Learning in Multi-Agent Systems

Abstract

Learning agents acting in a multi agent environment can improve their performance. These agents might decide upon their course of action by learning about other agents with whom they interact. The learning agents can learn about the others ’ information and rules of behavior. The agents will not need to plan their actions beforehand, each time they are asked to solve the same problem they have already solved or when dealing with similar problems. Our research focusses in finding learning algorithms for multi agent environments. In particular, we want to develop agents that learn how to cooperate with other agents and learn how to become experts. I am studying these themes in several domains to achieve a more broader understanding of how learning is influ-enced by different multi-agent domains and thus how learning algorithms can be developed. The research aims at answering the following questions: 1. How do agents learn how to act? I am looking at a teacher-learner model in which the teacher might also play the role of the learner, and the learner might also behave as a teacher. I am investigating two main directions. One direction regards a 2-phase model and we investigate how agents can be trained by other agents and then how they generalize the knowledge they acquired (Goldman & Rosenschein 1995). The second direction focusses on a dynamic learning model in which the agents learn from the others incrementally and also act based on the cur-rent knowledge they have. I am also investigating the behavior of learning agents that are implemented as automata (deter-minisitic and probabilistic). We investigated (Mor, Goldman, & Rosenschein 1995) how one learning agent can decide which action to play while he learns his opponent’s strategy. Currently, we are focussing on mutual learning In both cases, I am looking at reinforcement learn-ing, where the feedback is given to the agents by the