Collective learning of action sequences

Abstract

Learning in multiagent systems is a new research field in distributed artificial intelligence. The author investigates an action-oriented approach to delayed reinforcement learning in reactive multiagent systems and focuses on the question of how the agents can learn to coordinate their actions. Two basic algorithms, the ACE algorithm and the AGE algorithm (ACE and AGE stand for Action Estimation and Action Group Estimation, respectively), for the collective learning of appropriate action sequences are introduced. Both algorithms explicitly take into consideration that (i) each agent typically knows only a fraction of its environment, (ii) the agents typically have to cooperate in solving tasks, and (iii) actions carried out by the agents can be incompatible. The experiments described illustrate these algorithms and their learning capacities. >