Reinforcement Learning with Symbiotic Relationships for Multiagent Environments

Abstract

Studies on multiagent systems have been widely studied and realized cooperative behaviors between agents, where many agents work together to achieve their objectives. In this paper, a new reinforcement learning framework considering the concept of "Symbiosis" in order to represent complicated relationships between agents and analyze the emerging behavior. In addition, distributed state-action value tables are also used to efficiently solve the multiagent problems with large number of state-action pairs. From the simulation results, it is clarified that the proposed method shows better performance comparing to the conventional reinforcement learning without considering symbiosis.