o r experience (Kolodner & Simpson, 1984) . The memory we use organizes experiences (cases) based on generalized episodes (Kolodner, 1984, Schank, 1982). These s t r u c t u r e s hold generalized knowledge d e s c r i b i n g a c l a s s of s i m i l a r episodes. An i n ­ dividual experience is indexed by f e a t u r e s which d i f f e r e n t i a t e it from the norms of the c l a s s (those f e a t u r e s which can d i f f e r e n t i a t e i t from other s i m i l a r e x p e r i e n c e s ) . As a new experience is i n ­ tegrated i n t o memory, i t c o l l i d e s w i t h other ex­ periences in the same g e n e r a l i z e d episode which share i t s d i f f e r e n c e s . We c a l l such a c o l l i s i o n a " r e m i n d i n g " (Kolodner, 1984, Schank, 1982). This t r i g g e r s two processes. Expectations based on the f i r s t episode can be used in a n a l y s i s of the new

285

one (analogy). S i m i l a r i t i e s between the two episodes can be compiled to form a new memory schema with the structure just described (generalization). Figure 1 shows how the "orange d i s p u t e " (from the example above) is organized in a g e n e r a l i z e d episode associated w i t h " p h y s i c a l d i s p u t e s " (i.e., those over possession of a p h y s i c a l o b j e c t ) . At the top of the f i g u r e are the norms f o r " p h y s i c a l d i s p u t e s . " Below t h a t , the "orange d i s p u t e " i s i n ­ dexed by those f e a t u r e s which s p e c i a l i z e it in the " p h y s i c a l d i s p u t e " generalized episode. It is in­ dexed by f e a t u r e s of both of i t s i n t e r p r e t a t i o n s . As new cases are added to the memory s t r u c t u r e , sub c a t e g o r i e s w i l l be created in the s t r u c t u r e at the places in which other cases correspond to t h i s one. The "orange d i s p u t e " w i l l help form the b a s i s of norms associated w i t h those c a t e g o r i e s . R e t r i e v a l of cases from memory and i n t e g r a t i o n of new cases i n t o memory are accomplished through a t r a v e r s a l process. As a new case is processed, ap­ p r o p r i a t e g e n e r a l i z e d episodes are identified. Features d i f f e r e n t i a t i n g the new case from o t h e r s organized in the same category are used to c r e a t e new i n d i c e s . Indices associated w i t h f e a t u r e s already present in the indexing s t r u c t u r e are traversed. Any previous cases the new case c o l ­ l i d e s w i t h are a v a i l a b l e f o r f u r t h e r e v a l u a t i o n . A previous experience can thus be "remembered" if it is organized in the same g e n e r a l i z e d episode a new case is being i n t e g r a t e d i n t o and a l s o shares a s e t of d i f f e r e n c e s w i t h t h a t case. The "orange d i s p u t e " can be r e c a l l e d from the " p h y s i c a l d i s p ­ u t e s " g e n e r a l i z e d episode shown above by any physical d i s p u t e f o r which " d i v i d e e q u a l l y " f a i l s ( e . g . , the Sinai d i s p u t e ) , any time the o b j e c t of dispute is edible, etc. Memory processes are i n t e g r a t e d w i t h problem s o l v i n g processes as f o l l o w s : As a case is being considered, it is represented by a set of schemata (frames) which d e t a i l f e a t u r e s of the case ( i n the slot f i l l e r s ) . At any moment, memory has a v a i l a b l e to it the c u r r e n t problem r e p r e s e n t a t i o n . A memory t r a v e r s a l process running c o n c u r r e n t l y w i t h problem solving processes uses the c u r r e n t problem r e p r e s e n t a t i o n as a key i n t o memory. Cases the memory encounters in t r a v e r s i n g memory using the c u r r e n t case as i t s search key are a v a i l a b l e f o r case-based reasoning. As the problem r e p r e s e n t a ­ t i o n changes, search is d i r e c t e d to places in memory referenced by the most c u r r e n t problem representation. IV.

CASE-BASED REASONING'S VARIED ROLES

Before d i s c u s s i n g the roles case based reasoning can play in problem s o l v i n g , we need to duscuss the problem s o l v i n g framework i n t o which we enter our case based processes. In order to use case-based reasoning, a problem s o l v e r must be a b l e to receive and evaluate feedback about the r e s u l t s of i t s d e c i s i o n s . A problem solver t h a t suggests s o l u t i o n s to problems but never knows the outcome of i t s advice has no basis f o r e v a l u a t i n g its d e c i s i o n s , and thus cannot be expected to use i t s experience r e l i a b l y i n d e a l i n g w i t h later cases. This suggests t h a t the problem s o l v e r must have f o l l o w - u p procedures, i n c l u d i n g procedures for r e c o g n i z i n g , e x p l a i n i n g , a s s i g n i n g blame, and a t ­ tempting t o c o r r e c t f a i l u r e s . In addition to that and plan g e n e r a t i o n procedures normally included in

286

J. Kolodneretal.

problem s o l v e r s , we a l s o include a problem unders­ tanding phase in which a partially-specified problem d e s c r i p t i o n is e l a b o r a t e d to the e x t e n t necessary f o r problem r e s o l u t i o n . These processes are d e p i c t e d in Figure 2. Case-based reasoning can play a r o l e in any of these t a s k s . PHYSICAL DISPUTES

A. Case-based

reasoning

in

i nterpretation

Before a problem can be s o l v e d , it must be un­ derstood. D e t a i l s not presented in the problem d e s c r i p t i o n must be f i l l e d in through i n f e r e n c e or query, and schemata (generalized episodes) for r e p r e s e n t i n g the problem must be chosen. Problem schemata p o i n t to p o t e n t i a l s o l u t i o n plans and thus are c r u c i a l to d e r i v i n g good s o l u t i o n s . Analogy to a previous case can d i r e c t these processes in the f o l l o w i n g two ways: 1. 2.

by suggesting a d d i t i o n a l f e a t u r e s to be investigated by suggesting a l t e r n a t i v e i n t e r p r e t a t i o n s

The f i r s t step in understanding a problem is to make a hypothesis about a schema t h a t might d e s c r i b e the problem. A d o c t o r , f o r example, may make a hypothesis about the d i s o r d e r a p a t i e n t has by f o c u s s i n g on r e p o r t e d symptoms. A f t e r represen­ t a t i o n hypotheses are made, case-based reasoning can help in v e r i f y i n g them. An attempt is made to integrate the new case i n t o the hypothesized g e n e r a l i z e d episodes, causing r e m i n d i n g s . If the c u r r e n t case is missing d e t a i l s necessary to v e r i f y a hypothesized schema, and a remembered case i n c l ­ udes those d e t a i l s , an attempt is made to t r a n s f e r those d e t a i l s to the c u r r e n t case. If t h e i r assum­ p t i o n is c o n s i s t e n t w i t h the c u r r e n t case, they are filled i n , and the r e l e v a n t hypothesis (the schema associated w i t h the remembered case) is v e r i f i e d . A case one is reminded of may a l s o have had f e a t u r e s important t o i t s proper resolution that were not p r e d i c t e d by the schema. These f e a t u r e s are i n v e s t i g a t e d . On the other hand, t r a v e r s a l may r e s u l t in reminding of a case t h a t i n i t i a l l y was thought to be represented by the hypothesized schema, but was l a t e r found to need a d i f f e r e n t representation. I n t h i s s i t u a t i o n , the c u r r e n t case is evaluated to see whether it might a l s o be represented the second way, thus a v o i d i n g the previous i n t e r p r e t a t i o n e r r o r . B.

Case-based reasoni ng in plan g e n e r a t i o n

A f t e r a case is understood, a plan f o r i t s r e s o l u t i o n must be generated. Our approach, which we r e f e r to as plan i n s t a n t i a t i o n , c o n s i s t s of two major s t a g e s : p l a n s e l e c t i o n and plan r e f i n e m e n t . Plan s e l e c t i o n i n v o l v e s choosing the best from among a set of known plans or c o n s t r u c t i n g a plan by combining a p p r o p r i a t e pieces of known p l a n s . Plan refinement i n v o l v e s r o l e b i n d i n g and a d j u s t i n g the plan f o r the p a r t i c u l a r s i t u a t i o n . When several plans are f e a s i b l e , the expected r e s u l t s of using the plan must be generated and evaluated to choose the best one. Analogy to previous cases can help plan g e n e r a t i o n in f o u r ways: 3. 4. 5. 6.

suggestion of procedures to be f o l l o w e d suggestion of procedures to be avoided s e l e c t i o n of a means of implementing a plan p r e d i c t i o n of the outcome of a s e l e c t e d plan

Plans a s s o c i a t e d w i t h cases remembered d u r i n g i n t e r p r e t a t i o n are a v a i l a b l e to suggest plans to be f o l l o w e d or a v o i d e d . The p r e c o n d i t i o n s of any suc­ cessful p l a n are checked, and if a p p l i c a b l e , the p l a n is a t t e m p t e d . At the same t i m e , any p l a n used

J. Kolodner et

in a p r e v i o u s case and resulting p r o h i b i t e d in the c u r r e n t case.

in

failure

is

If more than one plan is suggested, the u t i l i t y o f p o t e n t i a l p l a n s must be evaluated and the most a p p r o p r i a t e one c h o s e n . Plan e v a l u a t i o n involves simulating the results of alternative courses o f a c t i o n and e v a l u a t i n g t h e m . Evaluation c a n b e done t a k i n g p r e v i o u s experiences into ac­ count. Simulating the results of using a plan provides a hypothetical situation which may be. similar to a previous real one. The r e s u l t s o f previous attempts at implementing the same plan under similar conditions provide a way o f e v a l ­ u a t i n g a p o t e n t i a l course of action. Additional information leading t o a b e t t e r s t r a t e g y may a l s o be provided d u r i n g e v a l u a t i o n . This process is an e x t e n s i o n o f w h a t Schank (1982) r e f e r s t o a s i n t e n ­ tional r e m i n d i n g , and i s a component o f W i l e n s k y ' s (1983) P r o j e c t o r . Any c h o s e n p l a n must b e refined or adjusted for the current situation. A s u i t a b l e plan im­ p l e m e n t a t i o n can b e d e r i v e d b y c o n s i d e r i n g now it was effected in a previous case. W e see t h i s i n t h e S i n a i e x a m p l e when t h e Panama Canal agreement is used in d e c i d i n g how t o c a r r y o u t t h e " d i v i d e agreeably" plan. C.

Case b a s e d r e a s o n i n g

i n error

recovery

Problem s o l v i n g e r r o r s u s u a l l y appear a s plan failures, but t h e y can r e s u l t f r o m i n i t i a l m i s i n ­ terpretation, poor implementation of the plan, i n c o r r e c t p r e d i c t i o n o f t h e r e s u l t s , new u n e x p e c t e d o c c u r r e n c e s , o r bad p l a n s e l e c t i o n . When s i m i l a r failures have happened previously, the work i n v o l v e d i n e r r o r r e c o v e r y can b e c u t d o w n . In a s e n s e , e r r o r r e c o v e r y can b e v i e w e d a s a n o t h e r i n s ­ tance of problem s o l v i n g , t h i s time i n t e r p r e t i n g t h e f a i l u r e ( e x p l a i n i n g i t ) and f i x i n g the faulty knowledge (remediation). A p r e v i o u s case can t h u s play r o l e s s i m i l a r t o those played i n initial in­ t e r p r e t a t i o n and p l a n n i n g : 7. 8. 9.

suggestion of an explanation for the failure s u g g e s t i o n o f a p l a u s a b l e new interpreta­ tion suggestion of a plan f o r recovery

The procedure by which a previous similar f a i l u r e may p r o v i d e a c l u e to tracking down the error in the current case is as f o l l o w s : Upon f a i l u r e r e c o g n i t i o n , the reasoner will first at­ tempt to r e c a l l a s i m i l a r previous e r r o r . Indices corresponding to the features of the case, the chosen plan and t h e f a i l u r e , a r e t r a v e r s e d . If a s i m i l a r f a i l u r e has o c c u r r e d previously, the ex­ p l a n a t i o n from t h a t f a i l u r e acts as a guide to con­ structing a hypothesis to explain the current failure. The r e m i n d i n g may a l s o i n c l u d e suggested plans for error recovery. Consider again our example in section one. The e r r o r r e c o v e r y t h a t o c c u r s a f t e r the mother's failure to correctly predict a s o l u t i o n to the Sinai d i s p u t e is as f o l l o w s : The m o t h e r remembers her previous failure with respect to the orange dispute. This reminding guides the c o n s t r u c t i o n of the f o l l o w i n g hypothesis: Perhaps t h e g o a l s o f t h e d i s p u t a n t s have b e e n i n f e r r e d i n c o r r e c t l y . If so, use " a g r e e a b l e d i v i s i o n " b a s e d o n t h e i r r e a l g o a l s .

al.

287

In this case, an a l t e r n a t e c l a s s i f i c a t i o n is sug­ g e s t e d (one based o n c o n c o r d a n t rather than com­ petitive g o a l s ) , a n e x p l a n a t i o n f o r the f a i l u r e i s suggested (the wrong g o a l s were b e i n g considered), and a p l a n f o r r e c o v e r y i s s u g g e s t e d ( f i n d t h e r e a l g o a l s and a p p l y t h e p l a n " a g r e e a b l e d i v i s i o n " ) . V. GUIDING THE TRANSFER 0F KNOWLEDGE Treating analogy as p a r t of a problem s o l v i n g p r o c e s s a l l o w s t a s k demands o f t h e problem solver t o g u i d e t h e t r a n s f e r o f k n o w l e d g e f r o m one c a s e t o another one. During i n t e r p r e t a t i o n , the problem s o l v e r needs t o i n f e r m i s s i n g information and to choose problem classifications. When inferring missing i n f o r m a t i o n , it looks at those f e a t u r e s of a previous case designated by the hypothesized generalized episode as necessary for clas­ sification. When enough information t o make a d e f i n i t i v e c l a s s i f i c a t i o n is a v a i l a b l e , the problem solver attempts to transfer the classification, checking t o make s u r e t h a t c o n d i t i o n s f o r r e c o g n i ­ t i o n are met. During plan g e n e r a t i o n , the problem solver attempts to transfer plans associated w i t h previous cases. To d e c i d e if a p a r t i c u l a r p l a n can be transferred, its preconditions are checked against the features of the c u r r e n t case. Error recovery involves f a i l u r e c l a s s i f i c a t i o n and then replanning. Missing i n f o r m a t i o n about the f a i l u r e n e c e s s a r y t o c l a s s i f y and p o s s i b l y explain it is i n v e s t i g a t e d and t r a n s f e r r e d f r o m a p r e v i o u s c a s e . Previous means of r e s o l v i n g the e r r o r ( i . e . , the remediation plan) are checked during replanning. Preventing the i n a p p r o p r i a t e t r a n s f e r of knowledge is complementary to enabling its transfer. Knowledge transfer i s p r o h i b i t e d when i t v i o l a t e s designated consistency constraints (e.g., precon­ d i t i o n s and e x c l u s i o n a r y p r i n c i p l e s ) . VI..

MEMORY SCHEMATA

T h e r e a r e two i m p o r t a n t i s s u e s t o d i s c u s s w i t h r e s p e c t t o l o c a t i n g c a s e s i n memory: what k i n d s of indices are necessary, and what kinds of g e n e r a l i z e d e p i s o d e s have p r o v e n u s e f u l ? Two t y p e s o f f e a t u r e s a r e used for indexing: features of the case and f e a t u r e s d e s c r i b i n g any f a i l u r e s encountered during analysis of a case. If blame is a s s i g n e d f o r a f a i l u r e and a correction has been found, the case is indexed by t h o s e f e a t u r e s w h i c h caused the f a i l u r e . When a second situation with those features is encountered, r e m i n d i n g o f t h e f i r s t p r o v i d e s a means o f a v o i d i n g f a i l u r e t h e second t i m e . When b l a m e has not been assigned, the features indexing the case by i t s d i f f e r e n c e s w i l l serve to direct reminding to a similar case. Though a p a r t i c u l a r procedure f o r a v o i d i n g f a i l u r e may n o t b e f o u n d , t h e f a i l e d ins­ tance will alert the reasoner of a potential problem. These two t y p e s o f i n d i c e s allow remin­ ding of (1) c a s e s w h i c h a r e s i m i l a r d e s c r i p t i v e l y to a current one and (2) those which failed similarly. Two t y p e s of g e n e r a l i z e d e p i s o d e s have p r o v e n useful f o r case-based problem s o l v i n g : descriptive and t a c t i c a l o n e s . Descriptive categories describe the types of problems the problem solver will encounter. Each s p e c i f i e s d e s c r i p t i v e f e a t u r e s and suggested remedies. In our problem solving f r a m e w o r k , two t y p e s o f p r o b l e m s a r e d e a l t w i t h by the problem s o l v e r : d o m a i n p r o b l e m s and f a i l u r e s . G e n e r a l i z e d e p i s o d e s c o r r e s p o n d i n g t o each o f t h e s e

288

J. Kolodner et al.

problem types are necessary. In the m e d i a t i o n domain, these include dispute types (e.g., p h y s i c a l , economic, p o l i t i c a l ) and types of media­ tion failures (e.g., "wrong goal inference", "incorrect classification", "wrong planning p o l i c y " ) . Figure 1 shows some of the norms and ap­ p l i c a b l e plans f o r " p h y s i c a l d i s p u t e s . " Similarly, the d e s c r i p t i v e f a i l u r e category "wrong goal inf e r e n c e " holds a normative description and s p e c i f i e s plans f o r i n f e r r i n g the r i g h t goal ( e . g . , "infer goal from r e s u l t i n g e v e n t s " ) . Descriptive c a t e g o r i e s are used d u r i n g understanding tasks -problem i n t e r p r e t a t i o n and f a i l u r e e x p l a n a t i o n . Tactical c l a s s i f i c a t i o n s of experience c o r ­ respond to p a r t i c u l a r p l a n s . They d e s c r i b e precon­ ditions, implementation d e t a i l s , and expected r e s u l t s of p l a n s , and organize cases in which the p l a n was used. In the m e d i a t i o n domain, we use tactical g e n e r a l i z e d episodes corresponding t o m e d i a t i o n plans ( e . g . , "divide equally", "divide i n t o d i f f e r e n t p a r t s " ) and r e m e d i a t i o n p l a n s , i . e . , plans f o r recovery from p a r t i c u l a r p l a n n i n g e r r o r s ( e . g . , "change p l a n n i n g p o l i c y " , " i n f e r goal from later actions"). T a c t i c a l g e n e r a l i z e d episodes are used d u r i n g the p l a n n i n g stages to determine whether a suggested plan is a p p r o p r i a t e (using i t s p r e c o n d i t i o n s ) , to f i n d a means of implementing a plan, and to p r e d i c t and e v a l u a t e the consequences of using a suggested p l a n . During e r r o r recovery, classifications a s s o c i a t e d w i t h r e m e d i a t i o n plans help in c r e a t i n g new problem i n t e r p r e t a t i o n s and selecting a l t e r n a t i v e resolution plans. VII.

CHOOSING THE BEST CASE

In a r i c h l y indexed e p i s o d i c memory, many remindings may occur d u r i n g problem s o l v i n g . While the memory s t r u c t u r e c o n s t r a i n s the p o s s i b i l i t i e s to some e x t e n t , t h e r e must be a way of choosing p o t e n t i a l l y r e l e v a n t previous cases from the set of r e m i n d i n g * . There are two ways to do t h i s . In the f i r s t method, the best case is chosen from the set of remindings through an "a p r i o r i " e v a l u a t i o n procedure, i . e . , one t h a t takes o n l y closeness of fit t o the c u r r e n t case i n t o account. Using t h i s method, if the s e l e c t i o n l a t e r proves to be inap­ plicable (e.g., due t o incompatible p r e c o n d i t i o n s f o r the suggested p l a n ) , a second choice is made by the same e v a l u a t i o n procedure. T h i s method is ac­ ceptable if f a i l u r e s are not expensive o r i r ­ recoverable. Because "a p r i o r i " e v a l u a t i o n may not always be r e l i a b l e , however, another method of choice must be used when more c a r e f u l l y t h o u g h t - o u t s o l u t i o n s are necessary. In t h i s case, an e v a l u a ­ t i o n procedure is used to rank cases, and a set of h i g h l y - r a n k e d ones are chosen. Suggestions from each of the s e l e c t e d cases are c o n s i d e r e d , a plan is generated based on each one, the generated plans are e v a l u a t e d , and a best p l a n is chosen. Since the m e d i a t i o n domain is one where f a i l u r e has l i t t l e consequence, we use the f i r s t method. The best previous case is chosen u s i n g an e v a l u a t i o n procedure based on C a r b o n e l l ' s (1982) invariance hierarchy. Priority i s attached t o d i f f e r e n t feature types. A l t e r n a t i v e cases are e v a l u a t e d according to a s e r i e s of e l i m i n a t i o n and ranking t e s t s . Our s e t of t e s t s is as f o l l o w s : 1.

E l i m i n a t e those cases where the goal r e l a t i o n s h i p i s d i f f e r e n t from t h a t i n the c u r r e n t case.

2. 3.

E l i m i n a t e those where the d e r i v a t i o n of the goal r e l a t i o n s h i p i s d i f f e r e n t from t h a t of the c u r r e n t case. Order the remaining cases by the f o l l o w i n g rankings: similar d i s p u t a n t argument » s i m i l a r d i s p u t a n t s » s i m i l a r d i s p u t e d ob­ ject

Consider how these work f o r the f o l l o w i n g example: The t h i r d - w o r l d and i n d u s t r i a l n a t i o n s both want r i g h t s to the m i n e r a l s in the w o r l d ' s sea beds. It has already been decided t h a t they w i l l be d i v i d e d . Neither s i d e , however, t r u s t s the other t o d i v i d e the r e s o u r c e s . Suppose our reasoner were reminded of three cases: (1) two c h i l d r e n f i g h t i n g over a candy bar — the s o l u t i o n is "one cuts the other chooses", (2) the d i s p u t e between I s r a e l and Egypt over the S i n a i , and (3) the d i s p u t e between the US and Russia over f i s h i n g r i g h t s o f f the US c o a s t . Using r u l e 1, (2) is e l i m i n a t e d s i n c e it i n v o l v e s a concordant goal r e l a t i o n s h i p r a t h e r than a c o m p e t i t i v e one. Rule 2 eliminates (3) since the goals of the d i s p u t a n t s are d e r i v e d d i f f e r e n t l y . In the c u r r e n t case, the goal is d e r i v e d from the d i s p u t a n t s ' d e s i r e to con­ t r o l and use a consumable r e s o u r c e , w h i l e in (3)» the goal d e r i v e s from an i n t e n t i o n to c o n t r o l a renewable r e s o u r c e . Case (1) is chosen as most ap­ p l i c a b l e , and an analogy based on the s u p e r f i c i a l s i m i l a r i t i e s between the c u r r e n t case and case (3) is eliminated. VI I I.

AN EXAMPLE FROM THE MEDIATOR

In the f o l l o w i n g example, we see the MEDIATOR (Simpson, 1985) r e s o l v i n g the Sinai d i s p u t e based on analogies to several d i s p u t e s it already knows about. I n i t i a l l y , the MEDIATOR is t o l d t h a t Egypt and I s r a e l both want p h y s i c a l c o n t r o l of the S i n a i , and t h a t m i l i t a r y means have been used p r e v i o u s l y t o attempt t o achieve t h a t . In attempting to c l a s ­ s i f y the d i s p u t e i n t o one of i t s known d i s p u t e types, the MEDIATOR is reminded of two previous cases, the Panama Canal d i s p u t e and the Korean con­ flict. Using i t s evaluation function (not ex­ plained here), it chooses the Korean c o n f l i c t as most a p p l i c a b l e since it shares more important features. Since t h a t d i s p u t e d o e s n ' t help it f u r t h e r i n t e r p r e t the problem, it goes on to the p l a n n i n g phase and attempts to t r a n s f e r the p l a n used s u c c e s s f u l l y to r e s o l v e the Korean c o n f l i c t . It checks t h a t p l a n ' s p r e c o n d i t i o n s f o r ap­ p l i c a b i l i t y to the new case. RECALLING PREVIOUS DISPUTES TO CLASSIFY THIS ONE .. reminded of the "Panama Canal D i s p u t e " because both d i s p u t a n t s are of type M-POLITY. reminded of the "Korean C o n f l i c t " because both o b j e c t s are of type M-LAND and both used M-MILITARY-FORCE to attempt *PHYS-C0NTR0L* Choosing the "Korean C o n f l i c t " ATTEMPTING TO SELECT A MEDIATION PLAN TO RESOLVE THE " S i n a i D i s p u t e " Using the "Korean C o n f l i c t " which was r e s o l v e d using " d i v i d e e q u a l l y " Checking f o r a p p l i c a b i l i t y o f t h a t p l a n . . . I suggest " d i v i d e e q u a l l y " be used.

J. Kolodneretal.

The MEDIATOR a s k s f o r f e e d b a c k a b o u t its decision and is told b o t h E g y p t ' s and I s r a e l ' s r e a c t i o n s . I t a t t e m p t s t o come u p w i t h a new solution, and considers the f a i l u r e of the suggested p l a n as the c u r r e n t problem to be resolved. It applies the same problem s o l v i n g process to the f a i l u r e to ex­ p l a i n i t and r e i n t e r p r e t .

Is t h i s a good s o l u t i o n ? (Y or N) No. * * * * DIVIDE EQUALLY n o t a c c e p t a b l e * * * * What happened? ~(we show t h e E n g l i s h e q u i v a l e n t ) I s r a e l says t h e want t h e S i n a i f o r s e c u r i t y E g y p t s a y s t h e y want i t f o r i n t e g r i t y ATTEMPTING TO EXPLAIN FAILURE AND FIND NEW SOLUTION. RECALLING PREVIOUS FAILURES reminded of " t w o s i s t e r s q u a r r e l over an o r a n g e " because i n b o t h " d i v i d e e q u a l l y " f a i l e d and b o t h o b j e c t s a r e o f t y p e M-PHYS-OBJ F a i l u r e was b e c a u s e o f M-WRONG-GOAL-INFERENCE. Transferring that c l a s s i f i c a t i o n to this f a i l u r e . A t t e m p t i n g t o use remedy c a l l e d " i n f e r goal from r e s u l t i n g a c t i o n s " U n a b l e t o use p r e v i o u s r e m e d y . C o n s i d e r i n g o t h e r r e m e d i e s M-WRONG-GOAL-INFERENCE Looking at " i n f e r goal from r e s p o n - e " Based o n t h e f e e d b a c k , I r e p l a c e ISRAEL'S g o a l w i t h a M-NATIONAL-SECURITY g o a l and EGYPT'S g o a l w i t h a M-NATIONAL-INTEGRITY g o a l . Remediation complete.

The MEDIATOR n e x t r e p r o c e s s e s t h e d i s p u t e . Because t h e p r o b l e m has been r e - i n t e r p r e t e d , there is no need to reference previous cases until plan selection. The r e m i n d i n g p r o c e s s (left out this t i m e ) r e t r i e v e s t h e same two c a s e s a s b e f o r e . This time the a d d i t i o n a l i n f o r m a t i o n about the goals of t h e d i s p u t a n t s c a u s e s t h e MEDIATOR t o focus on a different exemplar, the Panama Canal dispute. Using t h a t as a model, it suggests giving Egypt political c o n t r o l o f the Sinai but g i v i n g m i l i t a r y control to Israel.

R e c o n s i d e r i n g t h e p r o b l e m u s i n g new i n f o r m a t i o n . Considering the r e - i n t e r p r e t e d problem: I s r a e l and E g y p t b o t h w a n t t h e S i n a i , w h i c h has been p r e s e n t e d as ako M-PHYS-DISPUTE. ATTEMPTING TO SELECT A MEDIATION PLAN TO RESOLVE THE " S i n a i D i s p u t e " RECALLING SIMILAR DISPUTES . . . Reminded o f t h e "Panama Canal D i s p u t e " ... resolved using " d i v i d e into d i f f e r e n t p a r t s " . Checking f o r a p p l i c a b i l i t y o f t h a t plan . . . I suggest " d i v i d e i n t o d i f f e r e n t p a r t s " be used. U s i n g t h e "Panama Canal D i s p u t e " t o c r e a t e p l a n m a t c h i n g ISRAEL w i t h USA . . . m a t c h i n g EGYPT w i t h PANAMA... m a t c h i n g SINAI w i t h PANAMA-CANAL... matching (*G0AL* (*NAT-SECURITY* (ACTOR ISRAEL) (OBJECT S I N A I ) ) ) w i t h (*G0AL* (*MIL-C0NTR0L* (ACTOR USA) matching (OBJECT P A N A M A - C A N A L ) ) ) . . . (*G0AL* (*NAT-INTEGRITY* (ACTOR EGYPT) (OBJECT S I N A I ) ) ) w i t h (*G0AL* (*P0L-C0NTR0L* (ACTOR PANAMA) (OBJECT PANAMA-CANAL))) . . . t r a n s f e r r i n g o t h e r components unchanged.

IX.

289

SUMMARY

Though w e have p r e s e n t e d a f r a m e w o r k f o r u s i n g case-based reasoning in problem solving, we have not covered a l l p a r t s of the process in t h i s paper. The problem solver must b e a b l e t o keep t r a c k o f a t t e m p t e d s o l u t i o n s s o t h a t i t does n o t r e p e a t bad attempts. I t i s a l s o necessary t o develop c r i t e r i a to evaluate the feasibility of a case-based ap­ p r o a c h f o r any g i v e n p r o b l e m . T r a n s f e r must a l s o b e g i v e n more c o n s i d e r a t i o n . Direct transfer is often not enough. Sometimes o n l y p a r t o f a s o l u ­ t i o n s h o u l d b e t r a n s f e r r e d , and o f t e n t r a n s f e r itself is a n a l o g i c a l . M u l t i p l e cases r e p r e s e n t i n g a c o m b i n a t i o n o f a n a l o g i e s must a l s o b e c o n s i d e r e d i n i n t e r p r e t i n g or c r e a t i n g a plan to solve a problem. M u l t i p l e c a s e s may each p r o v i d e p a r t s o f a s o l u t i o n w h i c h must t h e n b e p i e c e d t o g e t h e r . This work derives from research in the c o g n i t i v e m o d e l l i n g o f l o n g t e r m memory ( K o l o d n e r , 1984, S c h a n k , 1982) and uses t h o s e t h e o r i e s a s its basis. At t h e same t i m e , w e m a i n t a i n c o n s i s t e n c y with relevant psychological work (e.g., Ross, 1982). Like Hammond (1983). who a l s o b a s e s h i s work on t h o s e m o d e l s , we p r o v i d e a model of case based reasoning which i n t e g r a t e s problem s o l v i n g , u n d e r s t a n d i n g and memory. While his work uses p r e v i o u s c a s e s f o r p l a n s e l e c t i o n and a v o i d a n c e , w e use them additionally f o r problem i n t e r p r e t a t i o n and e r r o r r e c o v e r y . Carbone11 (1983) i s c l o s e s t t o our r e s e a r c h , though h i s problem solving follows the search paradigm. I n our a p p r o a c h , t r a d i t i o n a l n o t i o n s of search are absent. Rather, we combine analogical transfer w i t h plan i n s t a n t i a t i o n . REFERENCES Anderson, J. R., B o y l e , C . F . , F a r r e l l , R . , and Reiser, B. (I984). Cognitive Principles in the d e s i g n of computer t u t o r s . I n P r o c e e d i ngs o f t h e S i x t h Annual C o n f e r e n c e o f t h e C o g n i t i v e S c i e n c e S o c i e t y , B o u l d e r , CO., 2 - 9 . B u r s t e i n , M. H. (1983). A model o f learning analogical reasoning and debugging. Proceedings of the National Conference A r t i f i c i a l I n t e l l i g e n c e , Washington, D . C,

W.

by In on 45

Carbonell, J. G., Jr. (1982). Metaphor: An i n e s c a p a b l e phenomenon i n n a t u r a l - 1 a n g u a g e c o m ­ prehension. In W. G. Lehnert and M. H. Ringle (Eds.) S t r a t e g i e s f o r N a t u r a l Language P r o c e s s i n g , H i l l s d a l e , NJ: Erlbaum A s s o c i a t e s . Carbonell, J. G., Jr. (1983). Derivational Analogy in Problem Solving and Knowledge Acquisition. In Proceedings of the International Machine L e a r n i n g Workshop. Mont i c e l l o , I l l i n o i s , 12-18. Gentner, D. (1982). Structure Mapping: a theoretical framework for analogy and similarity. In Proceedings of the Fourth Annual Conference of the Cognitive Science S o c i e t y , 13-15Gick, M. L. and Holyoak, K. J. Analogical Problem Solving. Psychology 1 2 , 306-355-

(I980)• Cognitive

290

J. Kolodner et al.

Hammond, K. J. (1983). Planning and Goal In­ teraction: the use of past s o l u t i o n s in present s i t u a t i o n s . In Proceedings of the N a t i o n a l Conference o n A r t i f i c i a l I n t e l l i g e n c e , Washington, D . C , 1 4 8 - 1 5 1 . Kolodner, J. L. (1984) . Retrieval and O r g a n i z a t i o n a l S t r a t e g i e s in Conceptual Memory: A Computer Model♦ H i l l s d a l e , NJ: Lawrence Erlbaum A s s o c i a t e s . Kolodner, J. L. and Simpson, R. L. (1984). Ex­ perience and problem s o l v i n g : a framework. In Proceedings of the S i x t h Annual Conference of the C o g n i t i v e Science S o c i e t y . Boulder, CO., 2-9. Polya, G. (1945) • How to solve NJ: P r i n c e t o n Univ Press.

it.

Princeton,

Reed, S. K. and Johnsen, J. A. (1977). Memory f o r problem s o l u t i o n s . In Bower, G. H. (Ed.) The Psychology of Learning and M o t i v a t i o n , New York: Academic Press. Rissland, E. (1982). Examples in the legal domain: h y p o t h e t i c a l s in c o n t r a c t law. In Proceedings of the Fourth Annual Conference of the Cogni t i v e Science S o c i e t y , Ann Arbor, M i c h i g a n , 96-99. Ross, B r i a n H. (1982). Remindings and Thei r Effects in Learning a Cognitive Skill. C o g n i t i v e and I n s t r u c t i o n a l Sciences Series CIS 19. Palo A l t o , CA: Xerox Palo A l t o Research Centers. Schank, R. bridge:

C. (1982). Dynamic Memory. Cambridge U n i v e r s i t y Press.

Cam­

Simpson, R. L. (1985). A Computer Model of Case Based Reasoning in Problem S o l v i n g . Ph.D. Thesis. Report # G I T - I C S - 8 5 / l 8 , School of ICS, Georgia I n s t , of Technology, A t l a n t a GA. Wilensky, R. (1983). Planning and Understanding: A Computational Approach to Human Reasoning. Reading, MA: Addison-Wesley P u b l i s h i n g Com­ pany. Winston, P. H. (1980). Learning and reasoning by analogy. Communications of the ACM, 23, 689 703.