Modular Learning Systems for Behavior Acquisition in Multi-Agent Environment

Abstract

There has been a great deal of research on reinforcement learning in multirobot/agent environments during last decades1. A wide range of applications, such as forage robots (Mataric, 1997), soccer playing robots (Asada et al., 1996), prey-pursuing robots (Fujii et al., 1998) and so on, have been investigated. However, a straightforward application of the simple reinforcement learning method to multi-robot dynamic systems has a lot of issues, such as uncertainty caused by others, distributed control, partial observability of internal states of others, asynchronous action taking, and so on. In this paper we mainly focus on two major difficulties in practical use : unstable dynamics caused by policy alternation of other agents curse of dimension problem The policy alternation of others in multi-agent environments may cause sudden changes in state transition probabilities of which constancy is needed for behavior learning to converge. Asada et al. (Asada et al., 1999) proposed a method that sets a global learning schedule in which only one agent is specified as a learner with the rest of the agents having fixed policies to avoid the issue of the simultaneous learning. As a matter of course, they did not consider the alternation of the opponent’s policies. Ikenoue et al. (Ikenoue et al., 2002) showed simultaneous cooperative behavior acquisition by fixing learners’ policies for a certain period during the learning process. In the case of cooperative behavior acquisition, no agent has any reason to change policies while they continue to acquire positive rewards as a result of their cooperative behavior with each other. The agents update their policies gradually so that the state transition probabilities can be regarded as almost fixed from the viewpoint of the other learning agents. Kuhlmann and Stone (Kuhlmann and Stone, 2004) have applied a reinforcement learning system with a function approximator to the keepaway problem in the situation of the RoboCup simulation league. In their work, only the passer learns his policy is to keep the ball away from the opponents. The other agents (receivers and opponents) follow fixed policies given by the designer beforehand. The amount of information to be handled in multi-agent system tends to be huge and easily causes the curse of dimension problem. Elfwing et al. (Elfwing et al., 2004) achieved the cooperative behavior learning task between two robots in real time by introducing the