A Memory Model for an Intelligent System for the ... - Semantic Scholar

1 downloads 0 Views 156KB Size Report
based reasoning (DARPA), Pensacola Beach, Florida, K.J.. Hammond (Edt.), Morgan Kaufmann, San Mateo, CA, 1989,. 181-187. 2] Aamodt Agnar and Plaza ...
A Memory Model for an Intelligent System for the 21st Century Isabelle Bichindaritz

LIAP-5, UFR de Mathematiques et Informatique, Universite Rene Descartes 45 rue des Saints-Peres, 75006 Paris France ABSTRACT

Originally, case-based reasoning emerged from Schank's theory of dynamic memory. It has then been presented as an arti cial intelligence methodology for processing empirical knowledge. Nevertheless, more recent case-based reasoning systems study how to take advantage from theoretical knowledge to process empirical knowledge more e ectively. In order to reach this goal, they propose memory models faithful to the ideal of dynamic memory, for which reasoning and learning are inseparable. The memory model presented here is composed of two parts, an experimental and a theoretical memory, expressed in a uni ed knowledge representation language, and organization. The components of the memory are cases and concepts, in the experimental part, and prototypes and models, in the theoretical part. The reasoning supported by this memory model can be various, and takes advantage of all the components, whether experimental or theoretical. It is strongly constrained by some specialized models in theoretical memory, called the points of view.

1. INTRODUCTION

Case-based reasoning is an AI methodology for processing experimental knowledge [2], a case being a set of empirical data. It has been originally proposed as an alternative to expert systems methodology [1], which is an AI methodology for processing theoretical knowledge. One way of de ning theoretical knowledge is that it is knowledge that has lost all links with the objects from which it has been learnt, whereas experimental knowledge has kept at least some of these links. If the rst case-based reasoners were totally empirical systems, they progressively integrated more and more theoretical knowledge. The role of theoretical knowledge is then to constrain case-based reasoning, which is the main kind of reasoning realized by such a system. From the beginning, research in case-based reasoning has bene tted the most from interactions with cognitive modelling, and from cognitive psychology. These domains constantly reassess that memory and intelligence are closely interrelated. In the same way, a look back to the precursory theories of case-based reasoning, essentially the theory of dynamic memory [23], gives the ideal of a system memory where learning and reasoning are closely linked. For that purpose, this paper proposes a dynamic memory model, where the two types of knowledge (experimental and theoretical) are represented

in a uni ed knowledge representation language, and memory organization. This permits to both theoretical and experimental knowledge to be represented and manipulated in an integrated framework. The second section gives the de nitions that are important for the rest of the paper. The third section presents the evolution of case-based reasoning towards the use of more and more theoretical knowledge. The fourth section deals with related research in cognitive psychology and cognitive modelling. The fth section proposes an integrated memory model where both theoretical and experimental knowledge are represented and processed. Finally, an application of this model to eating disorders in psychiatry is presented. It is followed by the conclusion.

Case

2. DEFINITIONS

A case is a set of empirical data either from an application domain, or created by a system to represent one of its experiences. Between these two extremes, raw cases from an application domain [9], and re ned cases, entirely built by the system, all levels of elaboration may exist.

Case-based Reasoning

A case-based reasoning system is a system that prefers, in order to process a new case, to use one or several previously met (or created) cases.

3. EVOLUTION OF CASE-BASED REASONING Theoretical Knowledge for Case-based Reasoning

Di erent approaches have been proposed to build systems taking advantage from both theoretical and experimental knowledge in arti cial intelligence. These approaches have been turned towards hybrid systems, combining case-based reasoning, for the processing of experimental knowledge, and rule-based [12,15,22] or model-based reasoning, for the processing of theoretical knowledge. The most collaborative work has been proposed between model-based and case-based reasoning. In some systems, model-based reasoning takes a prominent part in the reasoning process. In CASEY [19], model-based reasoning determining the etiology of a heart disease is available, and the instanciations of this model are progressively gathered in a case-base in order to short-cut model-based

reasoning, so that this reasoning is less and less used by the ory [23] is the achievement of these theories. In particular, system. Also, in KRITIK [14], the memory keeps the instan- this theory does not propose a hybrid solution between ruleciations of a qualitative model, as in CASEY, but also resorts based reasoning and an experimental type of reasoning, but to model-based reasoning for the adaptation step. a model where experimental and theoretical knowledge are Many case-based reasoners use a form of model-based rea- integrated in a single memory, so as to ease the interactions soning to constrain the reasoning at one step or the other : between them. This memory is represented in gure 1. It is for instance, ARCHIE [11] to evaluate the pertinence of cases a general model that has been only partially implemented during the extraction step, CASCADE [24] to validate the in the case-based reasoning systems that have followed in its retrieved cases, KRITIK [13] in the adaptation step, CHEF track [18,17]. [16] to repair its failures, and [10] for an explanation-based indexing. 4. MEMORY IN COGNITIVE MODELLING In these systems, case-based reasoning is the main reasoning, and model-based reasoning plays the role of a heuristic to con- Psychologists, among the most distinguished, have for a long strain the case-based reasoning whenever theoretical knowl- time studied the relations between memory and intelligence. edge is modelled. The interaction of the two methodologies The works of Piaget [21] were greatly interested in these relais tighter than the one between rule-based and cased-based tions. They laid stress on the ambiguities of the word memory, reasoning, because models and cases store similar knowledge and distinguished between two types of memory : [17] : large chunks of interrelated knowledge, and not frag1. the memory in its strict meaning only deals with remented knowledge as in rules. membering situations, processes, or objects which are singular, and either recognized or recalled as such ; its main distinctive property is that it references the past explicitly ; u-MOP generalizes Universal 2. the memory in its broad meaning deals with remem   bering schemes themselves, which are general ; for example,     the concept of square is a scheme, whereas square symbols 9 organizes are examples of this scheme ; its main distinctive property meta-MOP is that it faces the future. ? The authors try to separate memory from intelligence, with u-MOP generalizes organizes  no success. They conclude : \There is no immutable border     but a series of mobile borders between the mnesic act and the   act of intelligence in general : everything partakes of memory organizes ? 9 if one stands in the point of view of the memory in its broad MOP meaning, without which neither understanding of the present ? nor invention are possible.". Universal Scene This working unity between memory and intelligence is the generalizes organizes    aim of the dynamic memory model, and so the cognitive      model of case-based reasoning.  9

?

Generalized Scene

organizes

?

Script

The memory network dependent upon a domain in the theory of dynamic memory (u-MOP = Universal MOP).

Figure 1.

For an Integration of Experience and Theory

The precursors of case-based reasoning are the successive theories of Schank, that progressively have established the main role of memory for reasoning. The theory of dynamic mem-

5. A MEMORY MODEL FOR AN INTELLIGENT SYSTEM

In case-based reasoning, the integration of experimental and theoretical knowledge must be studied from a memory point of view. The composition of the memory, and its organization, direct the reasoning. The memory is a network of nodes linked by relations. The nodes may contain any entity composing the memory.

Memory Composition

The entities composing the memory are all expressed in a uni ed knowledge representation language, that of binary predicates. Since unary and n-ary predicates can be reduced to conjunctions of binary predicates, this formalization is not restrictive [25].

Experimental Entities: The experimental entities are concepts and cases. A concept is an entity learnt by the system from cases, by

generalization or abstraction [8]. Cases are entities which instantiate the concepts. They are linked to the concepts learnt from them. A case ci is represented by a conjunction of binary predicates, of two types : attributes describing static knowledge Att, and relations Rel between nodes no1r and no2r :

ci =

^ Relr (no1r ; no2r ) ^ Attt(valt) t

r

(1)

Each of these binary predicates is called a description element. A concept Cj is also represented by a conjunction of binary predicates, of the two types :

Cj =

^ Rels (no1s; no2s) ^ Attu(valu ) s

u

(2)

Experimental Memory: The experimental memory ME is a set of cases ci and concepts Cj organized in a network by relations Rk . ME =< K; C ; R > with K = f:::ci:::g, C = f:::Cj :::g and R = f:::Rk (n1k ; n2k ):::g with n1k and n2k 2 K [ C .

Theoretical Memory: The theoretical memory MT is a set of prototypes mi and models Mj organized in a network by relations Rk . MT =< P ; D; R >

Theoretical Entities: The theoretical entities are models with P = f:::mi:::g, D = f:::Mj :::g and prototypes. A model is a representation of a system from the application and R = f:::Rk (n1k ; n2k ):::g with n1k and n2k 2 P [ D.

domain, or independent from it, and is given to the system for granted; it is not learnt from its experiences. A prototype mi is an entity which instantiates a model, and is Relations: The relations are various. They can be classiconsidered as typical of this model. In domains where models ed in : are not available, prototypes are the only theoretical knowledge t for use. They are not, in this case, linked to models. 1. structural relations : they describe the disposition of objects with regard to one another, or with regard with a The representation of a prototype is the same as that of a set. They can be intrinsic structural relations, when they case : describe the disposition of objects with regard to one another (for example Has for abstraction/Is an abstrac(3) mi = Relr (no1r ; no2r ) Attt(valt ) tion of ! Is ? a(x; y)=Abstr(x; y) or Has as instance/Is t r an instance of ! Has ? inst=Is ? inst(x; y)). They can A model Mj is also represented by a conjunction of binary also be extrinsic structural relations, when they describe predicates, of the two types, where arguments can be varithe disposition of objects with regard to objects of a more ables : complex system including them all (for instance Creates the same e ect E as ! Effect ? E (x; y)). (4) 2. functional relations : they describe the roles played by Mj = Relr (no1r ; no2r ) Attt (valt ) t r objects with regard to one another, or with regard with the set to which they belong (for example Causes/Is caused Among these models, some have a prominent role. These by ! Causes=Has ? cause(x;y)). models are points of view. They are particular, coherent models. The role of a point of view is to discriminate, among all 3. correlationnal relations : they describe relations between objects that cannot be more di erentiated, or tothe possible representation elements that may exist in memtally abstract (for example Implies/Is implied by ! ory, those that are important for the task to realize. These Implies=Is ? implied(x; y)). description elements are then said to be pertinent to that 4. assessment relations : they express personal sensations, task. feelings, or ideas (for example Causes the sensation/The sensation is caused by ! Sensation=P ?sensation(x; y)). Entities: So the entities known by the system are concepts, cases, models and prototypes. They can be linked together, and in this case they become nodes in memory. A node can Reasoning also be a predicate argument, such as a constant. The functional architecture of the system in gure 2 is built from a classical case-based architecture. The main di erence Memory Organization in the entities retrieved from the memory at each reasoning The memory is a network of the entities previously presented. isstep here, these entities can be both experimental knowlIt comprises two complementary parts : an experimental mem- edge,: such as cases and concepts, and theoretical knowledge, ory, composed of cases and concepts, and a theoretical mem- such as prototypes and models. ory, composed of prototypes and models. The use of one or the other for the reasoning is not disordered, but closely controlled by these parts of the theoretical (5) memory M = ME MT called the points of view. A particular point of view is

^

^

^

^

[

input data

experimental memory

interpretation theoretical memory

extraction of candidate entities selection of the best candidate

...

success

explanation

Architecture of the system.

associated with each cognitive task realized by the system [7]. For some cognitive tasks, cases are priviliged during the reasoning, and for some other ones, concepts are priviliged. The theoretical entities are used whenever needed to constrain the reasoning.

6. AN APPLICATION EXAMPLE

The application for which this general system has been conceived is a complex real-world application in the domain of eating disorders in psychiatry. The system is a case-based assistant for clinical psychiatry expertise [5]. Clinical expertise is multi-faceted, combining expert diagnosis, treatment planning, patient follow-up and clinical research. All these di erent tasks are performed by the same experts, and closely interact with one another. This is why the system presented here proposes a memory model permitting such an interaction. In this paper, the interaction presented is that of the experimental and theoretical parts of the memory. Experts use both experimental and theoretical knowledge to perform any of their clinical tasks, and the status of knowledge, as pointed before, between experience and theory is not all or none, but rather a continuum.

A Case

...

FirstWeek(case?14; fried?egg; cooked?vegetables;toasted? butter; :::) memory updating / learning

Figure 2.

Hamilton ? depression(case ? 14; 10) ^ EDI (case ? 14; 35) ^ ...

use of the entity results evaluation

...

Apricot ? appreciation(case ? 14; I ? appreciate ? it) ^ Apricot ? avoidance(case ? 14; I ? dont ? avoid ? it) ^

none

proposition of a solution/interpretation based on the entity

failure

Inst(case ? 14; concept ? 28) ^ Inst(case ? 14; concept ? 9) ^ FamilyName(case ? 14; dufour) ^ FirstName(case ? 14; marie) ^

Several types of cases are needed : patients' cases, but also sta cases and control cases. The control cases are cases of non pathological persons, with which to compare the patients for certain questionnaires. An example of a patient's case is the following :

A Concept

Concepts are learnt by the system by conceptual clustering with an algorithm close to that of GBM [20]. So the experimental network is a shared-features network linking cases and concepts. An example of a concept is the following :

Concept ? 28(x) Is ? a(concept ? 28; concept ? 24) ^ Is ? a(concept ? 28; concept ? 11) ^ Abstr(concept ? 28; concept ? 31) ^ Has ? inst(concept ? 28; case ? 14) ^ Has ? inst(concept ? 28; case ? 36) ^ Diagnosis1(x; anorexia ? nervosa) ^ Prawn ? appreciation(x; I ? appreciate ? it) ^ Tripes ? appreciation(x; it ? disgusts ? me) ^ Rillettes ? appreciation(x; it ? disgusts ? me) ^ Dry ? sausage ? appreciation(x; it ? disgusts ? me) ^ French ? beans ? appreciation(x; I ? appreciate ? it) ^ Milk ? chocolate ? avoidance(x; I ? avoid ? it)

A Prototype

Prototypes are either theoretical descriptions of patients from a psychiatric classi cation such as the DSM-III-R [3] (several classi cations are used in the domain), for instance the typical anorexic, or theoretical descriptions from validated studies, such as the prototype of the normal person, gathering all the known means and standard deviations for numerical questionnaires. An example is the prototype of the normal subject :

Hamilton ? depression(x; 31:92; 6:6) ^ EAT (x; 11:1; 6:3) ^ ...

Apricot ? appreciation(x; I ? appreciate ? it) ^ Apricot ? avoidance(x; I ? dont ? avoid ? it) ^ ...

Prototypes are available for each cognitive task. For treatment, for instance, typical, theoretical treatments are described for the di erent diagnostic categories. Moreover, some prototypes gather synthetic information for the di erent diagnostic categories, and so for instance, a prototype of the anorexic is instantiated to the means calculated from all the

anorexic cases in the memory, for all the gures available.

A Model

THEORETICAL KNOWLEDGE Learning

In this application domain, no general model can be conceived, because psychiatry is a complex domain, where no A Clinical Research Reasoning patho-physiological model exists. Psychiatrists don't even know Here again, the words in capital letters are memory entities. whether such a model is conceivable. Nevertheless, particular, The interaction between the di erent entities is also tight. limited, models exist, such as taxonomies of signs and mental disorders. Points of view are also particular models, associ- Dynamically create a POINT OF VIEW from the POINT ated with each cognitive task realized by the system, such OF VIEW as diagnosis, but also with categories of knowledge, such as the four main categories that have proved necessary to the of the task realized, and from the input data system : behavioural, somatic, psychic and biological. An example of a point of view, that of the global eating disorders Extract the CANDIDATE CONCEPTS and CASES the POINT diagnosis point of view, is the following :

Bmi(20) ^ Normal ? increase(5) ^ Maintain ? refuse(5) ^ Anorexia ? denial(5) ^ State ? satisfaction(5) ^ Drive ? for ? thinness(5) ^ ...

OF VIEW

Select the best CANDIDATE CONCEPTS and CASES for the POINT OF VIEW Propose the CONCEPTS and CASES chosen Construct an interpretation of the CONCEPTS using

A Diagnostic Reasoning

In this example, and in the following one, the words in capital letters are memory entities. It shows the close interaction between the theoretical and the experimental entities of the memory.

the THEORETICAL and the EMPIRICAL KNOWLEDGE available Evaluate Explain the failures/successes for the POINT OF VIEW

Select the POINTS OF VIEW in theoretical memory

using both EXPERIMENTAL and THEORETICAL

Extract the CANDIDATE CASES for each

KNOWLEDGE

POINT OF VIEW Select the best CANDIDATE CASES for each POINT OF VIEW Propose one or several diagnoses for each POINT OF VIEW Construct an argumentation between these diagnoses Evaluate, and propose complementary EXAMINATIONS If failure, iterate If no other candidate, look for a PROTOTYPE in theoretical memory for each failing POINT OF VIEW Explain the failures/successes for each POINT OF VIEW using both EXPERIMENTAL and

Learning

7. CONCLUSION

The system presented here is an implementation of a memory model faithful to the ideal of dynamic memory developped by Schank in the theory of dynamic memory [23]. The memory components and the memory organization are similar for the experimental memory (a network of concepts and cases) and for the theoretical memory (a network of models and prototypes). They are expressed in a uni ed knowledge representation language. This permits to the reasoning process to take advantage tightly of both the experimental and the theoretical knowledge of the memory. The perspectives for this work are, among other, to study how the status of some knowledge may change, so that some experimental knowledge becomes theoretical, and vice versa. Another perspective is to model more complex points of view, in particular non coherent one. Points of view can then be contradictory, or redundant, and a set of problems related must be considered.

8. REFERENCES

soning (DARPA), Washington, D.C., R. Bareiss (Edt.), Morgan Kaufmann, San Mateo, California, 1991, 109-120. [1] Aamodt Agnar, \Towards Expert Systems that Learn [15] Golding Andrew R. and Rosenbloom Paul S., \Improvfrom Experience", In : Proceedings of a Workshop on caseing Rule-Based Systems through Case-Based Reasoning", based reasoning (DARPA), Pensacola Beach, Florida, K.J. Proceedings of American Association for Arti cial IntelliHammond (Edt.), Morgan Kaufmann, San Mateo, CA, 1989, gence, 1991, 22-27. 181-187. [16] Hammond Kristian J., \Case-Based Planning : A Frame[2] Aamodt Agnar and Plaza Enric, \Case-based reasoning : work for Planning from Experience", Cognitive Science, 14, Foundational issues, methodological variations, and system 1990, 385-443. approaches", AI Communications, 7(1), March 1994. [17] Kolodner Janet L., Case-Based Reasoning. Morgan Kauf[3] American Psychiatric Association, Diagnostic and stamann Publishers, San Mateo, California, 1993. tistical manual of mental disorders, Third Edition, Revised [18] Kolodner Janet L. and Riesbeck Christopher K. (Edts), (DSM-III-R). Washington DC, 1987. Experience, Memory and Reasoning, Lawrence Erlbaum [4] Bichindaritz Isabelle, \ A case-based reasoning system Associates, Hillsdale, New Jersey, 1986. using a control case-base", In : Proceedings ECAI-94, T. [19] Koton Phyllis, \Reasoning about evidence in causal exCohn (Edt.), 1994, 38-42. planations", In : Proceedings of the American Association [5] Bichindaritz Isabelle, \ A case-based assistant for clinical for Arti cial Intelligence, Saint-Paul, MN, 1988, 256-261. psychiatry expertise", In : Proceedings 18th Symposium on [20] Lebowitz Michael, \Concept Learning in a Rich Input Computer Applications in Medical Care, AMIA, WashingDomain : Generalization-Based Memory", In : Machine ton DC, 1994, 673-677. Learning : An Arti cial Intelligence Approach, Vol 2. R.S. [6] Bichindaritz Isabelle, Apprentissage de concepts dans une Michalski, J.G. Carbonell and T.M. Mitchell. (Edts.), Mormemoire dynamique : raisonnement a partir de cas adaptgan Kaufmann, Los Altos, CA, 1986. able la t^ache cognitive, Thesis of University Rene Descartes, [21] Piaget Jean and Inhelder Barbel, Memoire et intelliParis, 1994. gence. Presses Universitaires de France, Paris, 1968. [7] Bichindaritz Isabelle, \Case-based Reasoning Adaptive [22] Rissland Edwina L. and Skalak David B., \Combining to Di erent Cognitive Tasks", In : Proceedings ICCBR, A. Case-Based and Rule-Based Reasoning : A Heuristic ApAamodt and M. Veloso (Edts.), Springer-Verlag LNCS/LNAI, proach", In : Proceedings of IJCAI-89, Morgan Kaufmann, 1995, in press. San Mateo, California, 1989, 524-530. [8] Bichindaritz Isabelle, \Case-Based Reasoning and Con[23] Schank Roger C., Dynamic memory. A theory of receptual Clustering : For a Co-operative Approach", In : minding and learning in computers and people. Camdridge Proceedings 1st UK CBR Workshop, I. Watson and F. Mahrir University Press, Cambridge, 1982. (Edts.), Springer-Verlag LNCS/LNAI, 1995, in press. [24] Simoudis Evangelos and Miller James S., \The Applica[9] Bareiss Ray, Exemplar-Based Knowledge Acquisition. Acation of CBR to Help Desk Applications", In : Proceedings demic Press, Inc., San Diego, CA, 1989. of a Workshop on case-based reasoning (DARPA), Wash[10] Barletta Ralph and Mark William, \Explanation-Based ington, D.C., R. Bareiss (Edt.), Morgan Kaufmann, San Indexing of Cases", In : Proceedings : Workshop on caseMateo, California, 1991, 25-37. based reasoning (DARPA), Clearwater, Florida, J.L. Kolod- [25] Thayse Andre et al., Approche logique de l'intelligence ner (Edt.), Morgan Kaufmann, San Mateo, California, arti cielle. Tome 1. De la logique classique a la program1988, 50-59. mation logique. Dunod informatique, Paris, 1990. [11] Bhatta Sambasiva R. and Goel Ashok K., \Model-Based Learning of Structural Indices to Design Cases", In : Proceedings of the Workshop of the Thirteenth International Joint Conference on Arti cial Intelligence : Reuse of designs : an interdisciplinary cognitive approach, W. Visser and B. Trousse (Edts.), Chambry, France, 1993, A1-A13. [12] Branting Karl L., \Exploiting the Complementarity of Rules and Precedents with Reciprocity and Fairness", In : Proceedings of a Workshop on case-based reasoning (DARPA), Washington, D.C., R. Bareiss (Edt.), Morgan Kaufmann, San Mateo, California, 1991, 39-50. [13] Goel Ashok K. and Chandrasekaran B., \Use of Device Models in Adaptation of Design Cases", In : Proceedings of a Workshop on case-based reasoning (DARPA), Pensacola Beach, Florida, K.J. Hammond (Edt.), Morgan Kaufmann, San Mateo, CA, 1989, 100-109. [14] Goel Ashok K., Kolodner Janet L., Pearce Michael, Billington Richard and Zimring Craig, \Towards a CaseBased Tool for Aiding Conceptual Design Problem Solving", In : Proceedings of a Workshop on case-based rea-

Suggest Documents