A MULTI-AGENT ARCHITECTURE FOR GROUP LEARNING Eliane Pozzebon1,2, Janette Cardoso2, Guilherme Bittencourt1 and Chihab Hanachi2 1
Santa Catarina Federal University, DAS 88040-900 Florianópolis, Brazil, (eliane,
[email protected]) 2
UniversitéToulouse 1, IRIT F-31340 Toulouse Cedex, France (jcardoso,
[email protected]) ABSTRACT In this paper we propose a group management specification and execution method that seeks a compromise between simple course design and complex adaptive group interaction. This is achieved through an authoring method that proposes to the author predefined scenarios that already includes complex learning interaction protocols and in which student and group models use and update are automatically included. The method adopts ontologies to represent domain, student and interaction models and object Petri nets to specify the group interaction protocols. Dirung execution, the method is supported by a multi-agent architecture. KEYWORDS: Learning in Group, Collaborative Learning, Ontologies, Intelligent Tutoring Systems, Multi-agents Systems.
1. INTRODUCTION The field of Artificial Intelligence in Education (AI-ED) is going through an accelerated evolution process, mainly due to innovative computer technologies, such as hypermedia, Internet and virtual reality [Alpert 1999] [Brusilovsky 2000]. Nevertheless, the conceptual gaps between authoring systems and authors and between instructional planning and tutoring strategy for dynamic adaptation are challenges that have not yet been overcome [Mizoguchi 2000]. These challenges are specially complex in Intelligent Tutoring Systems (ITS) when one consider, beside an individual interaction, a group interaction. In this case, the ITS should not only support the domain presentation for a single student, but also manage the group interactions. ITS's that allow group work present different degrees of group interaction control. At one extreme, we have systems that only make available the communication tools that allow the group interaction (chat, mail, forum, cooperative editors, etc), leaving all the problem solving and coordination activities under human responsability. At the other extreme, we have systems that control all the details of the group interaction, following well defined and rigid protocols. In the former, the author task of instructional planning is at least as hard as in traditional group work planning. In the latter, the lack of flexibility makes it difficult to achieve dynamic adaptation and to share and reuse ITS components across domains. In this paper we propose a group management specification and execution method that seeks a compromise between these two extremes. The method is based, on the one hand, on a formal specification language, to be used by the authoring tool developers, that allows the definition of arbitrarily complex learning interaction protocols, here called scenarios. On the other hand, the method is intended to be applied to ITS's created using the multi-agent ITS building tool MathTutor [Cardoso 2004] [Frigo2005], in such a way that it can explore the known structure of the domain and student models in such ITS's. To tackle the compromise between simple course design and complex adaptive group interaction, the following project decisions were adopted in the development: (i) the explicit representation, using ontologies, of the knowledge that describes the domain, student and interaction models, including their relationship, (ii) the use of a multi-level control process to increase the flexibility of the behavior without sacrificing the specification simplicity, (iii) the use of an expressive formalism, Object Petri Nets (OPN) [Sibertin-Blanc
1985], to specify the group interaction protocols. OPN are a formalism combining coherently Petri nets (PN) technology and the Object-Oriented (OO) approach. While PN are very suitable to express the dynamic and possibly concurrent and open behavior of a protocol, the OO approach permits the modelling and the structuring of its active (actor) and passive (information) entities. In our case, actors correspond to author, students, while information are those describing the domain and scenarios. The paper is organized as follows: Section 2 presents an overview on the related work. Section 3 introduces the MathTutor ITS building tool. Section 4 explains how a group interaction scenario is specified and Section 5 presents a simple example of a scenario, discussing in particular what is a group interaction. Finally, in Section 6, we present some conclusions and future works.
2. RELATED WORK The concept of organization (also groups, institutions, communities, etc.) within MAS has been proposed in several papers [Ferber 2003], [Coutinho 2005], [Dignun 2004], [Zambonelli 2001], [Dignun 2004], [Esteva2004, Hubner 2002, Odell 2004], [Hubner 2005, Horling 2004, Silva 2004, Hanachi 2004], [Frasson 1998]. The concept of group has also been used in the context of ITS [Miao 2006, Frasson 1998, Thibodeau 2000], [Labidi 2003, Hernandez 1998]. We do not address in this paper the formation of a group into an automatic way. This issue is dealt for example in NetClass [Labidi 2003] using the model of the learner, the information of the authors and sociometric test (that measures the degree of cohesion among students). In WhiteRabbit [Thibodeau 2000], the groups are created from the analysis of the user model based on the keywords (about his projects, experience, etc.) and also of the conversations.
3. MATHTUTOR ITS BUILDING TOOL MathTutor [Cardoso 2004] is a domain independent tool to implement multi-agent Intelligence Tutoring System (ITS). Courses developed using MathTutor are based on the conceptual model MATHEMA [Costa 1995]. This model proposes an architecture of ITS that consists of three modules: the Tutoring Agents Society (TAS), the student interface and the authoring interface (fig.1). The student interface provides access to the system and the authoring interface allows the definition of the course structure and contents. The TAS consists of a multi-agent system where each Tutor Agent (TA) contains a complete ITS focused on a subdomain of the course target domain. Each one of the intelligent tutoring agents in the TAS is responsible by one sub-domain. The MATHEMA provides a modeling scheme for these sub-domains that is divided into two views: external view, and internal view.
Fig. 1. MathTutor System Architecture
The external view is a partitioning scheme, based on epistemological assumptions, that was designed to guide the author during course development.
The internal view proposes that the knowledge associated with each sub-domain can be organized into a set of curricula. Each curriculum is progressively refined according to three levels of detail: pedagogical units, problems and interaction support units. In the pedagogical units level, each curriculum, that describes a possible sequence of sub-domain contents to be presented to the student, is refined into a set of partially ordered pedagogical units, possibly with prerequisites relationships. In the problems level, each pedagogical unit is refined into a set of problems, also partially ordered and possibly with prerequisites relationships. Finally, at the interaction support units level each problem is associated with a set of interaction units with the student, that support the problem solving activities, such as explanations, examples and exercises. The domain knowledge of any ITS developed using the MathTutor tool presents the structure defined by this internal view. This fact allows the construction of group interactions that, although not domain dependent, can explore the domain structure, going beyond the simple communication support between group members and the author, because these group interactions can use the problem solving activities already defined in the context of the underlying ITS. A further advantage is that the student model used in group interaction management can be defined as an extension of the student model in the underlying ITS, in such a way that preferences and previous results of each student in the context of individual learning using the ITS can be explored by the group interaction manager. Both, domain and student models, are represented using ontologies. These ontologies are briefly described in the next subsections.
3.1 Domain Model The domain model contains definitions of all the concepts in the internal view of the MATHEMA model. A course is represented as an instance of the domain model and contains all the information provided by the author. This information is of two types: properties and contents. Examples of properties are prerequisite relationships, degree of detail, level of difficulty, etc. Contents is what is presented to the student, typically an interactive page encoded into predefined HTML pages templates. The ontology described in [Frigo 2005] include classes to define prerequisite order graphs, that can be used to define the relationship among pedagogical units or problems and classes to represent specific types of interaction support units, whose contents are also specified by the author. The classes correspond to the concepts in the internal view of the MATHEMA model.
3.2 Student Model The student model, proposed in [Frigo 2004], contains definitions of all the concepts necessary to characterize a student and her history of interactions with the system. Its contents include static, information such as education level, based on a preliminary test, and preferences, and also dynamic information that consists of descriptions of the student activities during all her sessions of interaction with the system. This model was extended extended in order to take into account the information necessary to group interaction, e.g. which is the student role in a given group.
4. GROUP MANAGEMENT The goal of the proposed group management method is to allow the specification and execution of complex learning interaction protocols, without burdening the author with the task of specifying how the student and group models should be taken into account and updated during the interaction. The group management process has two levels: the specification level and the execution level. The different concepts involved in these two levels are described by the ontology given in the fig.2. The goal of a group is to allow the interaction between several students that collaborate in order to solve a problem. A group (instance) follows a given scenario (protocol). In the context of MathTutor (ITS), there is a library of such scenarios. There are mainly two ways to create a group (in MathTutor): 1) by the course instructor, during a real cours (several groups of student in a classroom);
2) by an Student that wants to work with others students (taking advantage of the internet environment); in these case, he publish the goal of the group, and others informations (constraints, number of participants, etc.). In the case 1 above, the group can be automatically built formed by the system, see for example [Labidi 2003] that creates homogenous groups based on the model of the learner, the information of the authors and sociometric test. In [Thibodeau 2000] they are built from the analysis of the conversations and the keywords.
Fig. 2. Interaction Model
4.1 Specification Level The concepts involved in the specification level are: Scenario, Role and Supervisor. The main concept is the scenario one that corresponds to an interaction protocol that governs the interaction within a group. The concept of Role structures the agents into classes according to their capabilities in order to participate to a Scenario. Each Role is defined by: its name; the required skills (that an agent must meet to be authorized to play that Role); and the casting constraints (such as the maximum number of agents that may play that role, the condition required to enter into the Role), etc. Some examples of roles: Team Leader, Problem Solver, Plain Member (played by students) and Instructor. We distinguish between the author, the one that conceive the scenario, and the instructor, the one that choose a given scenario and provide the needed information to feed the scenario. Each Scenario may have: • some prerequisites, e.g., a minimal number of group members or the situation of these members with respect to the course of the underlying ITS; • some management decisions, e.g., who should take the initiative to start the group activity (system, course instructor, student), who decides when the group activity finishes, etc; • some contents, e.g., the prerequisite order of pedagogical units and problems for a given domain. All this information is provided by the instructor, using an authoring interface. The Supervisor is an operational specification of an agent (class) able to play the Scenario, whose behavior is represented by an object Petri net.
For each Scenario we record the Instructor, i.e., the scenario creator, the Students authorized to participate in it (and their Roles - Team Leader, Problem Solver, Plain Member) and the Supervisor able to play this Scenario.
Fig. 3. Specification Level
The specification process is a two-step process as shown by the fig.3. The first step consists in selecting a Scenario from the library and in entering the contents. These contents are instances of the domain model of the underlying ITS and can be used only in the context of group learning or in individual learning. The second step integrates this scenario with the Student model to produce the Supervisor agent. The designers of the scenario can introduce an adaptive character to the Supervisor agent through the use and updated of the student and group models.
4.2 Execution Level The execution level is defined by a multi-agent architecture inspired by [Ferber 2003] and represented on fig.4 The concepts involved in the execution level (see right side of fig.2) are: Agent, Student Agent, Coordinator Agent, Supervisor Agent and Group.
Fig. 4. Execution Level
An Agent is an assignment of a Role, as Student Agent, that belongs to a Group. A Group follows a Scenario and is supervised by an Supervisor Agent (an instance of a Supervisor). The Supervisor Agent is created by an Coordinator Agent. Each student is represented by an Student Agent that stores internally the information of the student model relevant to the group management, e.g., the group activities in which the student has participated, statistics about the status of the student in these groups (leader or not) and the number of group communications in which she was involved, etc. It can also consult the student model stored in the TAS agents of the underlying ITS (see Section 2.2). Each Group, during its existence, is represented by a (group) Supervisor Agent that supervises the group activities according to the Supervisor definitions in the specification level. It also stores internally all the relevant information about these activities. Finally, the architecture includes a general Coordinator Agent, the responsible by the creation and destruction of groups in execution time and also by monitoring individual learning in order to detect opportunities for group learning activities.
5. AN EXAMPLE OF SCENARIO To clarify the notions introduced in the previous section, we present an example of a simple general scenario and show how it can be instantiated into a concret group management process. The scenario is intended to develop the ``divide-and-conquer'' strategy in problem solving. It supposes a problem that can be particionated into a certain number of subproblems. Each subproblem can be solved independently and their solutions must be combined to solve the original problem.
5.1 General Description Following the concept of scenario in section 3, we must define the management decisions, the prerequisites and the contents. Management decisions : The scenario includes three roles: supervisor, team leader and problem solver. The supervisor role can be assigned either to the course instructor or to the group supervisor agent. The team leader role should be assigned to one or more students responsible by the integration of the solutions of the subproblems. The choice of these students can be done dynamically, e.g., the first to complete a subproblem solution or the best graded in the underlying ITS. Finally, each problem solver role is assigned to one or more students. Contents: The contents of the scenario, to be provided by the author through the MathTutor authoring interface (see Section 2), consist of the following interaction support units: • A general explanation of the problem and its subproblems. • An exercise that test the correctness of the combined subproblem solutions. • For each subproblem: a detailed explanation. one or more examples of solutions of similar problems. two types of exercises: one that test the correctness of the subproblem solution and one that verifies whether the solution correctly implement the expected interface with the other subproblem solutions. It should be noted that these contents are instances of the domain model ontology described in Section 2.1 and can be also used in the context of an individual interaction with the underlying ITS. Interaction Model: The specification of the scenario also includes the interaction model that in this case consists of the following steps: Presentation of the problem and its subproblems. Assignment of the problem solver roles associated with each subproblem. Monitoring of the subproblem solutions using the solution examples and correctness tests.
For each correct subproblem solution, monitoring of the interface correctness with respect to the other subproblem solutions using the appropriate tests. Coordination of the interaction among group members that incorrectly implemented the interface between their solutions. Assignment of the team leader role. Coordination of the interaction among group members to construct the combined solution and monitoring of its correctness. This protocol is represented by a predefined OPN associated with the scenario and it defines the behavior of the group supervisor agent.
5.2 Possible Instance To test the group management support we use an ITS on the domain of Information Structure, an undergraduate discipline of the Control and Automation Engineering course at the Santa Catarina Federal University, Brazil. This ITS is build using the MathTutor tool and respects the definitions in Section 2. A possible instance of the scenario above can be defined as follows: Problem description: given a language that suppport integer aritmetic operations, how to extend it to suppport operations for other types of numbers (float, rational, complex). Four subproblems: aritmetic operation packages for each of the three new types of numbers and a dispatch function package. Interfaces among subproblems: conversion functions from one type to another.
6. CONCLUSION We presented a method, based on ontologies, to allow the development of group learning in the context of ITS's. The underlying ITS's are built using the MathTutor tool. The fact that the structure of the domain model of these ITS's are known allows the method to automatically include all the student and group model management into the group scenarios. The group interaction scenarios, combined with the student model, are later explored by the multi-agent society responsible by the execution of the method to create an adapted interaction. Future work includes the enhancement of the student and group models and the implementation of a more friendly authoring tool to develop the group scenarios.
REFERENCES Alpert, S.R., Singley, M.K., Fairweather, P.G, 1999. Deploying intelligent tutors on the web: An architecture and an example. JAIE, 10:183–197. Bittencourt,G., Frigo, L.B., Pozzebon,E., 2004. Petri nets for authoring mechanisms. In XV Simpósio Brasileiro de Informática na Educação (SBIE’2004), ISBN 85-7401-161-4, pages 378–387, Manaus, AM, Brazil. Brusilovsky, P. 2000, Adaptative hypermedia: From intelligent tutoring systems to web-based education. LNCS, ITS00. Frigo, L., Cardoso, J., Bittencourt. G., 2005. Adaptive Interaction in Intelligent Tutoring Systems. In : HT-2005, Intern. Workshop on Combining Intelligent and Adaptive Hypermedia Methods/ Techniques in Web-based Education Systems, Salzburg, Autriche. Cardoso, J., Bittencourt, G., Frigo, L.B., Pozzebon, E., Postal. A, 2004. MathTutor: A multi-agent intelligent tutoring systems. In 1st IFIP Conf. on AI Applications and Innovations, WCC´04, pages 22-27. Costa, E.B., Lopes, M.A., Ferneda, E, 1995. Mathema: A learning environment based on a multiagent architecture. In LNAI - Advances in Artifical Intelligence, volume 991, pages 141–150.
Coutinho, L., Sichman, J.S., Boissier, O, 2005. Modeling Organization in MAS: A Comparison of Models. in: First Workshop on Software Engineering for Agent-oriented Systems, Uberlândia. Cuesta, P., Gomez, A., Gonzalez, J.C., Rodrıguez, F., 2002. A Framework for Evaluation of Agent Oriented Methodologies. The MESMA Approach for AOSE. Proceedings of Fourth Iberoamerican Workshop on Multi-Agent Systems (Iberagents’2002), at IBERAMIA’2002, the VIII Iberoamerican Conf. on Artificial Intelligence, Malaga, Spain. Dignum, M.V., V´azquez-Salceda, J., Dignum, F.P.M., 2004. OMNI: Introducing social structure, norms and ontologies into agent organizations. In P. Bordini & et al. (Eds.), PROMAS 2004 (pp.183-200). Heidelberg: Springer. Esteva, M., Padget, J., Sierra, C.,2001. Formalizing a language for institutions and norms. Proceedings of the Eighth International Workshop on Agent Theories, Architectures, and Languages (ATAL-2001), pp 106–119. Esteva,M., Rosell,B., Rodriguez-Aguilar, J.A., Arcos,J.L.2004.AMELI: an agent-based middleware for electronic institutions. In [AAMAS], pages 236-243. Ferber, J., Gutknecht, O., Michel,F.,2003. From Agents to Organizations: an Organizational View of Multi-Agent Systems. 4th Int. Workshop on agent-oriented software engineering, IV AOSE 2003, Melbourne, Australia. Frasson, C., Martin, L., Gouarderes, G., Aimeur, E.,1998. Lanca: A distance learning architecture based on networked cognitive agents. In Lectures Notes in Computer Science. Inteligent Tutoring Systems. Procedings of 4th International Conference, 1:594–603, August 1998. Hernandez-Dominguez,A.,1998. An Architecture of Cooperative Learning in a Distance Education context International Conference on Engineering Education, 1998, August 17-20, Rio de Janeiro, Brazil. Hanachi, C., Sibertin-Blanc, C.,2004. Active Middle-Agents in Multi-Agent Systems. Dans : Autonomous Agents and Multi-Agent Systems, Kluve Academic PublishersNetherlands, V. 8 N. 3, p. 131 - 164, avril 2004 Horling, B., Lesser. V.,2004. Quantitative Organizational Models for Large-Scale Agent Systems. Proceedings of the International Workshop on Massively Multi-Agent Systems, pp 297-312, Melbourne, LNCS 2935, pp 214-230, December 2004. Kyoto, Japan Hubner, J.F., Sichman, J. S., Boissier, O.,2005. S-moise+: A middleware for developing organised multi-agent systems. In Boissier, O., Dignum, V., Matson, E., and Sichman, J., editors, International Workshop on Organizations in MultiAgent Systems: From Organizations to Organization Oriented Programming (OOOP 2005), pages 107.120. Labidi, S., Souza, C.M., Nascimento, E., 2004. NETClass: Cooperative Learner Modeling inWeb-Based Environment 6th International Conference on Computer Based Learning in Science CBLIS, Nicosia, University of Cyprus. Miao, Y., Pinkwart, N., Hoppe,U.,2006. Conducting situated learning in a collaborative virtual environment. Proceedings of the 5th International Conference on Web Based Education, pp. 7-12. Anaheim, CA: ACTA Press. 2006 Mizoguchi, R., Bourdeau, J.,2000. Using ontological engineering to overcome common AI-ED problems. J. of AI in Education, 11(2):107–121, (2000). Odell,J., Nodine, M., Levy,R.,2005. A Metamodel for Agents, Roles, and Groups. Agent- Oriented Software Engineering (AOSE) V, James Odell, P. Giorgini, J¨org M¨uller, eds., LNCS, Springer, Berlin, 2005. Sibertin-Blanc, C., 1985. High-level Petri nets with data structures. In EuropeanWorkshop on Application and Theory of Petri Nets, pp 141–170. Silva, V., Choren, R., Lucena, C.,2004.A UML based approach for modeling and implementing multi-agent systems. In [AAMAS 2004], pages 914.921. Thibodeau, M., B´elander, S., Frasson, C., 2000. White rabbit.matchmaking of user profiles based on discussion analysis using intelligent agents. Procedings of 5th International Conference, ITS 2000, 1:113–122, June 2000. Montreal/Canada. Zambonelli, F., Jennings, N., Wooldridge, M., 2001. Organisational Abstractions for the Analysis and Design of MultiAgent Systems. In: Ciancarini P.,Wooldridge, M. (eds.): AgentOriented Software Engineering, LNCS 1957, SpringerVerlag.
