Cooperative Behavior Acquisition in Multi-agent Reinforcement Learning System Using Attention Degree

Abstract

AbstractIn a multi-agent system, it becomes possible to solve a complicated problem by cooperative behavior with others. When people act in a group, as they are predicting the others’ action, estimating the others’ intention, and also making eye contact with others, they are realizing cooperative behavior efficiently. In the present paper, we try to introduce the concept of eye contact into a multi-agent system. In order to realize eye contact, we firstly define attention degrees both from self to the other and from the other to self. After that, we propose an action decision method that self agent makes easy to choose a target agent and to choose actions approaching to the agent using the attention degrees. Through computer simulation using a pursuit problem, we show that the agents making eye contact each other pursue preys by approaching each other. Simultaneously, we compare the proposed system with the standard Q-learning system and verify the usefulness of the proposed system.KeywordsMulti-agent systemCooperative behaviorEye contactAttention degreeReinforcement learningPursuit problem