Inducing Domain Theory from Problem Solving in a Multi-agent System Eloi Puertas and Eva Armengol IIIA - Artificial Intelligence Research Institute, CSIC - Spanish Council for Scientific Research, Campus UAB, 08193 Bellaterra, Catalonia (Spain). email: {eloi, eva}@iiia.csic.es, Abstract. The approach of in this paper tries to model the scenario of how an agent with poor domain experience could improve its problem solving behavior. In contrast to other approaches, we do not permit that agents exchange domain knowledge (neither cases nor domain theory). The agent with poor experience takes benefit of the problem solving behavior of other agents to improve its performance. Thus, it requests other agents for solving known problems and then induces one domain theory per requested agent. Finally, the agent achieves a higher accuracy in solving problems by his own using the induced domain theories. 1 Introduction A multi-agent system (MAS) is composed of a collection of agents holding a set of properties [7] and also they are able to both coordinate and cooperate in order to achieve a goal. The introduction of learning capabilities into a MAS allows the improvement of the global prob- lem solving behavior. Some approaches use inductive learning methods for concept learning on a MAS. The goal of concept learning is to induce a domain theory compatible with all positive and negative examples. Therefore, the goal of concept learning in MAS is to build an integrated domain theory compatible with the positive and negative examples of all agents. For instance, Davies and Edwards [4] propose an extension of the Version Space [8] method for concept learning in MAS. The goal is to integrate the version spaces of each agent in order to build a domain theory consistent with all the local domain knowledge. A similar idea is introduced by Brazdil and Torgo [3]. Here the authors consider that each agent is able to induce a domain theory and then all the individual domain theories are transferred to one of the agents who integrates them. In both approaches [4, 3], the integrated theory can be used by any of the agents belonging to the system to independently solving new problems. Notice that agents have to share domain information to build the integrated theory. In our paper we propose that an agent can improve his accuracy by inducing a domain theory from the problem solving behavior of the other agents. Moreover, we do not permit the exchange of information among agents but only solutions of cases. Thus, one agent ask the others for solving problems and induces a domain theory taking into account only the descriptions of the problems solved correctly.