An Action Selection Method Using Degree of Cooperation in a Multi-agent Reinforcement Learning System

Abstract

In recent years, a concept of a dividual is proposed by a Japanese novelist to interact properly with another person. To construct a model of the dividual, the degree of cooperation is assigned to the corresponding dividual. By introducing the degree of cooperation into a multi-agent system, we evaluate what kind of changes appears in the agent behavior. In addition, we propose an action selection method by introducing the degree of cooperation into the soft-max method in a multi-agent system. Using the proposed method, we confirm whether the cooperative action is promoted or suppressed through computer simulations.