A MULTI-AGENT ARCHITECTURE FOR

MOBILE SELF-TRAINING

M. Ennaji*, H. Boukachour*, P. Gravé**

Laboratoire d’Informatique du Havre*, CIRTAI-NTIC**

25, rue Ph. Lebon

76058 Le Havre Cedex, France

Keywords: Multi-agent and multi-layer system, Case-Based Reasoning, teaching agent, Semantic features.

Abstract: This article is the result of an interdisciplinary meeting between sociologists and didacticiens on the one

hand and data processing specialists on the other hand. To develop the theoretical and methodological

principles of the design of a training environment, by putting the needs and the difficulties of the student at

the center of the design process and data-processing modeling, constitutes the common action of these two

research laboratories within the framework of this collaboration. To design a virtual tutor called “teaching

agent” in a system of remote formation implies the implementation of a flexible and adaptive system. We

propose an multi-agent multi-layer architecture able to initiate the training and to manage a teaching and an

individualized follow-up.

1 INTRODUCTION

A Computer Environment of Human Learning

(EIAH) is a computer system which has for

objective to favor the learning of a domain of

knowledge by a learning. The computer systems of

the learning assistant are traditionally structured

around an only educational module: an artificial

tutor. It possesses a domain expertise of knowledge

and applies a strategy of education to interact with a

student to help him to resolve a given problem. This

principle of functioning in autonomy of the couple

student-tutor can be satisfying until the moment

when the system reaches its limits; the presence of a

human teacher, even another student becomes then

essential. So, these systems can be used as a

supplement to the traditional teaching, in a class for

example. However, with the evolution of the

networks of communication such the Internet and

the services associated as the information servers

like Web type for example, the teaching and the

situations of learning move from the institutional

frame to the room of lessons towards the place of

residence, the company, etc.

So, it is now necessary to design EIAH which

take into account the mobility of the students, as to

ensure them an individualized follow-up to respect

their rhythm of learning and put in their arrangement

the human presence among all the accessible

educational resources.

This work is the result of an interdisciplinary

meeting between sociologists and didacticiens on

one hand and computer science specialists on the

other hand. It articulates around the pooling of the

skills of two research laboratories of the university

of the harbour: the Computer Science Laboratory of

Le Havre (LIH), and more particularly the research

group ARM (agents and major risk) and the

laboratory CIRTAI-NTIC which has experimented

some tools of learning and noted their limits. Our

contribution, in the field of the distance learning,

consists in designing and in realizing a computer

system able to introduce the learning and the

managing an individualized teaching and follow-up.

This article develops the multi-agent multi-layer

architecture resulting from the work of the computer

science specialists on the systems of decision

making for dynamic situations, which is adapted to

the conception and to the realization of an

"intelligent" virtual tutor also called "teaching agent

" in mobile learning.

2 TEACHING AGENT

The learning situation based on computer is thought

like a Human-Machine system. Classically it

343

Ennaji M., Boukachour H. and Gravé P. (2006).

A MULTI-AGENT ARCHITECTURE FOR MOBILE SELF-TRAINING.

In Proceedings of the Eighth International Conference on Enterprise Information Systems - AIDSS, pages 343-350

DOI: 10.5220/0002461003430350

 SciTePress

consists in examining the relations between different

components: teacher, learner, learning object and

computer. Sometimes it also includes a reflexion on

these relations and the institutional environment.

These learning system are focused on content but do

not attach importance to the career of acquiring

content. These learning systems generally rest on a

transmissive model rather than learning model. It

often reduces the e-learning to a learning activity

sequential organisation. So the interactivity of the

learner is limited by using simple navigation based-

function with the tool.

The working party NTIC of the CIRTAI

(ANNOOT et al. 2004) (BERTIN, 2004) (BERTIN

et GRAVÉ, 2004a) (BERTIN et GRAVÉ, 2004b)

shows, for the distant learning system, the specific

interactions between different components of an

ergonomic model based on a constructivist approach

(VYGOTSK, 1978). New interactions add to the

usually recognized learning interactions (teacher,

learner, learning object), because of the introduction

of the distance. Furthermore we have to add the

necessity of transferring in a virtual environment

interaction whose observation in a presential

learning have revealed the importance, particularly

peers relations, pedagogic tutoring and follow-up.

About this pedagogic tutoring, we enrich our

interactionist and ergonomic model (outline 1) by

addition of dimensions that allows the development

of learner cognitive and metacognitive abilities: this

main line research moves towards the conception of

“teaching agent”. We lean on the concept of

teaching agent, described by Philip Hubbard

(Hubbard, 1999), (Hubbard, 2000). It is an

informatic entity which, by its graphic, its

conception and its dynamic and well-timed

apparition mode, plays the virtual tutor role.

Together with Hubbard, such an agent has to

present certain characteristics:

• A physical presence and a personality,

• An expertise in the reference field,

• An aptitude for individualised learning,

• An ability to initiate learning.

The following outline situates the teaching agent

in the distant learning mediatized systems.

By its appearance and its operating mode, the

teaching agent is a sort of technological mediation of

human presence which might allow different uses:

• Aid in using activities and software;

• Methodological advices for learning better at

distance;

• Selective aid (dictionary, encyclopaedia…);

• Aid production and redaction (thesaurus for

language for example);

• Supply additional references;

• Watch learning operation in the background

(follow-up functions).

At the same time teaching specialist (the

pedagogue), expert and tutor (or companion), the

teaching agent is receptive at any time, without any

evaluative connotation unlike the teacher. “The

pedagogue in classical times was the slave who

escorted the children to school, uneducated slave

whose main task was to serve as bodyguard but who

could also help the learner in his homework, answer

Figure 1: Didactic ergonomic interactionist model (Berlin, 2001, 2004).

ICEIS 2006 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

344

questions, play games, or even give tests. But the

pedagogue never initiates: he comes forward when

summoned and, when the learner has had enough,

he goes back to his place. This is the role which

seems in some ways a natural one for the computer

to assume. After all, computers were built in the first

place to answer questions, not to ask them”

(HIGGINS, 1986). Its role is essentially formative.

This is a less explored domain of research for two

main reasons:

• The difficulty to formalize the theoretical nature of

a computing agent, capable giving the learner this

“cognitive scaffolding” (VIGOTSKI,1978),

• The problems for developing an appropriate

technology to these needs. We lean on an

experiment, carried out at the University of Le

Havre (France), with students in Economics and

Social Administration Bachelor’s Degree, who

received instruction in Organisation Sociology, in

two modes: part time in traditional presential

learning with the teacher, part time in self-training

situation with a CD-ROM. The same contents

were given. At the end of the training, students

filled a questionnaire about their perceptions of the

two pedagogic modes (processing of this data is

still in progress). As far as we are concerned, this

CD-ROM automatically generates follow-up files,

which memorize the student’s individual paths.

career during it self-training. The analysis of this

data files help understand the student’s path and

examine the pertinence to add a teaching agent or a

computing system, that can ensure these pedagogic

tutoring functions.

3 OUR APPROACH: DESIGN

AND IMPLEMENTATION

3.1 Introduction

There are numerous research works as well as

teaching practices coming from the teaching and the

computer experiments led in mobile learning and,

more particularly, in online training. These works

are interested in the conception and the

implementation of computer systems allowing to

attend the student in his training, BAGHERA

(Webber et Pesty, 2002), SIGFAD (Mbala and al,

2003).

The current platforms of e-learning do not have

tools allowing to make an individualized follow-up

of the student. This follow-up is essential because of

the great number of students which do not finish the

training. They use a more technical approach, by

proceeding to a simple transposition of a traditional

pedagogy in a computer portal, than a didactics one,

by integrating the contributions of the most recent

theories without limiting itself to the structures of

behaviourist kind.

In the framework of a self-training, the student must

have the freedom and the choice of the course to

follow in order to assimilate the object of the

learning by a building approach of the knowledge

instead of a massive use of documents. This freedom

must be assisted by a regular and continuous follow-

up. Traditionally, the follow-up is insured by a

teacher. The main objective of our training

environment, opened and distant (FOAD) is to

propose an alternative in the traditional teaching.

Consequently, the computer tool has to integrate the

assistance and the follow-up among its features.

This expected computer tool is an opened,

evolutionary system and a "translation" of the

training agent, described in the interaction model of

didactic ergonomics according to (Bertin, on 2001)

(Figure 1). To meet the needs of didacticiens, it

consists in conceiving and realizing a computer

system which has to take into account the following

characteristics:

Opened: the structure can change dynamically.

Evolutionary: because the components of such a

system are not known in advance, change during

time, and are essentially heterogeneous.

Autonomous: such a system has to take the role of

the teacher and to may be able to initiate the

learning.

Adaptive: the system is intended for an

individualized learning, so it is needed to take into

account the various profiles of the students.

The system is complex. This complexity, which

is also translated by the important number of data to

be treated and the dynamics of the situation to be

treated, led to us to choose the multi-agent systems

(MAS) as the modelling (Jennings and al., on 1998).

4 ANALYSING AND

DECOMPOSITION OF THE

NEEDED

4.1 Introduction

We are making closer the questioning on the

analysis of the career of the students in self-training,

and the decision system support which have to allow

to represent, to follow and to analyze the evolution

A MULTI-AGENT ARCHITECTURE FOR MOBILE SELF-TRAINING

345

of a dynamic situation. Such a system allows to

represent the observed situation but also its

evaluation.

Evaluating the situation can be performed by

calculating its possible consequences. This can be

carried out using previous situations whose

consequences are known. So, a reasoning based on

analogy can be used relying on the following

hypothesis: if a A situation looks like a B

situation, the consequences of the A situation

ought to be similar to those of the B situation.

The Case Based Reasoning (CBR) (Kolodner,

1993) is a methodology of resolution of problems

leaning on the re-use of the experiences spent to

resolve new problems. The decision making system

is one of the most promising domains of application

of the CBR. It allows to put in synergy the capacity

of resolution of problem of the man with the

capacity of the computer system. Memory of the one

and the other one strengthen mutually to participate

in the resolution of the problem.

In the framework of a mobile self training

offering the individualized follow-up of students, a

decision making system, allows to analyze the

course and the work of learning it to anticipate a

possible finishing of learning it or a "bad" learning

of this one.

The system which we propose, has to take into

account moreover, the evolutionary character and

the dynamics of the course to be analyzed. The

analysis is supported by the link which the system is

going to make, in a continuous way, between the

way of learning it and the past career .

The past tracks are described by scenarios

grouped together in a base called "base of

scenarios". They characterize, for every past career ,

all the determining aspects in its progress. We call

here determining aspect, a fact which played a

current role in the career the events have taken

place. So, every scenario contains a temporal list of

semantic features associated to the important aspects

of the careers.

The analysis of the career s has to be made in a

continuous career , in which we must use a multi-

agent architecture allowing the implementation of a

dynamic and incremental case reasoning for the

situations evaluation. This architecture allows the

real-time comparison of the situation observed with

past situations stored in a base of case.

4.2 Existing Works

In parallel, in this collaboration between computer

specialists and didacticiens, the areas of research of

the computer specialists of the LIH working on the

distance learning, concern the systems of decision

making that must estimate a dynamic situation.

Numerous applications of computer systems must

allow to represent an evolving situation in order to

be able to analyse it. This problem exists in different

application domains such as road traffic,

meteorology, risks management, etc. The

implementation of these systems requires to answer

to the following questions:

Which tools to model the situation?

Which tools to model and manage time?

Which architecture to manage the monitoring of the

evolution of the situation?

How to allow the users to have a clear and

understandable view of the state of the situation and

of its possible evolution?

The observed situation generally contains a great

number of dynamic parameters, that is to say

parameters whose value change over time. Systems

allowing the management of such situations must be

dynamic in order to be able to handle these

evolutions. As a consequence, to design these

systems, a flexible and adaptive architecture is

needed. This led us to choose a multi-agent

architecture(Jennings and al ., on 1998). We are thus

interested in the development of multi-agent systems

dedicated to the modelization and the evolution

forecast of dynamic situations.

Such a system must not only allow to represent

the observed situation, but also has to allow its

evaluation. Evaluating the situation can be

performed by calculating its possible consequences.

This can be carried out using previous situations

whose consequences are known. So, a reasoning

based on analogy can be used relying on the

following hypothesis: if a A situation looks like a B

situation, the consequences of the A situation ought

to be similar to those of the B situation. To perform

such a reasoning, we must elaborate:

• a multi-agent CBR. The representation of the

current situation is, in our context, based on a set

of agents.

• a dynamic CBR. The target case of the CBR

process is an evolving situation, so the CBR has to

take this evolution into account incrementally. In

other words, when the situation changes, it must not

be considered as a new target case.

Using such a dynamic multi-agent CBR, the

aim

of the system is to select as soon as possible the

cases of the base which seem to be the most similar

ICEIS 2006 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

346

to the current situation in order to be able to

anticipate its consequences. Of course, this selection

must adapted to the evolution of the situation over

time. Indeed, new information on the situation can

eventually modify the set of cases which have been

selected during the previous steps.

In the following figure (see figure 2), we present a

synthesis of the similarities and differences between

our approach and the CBR for dynamic situations.

More information can be found in (Simon and

Boukachour, 2004).

Standard CBR dynamique CBR Our approach

Target Case = set

of attributes

Target Case=

temporal description

Target Case =

temporal description

Static elaboration Static elaboration

continuous and

multi-agent

elaboration step

Indexation None None

Static recall step Static recall step

Multi-agent and

continuous recall

step

Adaptation Adaptation make by expert

Learnning Learnning Learnning

Figure 2: Similarities and differences between our

approach and the CBR for dynamic situations.

5 OUR PROPOSAL: PRINCIPLES

AND IMPLEMENTATION

5.1 Principles: The Different Kinds of

Agents Used

Our architecture is based on a multi-agent

architecture as proposed by Marcenac in (Marcenac,

1997). This kind of architecture uses several

hierarchical agent layers, a layer of the level n

having a view on the layer of the level n-1. Our

system use three different layers (see figure 3) :

the lowest one : it contains the agents allowing to

model the current state of the situation, that is to say

the informational agents,

the intermediate one : it contains synthesis agents

used to analyse the previous layer,

the highest one : it contains prediction agents which

must provide information about the potential

evolution of the situation using dynamic CBR

techniques.

5.2 Factuals Agents

First of all, the observed situation is modelled by a

set of "factuals" agents. This set of agents receives

pieces of information about the situation which are

sent to the system by actors or by distributed data

bases. Each factual agent is supposed to represent

one of these pieces of information which is called

“semantic features (SF)” (Jackendoff, 1993)

(Denhière and Baudet, 1992).. A SF is a three-part-

relation <object, qualification, value> representing a

partial aspect of the situation. A SF is also the

atomic data structure, i.e. the smallest piece of

information the system could deal with. (for more

details, see (Person and al., 2005) (Boukachour and

al., 2002). The main advantage to use agents to

represent information about the current situation is

that it allows to obtain a flexible representation

which can be easily adapted as the situation evolves

(ie as new pieces of information about the situation

are introduced in the system). Each agent must also

provide a temporal validity measure allowing to

evaluate the "freshness" of the piece of information

associated to its semantic features.

Each informational agent must provide numerical

measures of its evolution over time. More precisely,

these measures must allow to evaluate the level of

reinforcement of the agent inside the organisation it

belongs to. Indeed, it is supposed that the more an

agent is reinforced, the more its semantic features

must be taken into account in the evaluation of the

situation. This reinforcement must be based on a

similarity measure between items which can use

semantic, temporal and spatial aspects (Pesron and

al., 2005). These mechanisms allow to take into

account the fact that, for example, a piece of

information introduced very early in the system can

turn out to be non relevant later. On the contrary,

some can be given later to the system and finally be

judged as very representative of the current state of

the situation. More information can be found in

(Person and al., 2005).

5.3 Synthesis Agents

The goal of these agents is to provide a synthetic

view of the global behaviour of the factuals agents

layer in order to facilitate the comparison with past

situations stored in the scenarios base. This layer

helps to implement the standard target case

elaboration step of the case-based reasoning cycle.

This elaboration is, however, specific because of its

dynamic property.

A MULTI-AGENT ARCHITECTURE FOR MOBILE SELF-TRAINING

347

More precisely, the goal is to classify factuals agents

into groups. This operation can be done using

reinforcement and temporal validity measures

provided by factuals agents. Indeed, if one consider

a particular factual agent, the values of its measures

are not very significant. On the contrary, if these

measures are compared with those of the other

factuals agents, it allows to build groups of agents

with similar measures values. Theses groups can be

representative of important aspects of the current

situation which will be used by prediction agents to

manage the comparison with past situations.

The goal of synthesis agents is to dynamically

build these groups called clusters. In (Coma and al.,

2003), we propose a dynamic techniques for agents

clustering. Each cluster is modified over time

according to factuals agents evolution. For example,

it can increase if new informational agents seem to

be similar (from the measures point of view) to those

belonging to it. On the contrary, it can decrease, or

even disappear, if too few factuals agents belong to

it.

5.4 Prediction Agents

In order to be able to use CBR techniques, the

system must contain cases describing past situations.

Such cases are called "scenarios". These scenarios

must allow to characterize, for each past situation,

the set of decisive factors which seem to be related

to the career the situation went on. As a

consequence, each scenario contains a list of

semantic features associated to the decisive factors

of the past situation. This list can, eventually, be

organized temporally. These factors can be found

using experience feedback provided by domain

experts.

A prediction agent is associated to each scenario

stored in the system. The goal of the prediction

agent is to compare the course of the current

situation represented by the factuals agents with the

one described in the scenario. This comparison,

which must be made in real time, consists in

determining if the factors which seem to be

important in the current situation are similar to the

decisive factors of the situation described in the

scenario. In order to do that, the prediction agent

must know the factors, that is to say the items

associated to them, which are considered to be the

most representative of the current state of the

situation. Calculating these factors is the job of the

synthesis agents belonging to the intermediate

agents layer of the system.

The goal of the prediction agents layer is to

provide a continuous recall process of cases of the

case base, unlike the one used in CBR for dynamic

situations described before. Notice that, for the

moment, the adaptation step will be done by domain

experts which will be in charge to evaluate if the non

matched part of the recognised scenario can be used

for the current situation. Indeed, the continuous

evolution of the analysed situation may decrease the

relevance of the adaptation process result. That's

why, after having discussed with experts, it has been

chosen to give, to the expert, elements about the

main similarities and differences between the

Figure 3: Multi-agent and multi layer architecture.

ICEIS 2006 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

348

scenario and the current situation which he can use

in order to manage his own adaptation. More

information can be found in (Simon and

Boukachour, 2004).

6 MULTI-LAYER MAS

GENERICITY AND

SPECIFICATION

In the previous section, we have presented the multi-

agent architecture allowing the implementation of a

dynamic and incremental CBR for the evaluation of

the potential evolution of an dynamic observed

situation. The architecture is based on two parts:

• a generic part: 3 layers of agents;

• a bound part to the domain: ontology, measure of

nearness.

The genericity is one of the objectives of the multi-

layer system. The genericity of the system must be

understood here as being the separation between a

part of the mechanisms considered as being relevant

independently from the domain. The generic aspect

covers:

• the use of semantic features as atomic granules

of information piece: at a given time, the current

situation is representable by a collection of

semantic features;

• a factual agent takes care of a semantic feature;

• a factual agent arranges several internal

indicators which inform its states.

Some aspects of the internal indicators of a factual

agent are generic. For example, an evolution of the

speed will induce an update of the acceleration

independently from the domain.

Automaticaly, the ontology is specific to the domain.

The categorization of the semantic features is

specific too, as the choice of the valid transitions

from the generic automaton, the setting parameter of

thresholds and the actions were associated to the

transitions.

The collaboration between didacticiens and

computer specialists allow a clarification of the

concepts and the vocabulary of the common domain.

To initiate the learning and to manage an

individualized teaching in a multimedia environment

supposes to translate the characteristics of the

teaching agent into multi-agent system. Hubbard

isolates four characteristics of the virtual tutor which

we suggest to take into account in the multi-agent

system in the following career :

the "physical" presence and the personality of the

virtual tutor correspond to an adaptive and

intelligent human-machine interface;

the expertise in the field of reference corresponds in

the MAS to a knowledge base (base of scenarios and

ontology);

the capacity in an individualized teaching

corresponds to the implementation of a reasoning by

dynamic CBR supported by the architecture in three

layers;

the capacity to introduce the learning corresponds to

an autonomy of the system and the intrinsic

proactivity of its agents.

The low layer of the system, the factual agents,

contains agents carrying the semantic features bound

to the various actions of student. The career of the

student is so represented by a set of agents draw.

The ontology of the domain and particularly, the

object of the learning, allows to define the semantic

features.

The intermediate layer, the synthesis agents, consists

in placing the agents or the groups of factual agents

in regard each others. This layer participates in the

phase of elaboration of the target case of the CBR,

by keeping that the striking elements of the career

of student.

The highest layer, the prediction agents, its role is

to build a continual process and incremental of recall

step. At each scenario is associated a prediction

agent. The purpose of an agent of prediction is to

estimate continuously the degree of similarity

between the career of the student and the scenario to

which it is associated. A scenario contains the

determining facts of a known career as well as the

result of the evaluation of this career .

7 CONCLUSION

In this article, we have presented a multi-agent

architecture allowing the implementation of a

dynamic CBR for the evaluation of the potential

evolution of an observed situation. This architecture

relies on 3 layers of agents with a pyramidal

relation. The lower layer allows to build a

representation of the target case, i.e. the current

situation. The second layer allows to implement a

dynamic elaboration of the target case. Finally, the

upper layer implements a dynamic process of source

cases recall allowing the search for past situations

similar to the current one.

The system bases itself on heterogeneous data. It

is a question of going from an exhaustive and factual

description of the situation ( current work) in a level

A MULTI-AGENT ARCHITECTURE FOR MOBILE SELF-TRAINING

349

description knowledge allowing to characterize

synthetically this situation. The continuous treatment

of the information from the environment allows to

suggest to the actors (students and tutors) the

possible evolutions of the current situation. For that

purpose, we have to formalize the representation of

the successful information. To represent the current

situation, it is necessary to proceed to the

construction of an ontology of the domain to be able

to categorize the various semantic features

(elementary information).

The experiment of a tool of self-training by the

sociologists of the university of Le Havre produce a

set of files tracks. These files represent the career s

of a student to make a study case. The current work

consists in analyzing these files, to build the

ontology of the domain and specify the low layer by

identifying the semantic features.

REFERENCES

Annoot Emmanuelle, Bertin Jean-Claude, Gravé

Patrick,2004 « Quelles médiations dans les formations

à distance avec les nouvelles technologies dans

l’Enseignement Supérieur ? », Rapport de recherche

dans le cadre du CPER Pôle SHS, Universités du

Havre et de Rouen..

Bertin Jean-Claude, 2004 « L’ergonomie didactique face

au défi de la formation ouverte et à distance », ASP,

numéro spécial RANACLES.

Bertin Jean-Claude, Gravé Patrick, 2004a « Didactic

ergonomics and Web-based materials design »,

communication à CALICO 2004, Carnegie Mellon

University, Pittsburgh, USA.

Bertin Jean-Claude, Gravé Patrick, 2004b « Didactic

ergonomics and Web-based materials design: in

favour of a conceptual model », communication à Xith

International CALL Conference, CALL & Research

Methodologies, University of Antwerp (Belgique).

Boukachour H 2002. Système de veille préventive pour la

gestion de situations d'urgence: une modélisation par

organisations d'agents. Application aux risques

industriels. PhD Thesis, University of Le Havre.

Boukachour H., Simon G., Coletta M., Galinho T., Person

P., Serin F.,2002 Preventive Monitoring Information

System: a Model Using Agent Organizations SCI2002,

Orlando, USA.

Boukachour H., Simon G., Serin F., Galhino T., Coletta

M., Person P. and Fournier D.,2003 Vers une

architecture multi-agent pour la représentation et

l'évaluation de situations dynamiques, CCGEI'03

Montréal.

Cardon A.,1997. Les systèmes d'information et de

communication de gestion de crise: une modélisation

par agents. Ingénierie des Systèmes d'Information,

5(2), p. 167-193.

Coma R., Simon G, Coletta M., 2003 « A multi-agent

architecture for agents clustering » Agent Based

Simulation ABS’2003, Montpellier.

Denhière, G. and Baudet, S. (1992). Compréhension de

texte et science cognitive . Editions Puf.

Durand S., 1999. Représentation des points de vues

multiples dans une situation d'urgence : une

modélisation par organisations d'agents. PhD Thesis,

University of Le Havre.

Higgins John, 1986 « Smart learners and dumb machines

», Systems, vol. 14, n° 2,1986 pp. 147-50.

Hubbard Phillip, 1999 « Teaching agents in CALL

tutorials », CALICO’99, Oxford, Ohio, USA.

Hubbard Phillip, 2000 « Taming teaching agents, meaning

technologies and participatory dramas”, CALICO

2000, University of Arizona, Tucson, USA

Jackendoff, R., 1993. Semantics and Cognition

Cambridge, M.I.T.Press.

Jennings N., Wooldridge M., Sycara K., 1998 A roadmap

of agent research and development. Autonomous Agent

and Multi-Agent Systems, 1(5), p 7-38.

Kolodner J., 1993 Case-based reasoning, San Mateo CA :

Morgan Kaufman.

Mbala A., Reffay C. and Chanier T. 2003 SIGFAD : un

système multi-agents pour soutenir les utilisateurs en

formation à distance. In Actes de la conférence

Environnements Informatiques pour l'Apprentissage

Humain (EIAH'2003), Strasbourg, France, pages 319-

330.

Marcenac P., 1997 Modélisation de systèmes

complexes par agents. Techniques et sciences

informatiques, p 1013-1037.

Person P., Boukachour H., Coletta M., Galinho T.

and Serin F, 2005. From Three Multi-agent

Systemsto One Decision Support System.

IICAI’05. Inde

Simon G and Boukachour H, 2004 Towards a Multi-Agent

Architecture for Dynamic Case-based Reasoning".

ICKEDS'04. International Conference on Knowledge

Engineering and Decision Support.Porto (Portugal).

Simon G, Boukachour H., and M. Coletta, 2002. Vers une

architecture multi-agent pour la modélisation et

l’évaluation de situations dynamiques. Technical

report, LIH, Université du Havre.

Vygotski Lev Sémionovitch, 1978 Mind in society : the

development of higher psychological processes,

Harvard University Press, Cambridge, MA.

Webber, C., Pesty, S. 2002 Emergence de diagnostic par

formation de coalitions - Application au diagnostic

des conceptions d'un apprenant. In: Journées

Francophones pour l'Intelligence Artificielle

Distribuée et les Systèmes Multi-Agents

J.P.Muller(ed), Hermes, Lille, pp.45-57.

Wooldridge M., Jennings N.R., 1998 Pitfalls of

agentoriented development, 2nd International

Conference on Autonomous Agents, pp. 385-391,

Minneapolis.

ICEIS 2006 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

350