Keywords: CBR, adaptation, re-use, capitalization, knowledge, system, reasoning, case-based.

Abstract: This paper summarizes a new approach of the Cased-based Reasoning. The cases are not stored. The

problem case solution is built as a puzzle. The puzzle obtained corresponds to the required solution. Each

part is carrying information and has an associative behaviour. A piece seeks the piece which can be

associated in width and in depth method. This associative behaviour is determined by several mechanisms:

engine of expert system to binary rules, model of multicriterion choice of ordinal outclassing, search for

close indices. A puzzle can thus have a complex mode of reasoning; each piece has a specific behaviour.

The tool was tested on two applications of decision-making aid: identification of malaria facies and

assistance to the specification of habitats. These applications made it possible to check the interest of this

original framework. In particular it brings an elegant solution to the phase of adaptation in CBR technique.

1 INTRODUCTION

Research in the field of the C.B.R. (Case-Based

Reasoning) is developed by Kolodner (Kolodner,

1988) then by Richter (Richter et al., 1993) and

Veloso (Veloso et al., 1995). Paradigm C.B.R. is

based on the model of Aamodt (Aamodt et al., 1994)

represented by figure 1. The best of the knowledge

is given by Watson (Watson, 1997). The C.B.R. is

characterized by a adaptation step of an existing

solution. It is allowed that the intervention of the

user is necessary. The authors come up against a

difficulty of formalizing this step in a generic way.

Many authors resume works of Carbonell

(Carbonell, 1986) on the analogy. The process of

adaptation were exposed by Voss (Voss, 1997) and

(Wilke et al., 1998). But the organization of

knowledge in hierarchy (Kayser, 1997) opened new

ways (Lieber, 2002). Let us examine the best of the

knowledge of the re-use and the adaptation in the

case-based reasoning systems.

2 ADAPTATION AND RE-USE

THE BEST OF THE

KNOWLEDGE

From the first research work on the CBR, a lot of

proposals were about the problem of adaptability

[Hinrichs 89][Turner 89]. Kolodner has proposed a

classification of the types of adaptation [Kolodner

93]. He distinguishes the differences of methods and

the differences of search for adaptation rules.

In 1996, Voss [Voss 96] adds to the cases

identification criterion, the characteristics of the case

type, with an estimation of the adaptation. [Purvis

and al] suggest to exploit the constraints to guide the

adaptation. This approach wants to assure the

obtaining of consistent solutions. This implies that

whole of constraints is complete and correct.

The Trinety College’s team of Dublin has studied

the adaptation for the complexes systems [Smyth

and al 93]. They are proposing in particular to learn

rules of adaptation from a base of cases [Hanney and

al 97].[Veloso 97] describes a mechanism which

adapts the parts of case with a part of other cases.

An interesting contribution has been brought by

[Hanks and al 95] with an algorithm of plan based

on cases and independent of the field.

[Leake and al 97 b] have imagined a process of

adaptation with three types of learning: case learning

by creation of responses plans, learning by

indexation of case according to their use, case

learning by search for similarity about the adaptation

cost. But we notice that the improvement of abilities

of adaptation

and re-use go through new

organizations of case-based reasoning systems

[Kayer 97] [Lieber 2002].

For this reason the model P is proposed, with a new

organization of knowledge.

MODEL P : AN APPROAC

H OF THE ADAPTABILIT

OF CASE-BASED REASONING SYSTEMS

Mathilde Billy, François-Xavier Magaud, Claude Petit and Laurent Combasson

CESH - LASS, Université Claude Bernard Lyon 1

357

Billy M., Magaud F., Petit C. and Combasson L. (2004).

MODEL P: AN APPROACH OF THE ADAPTABILITY OF CASE-BASED REASONING SYSTEMS.

In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 357-362

DOI: 10.5220/0002616103570362

 SciTePress

In many applications, the base of knowledge can be

divided into several subsets, which are common with

several cases. It is interesting to store these

fragments not in several distinct cases, but in the

form of pieces being able to be assembled. As for a

puzzle, a case is reconstituted starting from pieces.

The idea is not any more to store cases but pieces,

which will be assembled and will form a new case :

a puzzle.

3 THE CONCEPT OF PUZZLE

A puzzle is a progressive assemblage of pieces. For

each problem, the user provides information to

reconstitute a case (puzzle) which will bring a

solution to the problem. Pieces, and their possibility

of bringing together, will have been defined

beforehand by an expert. The process, which leads

to the formation of the puzzle-solution, is therefore

directed by the choices of this expert. Its work

establishes implicit bonds between the pieces, which

guarantee to the user not to omit anything in its

search for solution.

4 THE BASIC ENTITY: THE

PIECE

A piece is a cellular object, composed of two

information types:

- A descriptive information, more or less complex,

on a precise subject, which can for example be

materialized, in the simplest cases, by an image or a

page

HTML. That corresponds to a piece of puzzle.

- Information of pairing, guiding the choice of the

pieces which will be connected to it. This

information allows an automatic aggregation of the

pieces, to form the puzzle. That corresponds to

contours of a piece of puzzle. The pieces do not

behave like passive objects. They have their own

technique of selection of the successors and, as we

will see it later, are able to adapt their search

according to already selected pieces. The pieces can

thus be compared with intelligent agents, holding

account of their environment and semi-autonomous

(because to user is sometimes solicited).

Here are two examples, on the topic of malaria.

4.1 Example 1: the piece "Facies

river"

4.1.1 Descriptive information, in the shape

of a html page

The facies River

The bank of the rivers are very touched zones.

Indeed, many water holes are formed along

the rivers: if these rivers are in withdrawal

(like most of the time), the conditions are

ideal for the development of the anopheles.

Here are some photos:

River in pre-forest

zone (Cameroon)

Savane river

The zones at the banks of the rivers are high-

risk zones of transmission of malaria in any

season.

4.1.2 Pairing of the pieces

The pieces "Facies river" uses the method of

indexation. This method, which uses the key words

"river" and "malaria with permanent transmission",

From case

extract pieces

Figure 1: Principle of CRB Technique

Figure 2: Principle of Modèle P

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

358

will be presented in the section “Techniques of

pairing of the successors”, as for all the terms in

italic. This piece belongs to the field "geo-climatic

zone".

5 CONSTITUTION OF A PUZZLE

The aim is to bring, starting from the pieces of the

base, a solution to the problem of the user.

The used

algorithm is recursive and built a puzzle: starting

from the choice of the first piece, the relations with

the not yet visited pieces of each piece are explored.

The mechanism builds a tree with an in-depth

exploration initially. The target piece represents the

set of information provided by the user to initialise

the search. It does not have descriptive information,

and thus did not appear in the final puzzle.

The initialisation of the search for the first piece

returns to the user. He has the choice of the

technique of pairing of the successors, as well as

filtering. According to the technique chosen, he can

be asked to give more precise criteria. He defines the

first piece of the puzzle, called target piece.

The search is done in a logical order : either from

most general to more detailed, or in an order

imposed by the nature of the problem. The choice of

the possible successors based on two processes:

- Filtering : the choice is reduced at pieces having a

common criterion.

- Filtering by fields: The pieces belong to a field,

represented by a name. They also have a field of

call, in which the search for the possible successors

is carried out.

- Filtering by dates: the base evolves with time and

new information, therefore certain pieces can

become null and void, or less adapted to the

situation. Each piece having a creation date, the

choice is restricted at pieces former/posterior on a

selected date.

- Techniques of pairing : In a field, the piece-

successors are chosen either by asking the user

for the contest, or by seeking the closest piece,

according to criteria defined by the expert:

- Techniques by interrogation: selection on

photo/description/HTML page, expert system, etc...

- Techniques of the nearest neighbour : indexation,

multicriterion distance from outclassing, etc ...

Each piece indicates, according to the technique of

pairing associated, a set of successors, which will

seek another set of successors. This tree structure

establishes relations which bring closer the

information. Let us notice that it is impossible to

take again a piece already selected under penalty of

creating a circuit.

The combination of these two techniques of pairing

makes easier the work of the expert : he can divide

the general problem and classify the pieces in field,

according to the technique of search.

This mixed method has an interest: the user is

guided, directed automatically to certain tracks by

the expert, and can redirect his search, with his

initiative, using the techniques by interrogation.

Termination of the algorithm: There are two

possibilities:

- The stock of pieces is exhausted

- All the started branches are finished, ie for each

branch a sheet has been reached. Indeed, certain

pieces can be defined as sheets by the expert. The

pieces-sheets are those which do not search for a

successor.

Associative bonds of the pieces: It is interesting to

bind pieces whose association is necessary to the

good understanding of the problem. These bonds are

not commutative. Two associated pieces will thus be

called in different contexts, but the selection of a

piece will automatically involve the selection of its

bonded piece.

Example from the application to the facies of

malaria : In the struggle against malaria, the

countries are classified in 3 types, corresponding to

the recommended treatment. Burkina Faso pertains

to group 2.

The piece Burkina Faso is related to the piece

"country of group 2" (the choice of the piece

"Burkina Faso" involves the automatic selection of

the piece "country of type 2" as a successor).

This method makes it possible to choose at each

stage the most adapted information, and provides in

the end an answer "made to measure" to the

question. The adaptation to the problem was made in

a generic way, progressively. This approach which

builds a search for solution as a puzzle is a great

improvement of the stage of adaptation.

6 THE TECHNIQUES OF PAIRING

OF THE SUCCESSORS

This chapter is not the best of the knowledge of

these techniques. It presents the mechanisms

implemented and tested in two applications. Two

robust and innovating techniques are proposed: a

multicriterion search and a reasoning with rules of

production.

MODEL P : AN APPROACH OF THE ADAPTABILITY OF CASE-BASED REASONING SYSTEMS

359

6.1 The nearest neighbours

techniques

6.1.1 Indexation

Key words can be associated with the pieces, such as

descriptors of the situation represented by the piece

or the topics in connection with his subject.

The very close piece will have common key words,

and more numerous they will be, closer the pieces

will be, in the semantic network formed by the key

words.

Indexation consists in listing the key words of the

target piece, then to search for the pieces comprising

the most key words of this list. Let us examine the

following example :

The piece C calls pieces of the D1 field. It has the

key words "i1", "i3" and "i7". The pieces of the D1

field are the pieces A, B, C, D, E, F

The pieces whose key words belong to the preceding

list are: A ("i1" and "i3"), B ("i1") and

D ("i1"). If

the piece A is not taken yet, it will be chosen as a

successor of the piece C. If not, the pieces B and D

are retained jointly, with an equal number of

common key words. (If they are not available any

more, the piece C does not have any successors).

When there is filtering by field, if no piece of the

field have common key words with the piece

considered, it is possible for the user to widen the

search to all the pieces of the base. This possibility

also exists for multicriterion search.

6.1.2 Multicriterion search

This technique of pairing corresponds to certain

problems. The idea is as follows: Several

combinations of criteria can lead to the same choice,

with the same solution. Several criteria can be

associated to the pieces, which allow to make a

multicriterion calculation of distance with ordinal

outclassing. This outclassing avoids the circuits.

Therefore the mechanism is robust. Two pieces will

be close if they obtain a nearby total score.

Let us explain on an example: The piece B calls

pieces of the D1 field, i.e. A, B, C, D, E, F. The

criteria of the piece B are c1 (value=6), c2

(value=3)et c3 (value=1). The pieces having these

same criteria are D (c1=2, c2=7, c3=4) and F (c1=6,

c2=4, c3=1).

The following table is drawn up:

The outclassing number for B is : 1 on c1 + 0 on c2

+ 0 on c3 =1

Score of outclassing of each piece :

B : 1

D : 4

F : 2

F is the piece to which the score is closest to B.

Therefore F is B's successor. The tool also proposes

to the user to spread the selection to the 2 or the 3

closest pieces.

That obviously requires to compare pieces having

exactly the same criteria. As in the case of

indexation, the first stage consist in finding the

pieces having all the criteria of the target piece, then

calculation is carried out.

This method can be applied to non-numerical

criteria, but with an order. Example : code for

colours.

Dark blue=1, light blue=2, green=3, yellow=4,

orange=5, red=6, …

6.2 Techniques by interrogation

6.2.1 Expert system with rules of production

The engine of the expert system exploits a base of

binary rules, with a no monotonous reasoning. It can

use traditional rules of production which are

translated into binary rules or directly binary rules.

The absence of combinatory explosion in the engine

allows a great robustness of operation. It always

obtains a result and in a very fast time. The expert

provides the rules and can associate piece-successors

to the conclusions.

The user thus will traverse the decision tree, while

answering the questions until he reaches a sheet,

itself connected at a piece-successor. The presence

of isolated rules (under trees) can involve the

selection of several pieces-successors.

C1 C2 C3

B 6 3 1

D 2 7 4

F 6 4 1

Figure 3: Transaction of the production rules into

decision tree

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

360

6.2.2 Manual selection

This mode of selection consists in choosing among

specifications the successors of the target part. The

choice of the user is guided by descriptive

information of the pieces: photographs and/or

HTML page.

The expert can decide to declare a piece without

successor. It can represent independent information,

without possible continuation.

7 THE MODEL P IN ACTION

The application starts with a manual selection on the

D1 field. The pieces proposed are A, B, C, D, E, F.

A and E are chosen.

Study of the part E: The piece E is a sheet. However,

this piece being associated at piece P, this one is

automatically added in the solution tree. The piece P

is a sheet.

Study of piece a: The technique of piece A is

indexation, on the D1 field, with the key words "i1",

"i2" and "13". The pieces successor are B and C,

with two common key words ("i1 "and" i2 "for B,"

i1 "and" i3 "for C).

Study of piece b: This piece calls the pieces of the

D1 field with the multicriterion technique.

Calculation is carried out on the criteria c1, c2, c3,

and

the closest pieces are D and F (scores:

score(B)=2, score(D)=1, score(F)=1).

Study of the piece D: This piece requires a manual

selection on the D2 field, whose only piece is G. G

is selected. G is a sheet.

Study of the piece F: This piece is an example of

expert system.

What is the value of fact 1? answer = yes (to fact 2)

What is the value of fact 2? answer = no (to fact 4)

What is the value of fact 4? answer = yes

(conclusion 2)

conclusion 2 is connected at piece N, which is a

sheet.

Study of the piece C: The piece C uses a indexation

on the D1 field. All the pieces of this field have

already been selected. The user is asked if the field

of search to have to be widen. After confirmation,

search fails, because no piece have a common key

words with those of C. C becomes a sheet.

A response or a different choice for a piece which

has a technique of questioning can make trees of

rather distinct form and contents. This shows the

many possibilities of construction of the model. But,

an already selected part cannot be selected twice.

8 CONCLUSION

This paper presents a new approach of the

architecture of CBR tools. It wants to contribute a

share to the problem of adaptability in techniques

CBR. The mechanism proposed is based on a

postulate: the cases of a field are decomposable

(entities, sub problems, processes, diagrams, ...) and

a component can be divided into one or more other

cases. Only the components of a field are preserved.

The case solution is built automatically as for a

puzzle. Each part of the puzzle brings an element of

the solution and associates the part in width and in-

depth. A part has an information part and an

associative behaviour. The mechanism is recursive.

The depth of the puzzle is not limited. Several

models of reasoning were implemented: engine with

binary rules of production, indexing, multicriterion

search of a case. In the same puzzle, several types of

reasoning can cohabit. On two applications

(detection of facies of malaria and identification of

habitat), it showed its great flexibility of adaptation

with a context. A greater effectiveness is obtained.

The user does not seek a case among a multitude of

case but reconstitutes the nearest case with several

possible reasoning. The updates relates to the stored

parts, subsets of the puzzle, and the parameter

setting of the models of reasoning.

Two applications have been developed. They permit

to build up the know-how and to exploit it. Once

concern the admission of the facies of the malaria.

The other touch on the architectural expert or

buildings. They have permitted to confirm the new

concepts suggested.

REFERENCES

Aamodt, A. and Plaza, E. 1994. Case based reasoning

: Foundational issues, methodological variations,

and system approaches. AICOM, vol 1, pp 39-59.

Carbonell, J.G. 1986. Derivational analogy : A theory

of reconstructive problem solving and expertise

acquisition. In Readings in Knowledge Acquisition

and Learning, Morgan-Kaufmann.

Hanks S. et Weld D. S. – A domain independatn

algorithm for plan adaptation. Journal of Artificial

Intelligence Research, n°2, 1995, pp. 319-360.

Hanney K. et Keane M. T. – The adaptation

knowledge bottleneck : how to ease it by learning

from cases. Second International Conference on

Case-Based Reasoning, ICCBR 1997, éd. par

Leake D. et Plaza E. pp. 359-370. – Springer-

Verlag Berlin, Germany, 1997.

MODEL P : AN APPROACH OF THE ADAPTABILITY OF CASE-BASED REASONING SYSTEMS

361

Hinrichs,T.R. 1989.Strategies for adaptation and

recovery in design problem solver. Workshop on

case-based Reasoning, DARPA 89. pp. 115-118. –

Morgan-Kaufmann, San Mateo, 1989.

Kayser, D. 1997. La représentation des

connaissances. Hermes, Paris.

Kolodner, J.L. 1988. Workshop on case-based

Reasoning, DARPA 88. Morgan-Kaufmann.

Kolodner, J. L. 1993. – Case-based Reasoning. –

Morgan Kaufmann Publishers, 1993.

Leake D. B. , Kinley A. et Wilson D. – Learning to

integrate multiple knowledge sources for sase-

based reasoning. Proceedings of the 15

International Joint Conference on Artificial

Intelligence. – Morgan Kaufmann, 1997.

Lieber, J. 2002. Strong, Fuzzy and Smooth

Hierarchical Classification for Case-Based

Problem Solving. Proceedings of European

Conference on Artificial Intelligence, pp 81-85.

Purvis L. et PU P. – Adaptation Using Constraints

Satisfaction Techniques. Case-Based Reasoning

Reasearch and Development. Proceedings of the

First International Conference on Case-Based

Reasoning- ICCBR-95, éd. Par Veloso M. et

Aamodt A. pp. 289-300. – Sesimbra, Portugal, 23-

26 octobre 1995.

Richter, M.M. Wess, S. Althoff, K.D. and Maurer, F.

1993. First European Workshop on Case-based

Reasoning. University of Kayserslautern,

Germany, Lecture Notes in Artificial Intelligence,

volume 837, Springer Verlag.

Smyth B. et Keane M.T. – Remembering to forget a

competence-preserving case delection policy for

case-based reasoning systems. IJCAI-95.

Proceedings of the Fourteenth International Joint

Conference on Artificial Intelligence, éd. par

Mellish C. pp. 377-382. – Morgan Kaufmann

Publishers San Mateo, CA, USA, 1995.

Turner,R. M. 1989. Case-based and schema-based

reasoning for problem solving. Workshop on case-

based Reasoning, DARPA 89. pp. 341-344. –

Morgan-Kaufmann, San Mateo, 1989.

Veloso, M. and Aamodt, A. 1995. Route planning by

analogy, First International Conference on Case-

Based Reasoning, ICCBR'95. Sesimbra, Portugal,

Springer-Verlag.

Veloso, M. – Merge strategies for multiple case plan

replay. Second International Conference on Case-

Based Reasoning, ICCBR 1997, éd. par Leake D.

et Plaza E. pp. 413-424. –Springer-Verlag, Berlin,

Germany, 1997.

Voss, A. 1996. Structural Adaptation with TOPO.

Workshop on Adaptation in Case-Based

Reasoning, ECAI-96, éd. Par Voss A. , Bergmann

R. et Bartsch-Spörl B. – Budapest, Hungary,

August, 1996.

Voss, A. 1997. Case Reusing Systems – Survey,

Framework and Guidelines. Knowledge

Engineering Review, vol. 12, n°1.

Watson, I. 1997. Applying Case-Based Reasoning :

Techniques for enterprise Systems. Morgan-

Kaufmann.

Wilke, W. and Bergmann, R. 1998. Techniques and

knowledge used for adaptation during case-based

problem solving. IEA'98.

Model P framework

phase 1

: identification of the possible piece for a

puzzle (experts : Delphi method...)

phase 2

: Construction mechanism of the puzzle and

resolution of the problem

complete puzzle

solution of the

problem

search for the first piece

with technical choices of

preliminary pairing

PROBLEM

information concerning the

problem (puzzle target)

There is not possible

pairing

search for the following piece

with mechanism of pairing

associated with the preceding

piece

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

362