FORSIC: A SELF-ORGANIZING TRAINING SYSTEM Jo LINK-PEZET*, Marie-Pierre GLEIZES**, Pierre GLIZE** *Unité Régionale de Formation à l'Information Scientique et Technique 11 Rue des Puits-Creusés - 31000 Toulouse, France Email : pezet@cict.fr and **IRIT - Université Paul Sabatier 118 Route de Narbonne - 31062 Toulouse Cedex, France Email : gleizes@irit.fr glize@irit.fr ABSTRACT The main goal of this presentation is the implementation of a collective knowledge management system. With networks, intellectual practices have changed, as well as the associated information practices, through the constant possibility of browsing the authorized sources and of collecting new, re-usable information. This implies a certain information culture or background in the use of information systems (content, organization) as well as knowledge about the topic (relevance) from the searchers. We assert that massive access of electronic networks necessitates training in information retrieval and use. Since 1997, the French universities are supposed to provide this type of teaching to the students, which implies the existence of teachers and of programs. This paper derives from three convergent observations: the necessary adaptive ability of a system in a dynamic environment, the role of the internal organization between the components for the global behavior, the performance improvement at a society level when cooperation is observed at the agent level. We have defined an adaptation process for artificial systems based on self-organization between its parts (agents) : cooperation is the criteria used to evaluate and reconsider the current organization (a collective knowledge management system to help the trainers to train the students). Keywords: Collective knowledge management, adaptive network, self-organization, cooperation, multi-agent learning. 1. THE CONTEXT: A NETWORKED ENVIRONMENT Networks provide an important source of available and valuable information. The network of networks is a unique, distributed, real-time system constituted by highly connected, heterogeneous elements. Its main characteristics stem from its complexity and dynamics. The challenges at a trainer’s level are : to describe the relevant information that a student should receive, to identify training needs, to find support on a “knowledge community” and to identify the information literacy as a discipline, to be able to create the pedagogical resources and to transmit the knowledge he has on the information resources he uses (cd-roms, databases, Internet) in a way that is adapted to the students’ needs, to be able to use the competencies of other teachers and to multiply them through cooperative work so as to adapt to new situations and needs, to provide a “learned” culture of information for a larger audience (a campus) as a social, political, economic and scientific goal, to collect all relevant new information according to the trainers’ profile. Achieving this vision requires a radical restructuring of social relationships and an insight across a range of relevant research areas. Success is expected to depend on a broad and interdisciplinary perspective, pulling together expertise from various disciplines such as information science, organizational theory, distributed cognition, cooperative multi-agent technology and ethnomethodology. We have been inspired by Jean Lave’s work on apprenticeship [1] and E. Hutchins [2] on cooperative work in a distributed environment. Our project is a complement to FORMIST (http://formist.enssib.fr/), an inventory of pedagogical resources on the same subject : both projects are financed by the French Ministry of Education. 2. THE PROJECT The basic principle is to create an adaptive and cooperative system composed of technological devices, to plunge it in the considered dynamic environment constituted by a numerous community of individuals (teachers and learners) in order to identify, increase and improve the teaching material, as well as its accessibility A Well-trained population of professional information users The URFIST (Unité Régionale de Formation à l'Information Scientifique et Technique) was created in 1983 by the French Ministry of Education in order to train information professionals in the use of electronic information. They are specialized both in the use of information tools and in various scientific domains. Because of the urgent need to train the university students in bibliographic research, the URFIST of Toulouse (France) has trained quite a number of these professional experts to convert their professional knowledge into a pedagogical skill or competency [3], so that they can in turn train students in the use of bibliographic resources for their research. Considering that these information specialists have all been to some extent the authors of pedagogical resources, and have competency in using specific resources, we have conceived a cooperative knowledge management system called FORSIC (Formation et Recherche en Sciences de l’Information et de la Communication). In fact, all these persons have been trained in what we call cooperative thinking, as they are using various tools (software, devices), are used to work in teams, are able to judge the relevance of information, rank, map, datamine it [4], [5]. We have tried to imagine how to tame these technologies to meet a user’s needs in a highly dynamic environment and to favor interaction in a community of peers. The practice of knowledge conversion, as part of a learning organization, is a necessary in order to follow the evolution of the environment. We use various technologies in what we call a learning organization, because we want to experiment them in real life. We hope to be able to bring some suggestions or new solutions to certain training challenges and to see the emergence of a pedagogical referral. The Description of the learning organization The basic idea has been to build a virtual and dynamic hypertextual system, in which the knowledge of each person is described and where every single item of information is both linked to other items (context creation) and to other persons (networking). Each item describes a facet of the knowledge of each person belonging to the community, and shows in what way the knowledge of each person is distinct from other persons’ knowledge. All information items are memorized, both indexed (according to a pre-set list of competencies) and described in natural language in order to be searched, interpreted, and used. We have identified four steps corresponding to four different uses of information. Each function corresponds to a different level of technology. 1. Identify the knowledge capital of a community : to make community members aware of what they know at an individual level, to make the collectivity aware of the collective knowledge it possesses. It concerns the individuals, their intellectual production (pedagogical resources, their contributions to individual and collective projects, the resources they know, use most frequently, advise for a precise use, etc). It is a multidimensional data warehouse and a communication tool. It will be built with Lotus Notes Domino. 2. Represent : we want to create a visible, referential system using See-K, a software and a knowledge management method produced by the society Trivium. This knowledge management software represents the knowledge of the community in the shape of a tree and the pedagogical resources in the shape of interconnected islands. Therefore it is possible to obtain a representation of the cognitive capital of a community and to visualize the contents of the teachings they make. It is a good resource management system to organize the training sessions (considering the great number of teachers and training sessions) according to the needs, to see the real emergence of a cognitive capital. 3. Transform : often the numerous pedagogical resources are similar. The mapping of the pedagogical resources is a good step towards visualizing the contents, and it is possible to datamine the relevant resources in order to answer a specific problem. We want to be able to re-engineer them and re-use them. We currently cooperate with two researchers specialized in authoring systems and agent technology. We are developing a domain model for information teaching, a model for the trainers and using an inference engine for the creation of individual personalized resources. 4. Create and animate an intelligent collective through the cooperative interweaving of links to create new resources by learning and adaptation. In the following paragraphs, we are going to develop this last aspect. A Collective intelligence process for emerging information system In such an environment, the system needs to be able to react and readjust its behavior to ever-changing circumstances and evolving needs. For this reason, we chose a decentralized system to allow the dynamic creation, deletion, combination and recombination of the various elements of the system through self-stabilizing mechanisms and without loss of identity into the collaborative structure. A trainer’s profile corresponds to the information about the trainer and his activity. It can be constructed explicitly or deduced by the system (trainer’s agent) by observing the trainer’s behavior. In order to learn autonomously, the user’s agent is “looking over his shoulder”[6]. Information filtering is a name used to describe a variety of processes involving the delivery of information to people who need it [7]. Filtering can reduce the information load on the user, provide better coverage of news streams, support ongoing information interests, and provide a mechanism for locating resources and people associated with a particular area of interest. Collaborative filters help people to make choices based on the opinions of other people. We want to try alternative ways to map the available information corresponding to users' interests (difficult to overcome the indexing and vocabulary problem). Building models of users' interests (which we do in the case of the individual resource production) is also extremely difficult because of the differences between each individual user (his knowledge and cognitive model), and because these are constantly changing. Checking the patterns of keywords is not sufficient for modeling concerns because we need semantic and contextual information. On the opposite hand, we believe that contextual information is difficult to express and to use. It is the reason why we apply a theory based on cooperative self-organization to solve this problem. To do this, only textual description are used. The idea is to find the relevant information needed by each trainer, on the basis of the beliefs he/she has of his/her own needs. The profile is built from the observation of the user’s searching or teaching habits : a personalized assistant is implemented on the user’s regular browser which will learn dynamically and permanently adapt to match the needs using the agent technology, e.g. to reach a state of functional adequacy on a cooperative work device [8]. The evolution of the user’s needs and creation/deletion/modification of pedagogical resources precludes a classical centralized control and learning mechanisms are based on self-organization. An Assistant for the creation of pedagogical resources When a user wants to use FORSIC, he downloads an agent assistant from a predefined site on the Web and launches the install procedure. He now possesses his own agent assistant which continuously monitors his behavior and provides him with advice. The user can describe his subjects (topics) of interest that his agent assistant can learn. When the user describes a subject in which he is interested, the agent assistant guides him for the formulation and suggests related terms issued from the resources description. The system will spontaneously offer new information according to the user’s known fields of interest. The resources are listed and ranked according to the relevance degree to the subject, but also with the “semantic” links between them. The user is able to give a simple evaluation of the previous transactions he has made, of the relevance of the various retrieved documents according to his needs. The result of this information exchange will be used to improve the global knowledge of the system about these sites. As a consequence, an emerging "infosystem" constantly scales up or down, evolves and adapts in order to meet the changing demands of the individual user and creates an extended, larger and highly dynamic knowledge production. Figure 1 - The user' view for a pedagogical resource presentation Figure 1 gives a representation of the user-system interaction. In the frame located on the left part of the browser Window (here Netscape), we find the list of relevant resources structured as a document. There are hypertextual virtual links obtained from the learning mechanism in the FORSIC tool. For example this document is made by the aggregation of many subsets from different courses identified in FORSIC matching the end-user’s needs. We insist upon the fact that these links are not predefined physical links between pages corresponding generally to the global course document of a trainer in FORSIC. On the right part of the screen, the user has opened one of these pages. The current page is indicated in the frame by an inverse video line. The organization of the virtual document is derived from the beliefs contained in the user profile. The glossary gives the list of terms related with the trainers’ needs and contained in all the pages of the virtual document. These relations are also deduced from the thesauruslike user profile. When connecting to the data base containing the description of the expert users, the final user and the multi-agent technology create a system that learns all the time, refines its global vision of the world and supports the dynamic creation of new types of relations and activities : it verifies, advises, redirects and gives relevance feedback. In doing so, it creates knowledge that can be used by the community, value and degrees of scalability, sustainability and robustness that exceeds any individual knowledge. The experiment is just starting and we believe the benefit will be great to the trainers’ community. Our goal is to extend FORSIC to a larger audience (the students on a campus, for example) who can use « certified » information and have the possibility to communicate directly with human experts. The created environment allows the knowledge and capabilities of every single information user to be enhanced and dynamically recombined with other persons’ knowledge to satisfy the goals and intentions of the individuals, groups or organizations on whose behalf they operate. 3. A SELF-ORGANIZING ARCHITECTURE FOR AN ADAPTIVE NETWORK Emergent activities in complex adaptive systems Many systems exhibit sophisticated collective information-processing abilities that emerge from the individual actions of simple components interacting via restricted communication pathways. Some often cited examples include efficient foraging and intricate nestbuilding in insect societies, the spontaneous aggregation of a reproductive multi-cellular organism from individual amoebae in the life cycle of the Dictyostelium slime mold, the parallel and distributed processing of sensory information by assemblies of neurons in the brain, or the optimal pricing of goods in an economy arising from agents obeying local rules of commerce [9]. These coherent global activities are realized by entities having only local views of their environment, as in FORSIC which works with well-known identified partners willing to participate to the community activities. Erroneous behavior or unexpected events are not controlled by an individual but must be attended to by adaptation processes. Adaptability is necessary to respond to unexpected events, to deal with irrelevant and conflicting information from many sources and in order to act when all the relevant information is not yet available. Computer scientists are nowadays confronted with the development of artificial systems which are not easy to use or understand. These difficulties are due to the performance of the machines and their networks. Commonly speaking, this refers to the broad use of complexity as an intuitive notion to speak about intractable systems [9]. We suggest that the way to reduce complexity at the conception phase of artificial systems is by avoiding to program at the higher global level. Thus, the computation is built up from microscopic level in a bottomup fashion, and the macroscopic structure emerges from the microscopic computation [10]. In the case of FORSIC, we consider that complex systems are notoriously difficult to maintain and that there are many problems to predict the emergent behavior of large numbers of interacting entities leading to the desired collective behavior satisfying the requirements defined in the previous paragraph. The Theoretical framework The theory followed by the cooperative assistant in this work is based on the following theorem: For any functionally adequate system in a given environment there is a system having a cooperative internal medium which realizes an equivalent function [11]. The consequence of this theorem is that the entities composing a system only pursue an individual goal (being cooperative with others) while adapting themselves to possible external perturbations through learning. Although no finality is explicitly programmed in the behavior of the system components, the self-organizing system works to provide an appropriate result. Multi-agent systems consist of several agents which are able to interact with each other and with their environment. Each agent has a local view of the environment, has specific goals and is unable to solve alone the global task of the system. The global characteristics of such a system thus arise from the cooperation between its elements. The result of cooperation, in turn, acts on the interactions between agents and modifies the properties of the system in an intricate manner. Designing a cooperative multi-agent system consists in defining for each component - the agents - all possible uncooperative states and the actions associated to these states in order to eliminate them. When the system is running in a dynamic environment, an internal process is observed in which the relations between agents are modified. Thus, the reorganization of the partial functions achieved by each agent leads to a modification of the function of the global system, and non cooperative states due to unexpected events are progressively eliminated. In this approach, the goal of each autonomous agent is to find the right place inside the organization in order to interact cooperatively with other agents. The behavior of an agent at each time step is determined by the fact that, from its viewpoint, only cooperative interactions are achieved. In order to test a theory (by falsification or validation) many experiments must be effectuated in various fields. For this reason, we have implemented our theory on cooperative self-organization in three applications: the first was the Tileworld game [12] in which we have experimentally verified that cooperative agents have better results than selfish ones. the second concerns an application in electronic commerce with France-Telecom [13], and with Deutsche Telekom [14], [15]. In these applications, the agents representing the users and the services create a dynamic network of mutual interest based on the cooperative self-organization process. the last concerns a national multi-disciplinary project about natural and artificial collective intelligence. The first results of a cooperative self-organized ants society application gives performances of at least better than natural simulated insects [16]. A Network of collective intelligence The network of networks as a real time distributed unique information system, a highly-complex system (http://lpsl.coe.uga.edu/Jacobson/CTCS/Resources/links.html): it possesses a great number of densely interconnected elements. As a consequence, the control of the global activity is not provided by a classical top-down conception. The self-organized information systems are innovative in the world of artificial systems conception, because they have the possibility to control a dynamic complex system through emergence [17]. This is how we guarantee the functional adequacy of a system through the co-operative self-organization of its parts [18]. We use our theoretical work on learning based on self-organization and actually applied in the ABROSE project [15]. The complex dynamics of FORSIC, which is due to the frequent evolution of the user’s needs as well as the creation/deletion/modification of pedagogical ressources, precludes a classical, centralized control and learning mechanism. Therefore, the technical approach is based on self-organization in which the links between system components reflect the current state of the system. TA 1 TA 2 . . . TA n MA TA 2 . . . TA n MA TA 1 TA 2 . . . TA n MA Figure 2 - Multi-agent architecture for a self-organizing Network To do this, we have designed a three level structure: 1. The middle level is the most implicit system, it is composed of a set of assistant agents which represents the FORSIC users (named TA for Transaction Agent in the Figure 2). A user can search for information (a learner) or provide information (a teacher). The assistant agent goal is to continuously (and transparently) guide the user in his searching activities. It decides autonomously to interact with others assistant agents to answer their needs or to ask help. This is possible because they benefit from a cooperative social attitude which allows a collaborative information search process. For example, an assistant agent can spontaneously send information to another agent, on belief that the other agent is interested in this information. For that, it possesses beliefs on others’ assistant agents. 2. At the top level, there are the mediation agents (MA in the Figure 2). A mediation agent mainly represents a set of pedagogical resources. It has a larger, although still local, representation of the Network obtained through the learning activities of the system based on the beliefs provided by the transactions evaluation. This evaluation is given by a user and is sent by its assistant agent. A mediation agent interacts with all the assistant agents located on its site, in order to help them in the searching function. In the global project we suppose that each Urfist in France manages its own pedagogical resources on a local MA. Currently, only one MA is developed in the URFIST of Toulouse. 3. In each mediation agent and each assistant agent, there is a self-organizing multi-agent system (this is what we referred to as the User Profile). This agent in the multi-agent system describes the organization of the agent society. The agents express the viewpoint that an agent of the others levels can have about themselves or about the others. They link the shared concepts of two agents and confer them on an agent to hold a dynamic representation of its environment. This multi-agent system manages all the terms used by users in order to organize them as a thesaurus with links of genericity, specificity, etc… The knowledge of a MA or TA concerning others agents is continuously updated on the basis of the transaction flow. Thus, each layer improves its internal organization in order to efficiently provide relevant information at any time. Because at each level of this organization, the system can adapt itself, a self-organized multi-agent system is present at each level. This type of system perceives its environment which is made of users and information resources. This perception occurs through the observation of users‘ transactions. Its action on the environment is based on the activities of the users, it gives recommendations. The system has a representation of its environment at all the levels of the architecture. It is never a perfect representation of the system because of its high dynamics, but nevertheless this representation is quite correct. Consequently, the system will provide answers which are close to the best possible answers. Matching a network of collective intelligence and the ideal requirements Taking into account the remarks put forth, cooperative self-organizing systems will converge towards zero default. 4. The learning capability of agents is due to self-organization and not to a linguistic approach. The systems learn at the same time as the end-users interact that with it. As a consequence, the system is capable of working with people from different countries and from any technical field . 5. The system having a representation of the information domains and of the user’s needs is able to build relevant matches between them. 6. It learns continuously in order to refine its global representation of the world. 7. Any page is physically linked to other pages on its original site, but the user disposes of logically linked pages not necessarily on the same site. The page organization displayed on the screen is based on the current page description the system possesses and on user needs. 8. The agent assistant is generally able to pre-load pages which may be needed in the near future by the user in using this logical organization and the current state of the user’s work. Using a cache located on the user’s computer avoids waiting and decreases communication costs. 9. The user is able to give a simple evaluation of the previous transactions he has made, about the relevance of some sites according to his needs. This result will be used then to improve the global knowledge of the system on these sites. 10. Because the system knows the current needs of any user, it is able to alert them of any new information which was not already present. Because an agent assistant is still at work when the user is not logged on, it is able to perceive new relevant information. 4. DISCUSSION AND CONCLUSION In the project, the environment of any user will be composed of: His usual browser to access to the URL of the FORSIC site (a Mediation Agent) on the Web, His assistant agent which is able to guide him permanently. This is only the new software from the user’s point of view, but there is no specific technical competency is needed to install and use it. In FORSIC, in 2000-2001, we will work on the base of personalized contact with identified end-users who habitually work with the URFIST. We can work with fifty trainers both in information systems and disciplines, for precise and controlled evaluation. In 2002, we will provide a larger access to the system to different sites in France (there are seven URFIST in France) and will observe the internal functioning of the global system. The goal is to then open the system to students (self-training and information research). By breaking the classical boundaries of classifications, categories, this initiative aims to foster the creation of a new trend for the research community. A sophisticated collective information-processing ability emerges from the individual actions interacting via communication. The global activity is realized by persons and systems having only a local view of their environment. Erroneous behavior or unexpected events are not controlled by an individual, and will find an answer through an adaptation process when meeting inexact and conflicting information from many sources, allowing to act before all relevant information is available. The learning capability of agents is due to selforganization and not to any linguistic approach. The networking of society and the evolution of technological supports invite today’s scientists, philosophers, technology specialists and information scientists to think about a new model of information treatment and its use in an heterogeneous, distributed, dynamic environment. It is no longer possible to think in terms of a hierarchic, a priori, closed model of knowledge (or information) production and organization. With the networking of information resources, the individual and his environment negotiate constantly the knowledge model used and its organization a posteriori. Electronic background favors the emergence of a new type of knowledge organization through the interaction and cooperation of individuals. Each person thinks and works from his “own locus of knowledge”, that from what he/she knows about his/her needs, strategy, project. Consequently, a great political challenge of the information society is the training of the individuals in accessing electronic resources and to provide an equal access to these resources. It is also the responsibility of society to set pedagogical projects of different kinds, tools of different types: this is our goal with FORSIC. 5. REFERENCES [1] Lave, J., Wenger, E. (in preparation). Situated Learning: Legitimate peripheral participation. [2] Hutchins, E. Cognition in the Wild. Cambridge. Mass: MIT Press, 1995. [3] Takeuchi, H., Nonaka, I. The Knowlege Creating Company. New York: Oxford University Press, 1995. [4] Link-Pezet, J. “Coopération et auto-organisation: quelques élèments de réflexion pour une nouvelle approche du travail intellectuel”,Solarisn°5http://www.info.unicaen.fr/bnum/jelec/Solaris/d05. [5] Link-Pezet, J. (in preparation) “Mémoires et intelligence collective”. Mémoire d’habilitation à diriger des recherches [6] Maes, P. “Agents that Reduce Work and Information Overload”, Communications of the ACM Vol. 37 n°7, 1994. [7] Belkin, N. J., Croft, W. B. “Information Filtering and Information Retrieval: Two Sides of the Same Coin?”, Communications of the ACM Vol. 35 n°12, 1992. [8] Gleizes M. P., Camps V., Glize P. “A Theory of Emergent Computation based on Cooperative Self-organization for Adaptive Artificial Systems”, 4th European Congress on Systemics - Valencia, Septembre 1999. [9] Crutchfield J. P. “The evolution of emergent computation”, Santa Fe Institute Technical Report 94-04-012, 1994. [10]Forrest S. “Emergent computation : Self-organizing, Collective, and cooperative phenomena in Natural and Artificial Computing networks”, Special issue of Physica D - MIT Press / North-Holland, 1994. [11]Camps V., Gleizes M. P., Glize P. “A self-organization process based on cooperation theory for adaptive artificial systems”, PERVS, International Conference – Krakovie, 1998. [12]Piquemal-Baluard C., Camps V., Gleizes M. P., Glize P. “Cooperative agents to improve adaptivity of multi-agent systems”, Intelligent Agents Workshop of the British Computer Society, Edited by N.S.Taylor and J.L.Nealon, Oxford, 1995. [13]Camps V., Gleizes M. P. “Cooperative and mobile agents to find relevant information in a distributed resources network”, Workshop on Artificial Intelligence-based tools to help W3 users, Fifth international conference on World Wide Web. Paris: 1996.http://www.info.unicaen.fr/~serge/3wia/workshop/papers/paper30.html [14]Camps V., Gleizes M. P., Glize P. “Une théorie des phénomènes globaux fondée sur des interactions locales”, 6 ièmes Journées francophones sur l'Intelligence Artificielle Distribuée & les Systèmes Multi-Agents, Éditions Hermès, 1998. [15]Camps V., Glize P., Gleizes M. P., Léger A., Athanassiou E., Lakoumentas N. “A Framework for Agent Based Information Brokerage Services in Electronic Commerce”, EMSEC Conference, 1998. [16]Topin X., Fourcassié V., Gleizes M. P., Théraulaz G., Régis C., Glize P. “Theories and experiments on emergent behaviour : From natural to artificial systems and back”, European Congress on Cognitive Science, 1999. [17]Wooldridge M., Jennings N. “A Theory of Cooperative Problem Solving”, Intelligent Agent Workshop of the British Computer Society, 1995 .