Generation of roles in reinforcement learning considering redistribution of reward between agents

In this paper, we propose a method for generating roles by redistribution of reward between agents in reinforcement learning of multiagent system. In the proposed method, there are two kinds of reward; (1) reward obtained from environment when agents reach the goal and (2) reward received from other agents. Agents can learn actions to work cooperatively with other agents by the reward received from other agents, and the roles of agents are generated. In the proposed method, agents can decide how much reward to which agent to give from the obtained reward by using the history that records actions at the time when the change of environment are recognized. We carried out computer experiments for two tasks; (1) path finding problem and (2) transportation problem, and confirmed that roles of agents are generated by redistribution of reward between agents in the proposed method. Moreover, we confirmed that the proposed method can learn as similar as when the designer decides conditions for redistribution of reward.