Emergence of Roles for Cooperative Behavior by Multi-Agent Adversarial Inverse Reinforcement Learning with Task Auxiliary Reward

Abstract

Humans in a group play different roles and cooperate to achieve a goal of tasks. For instance, group members who have a role variation (e.g., decoy and seeking) show better performance than the members who have the same role, such as ran away only in the tagging task. To realize such sharing roles in the group behaviors, each member should acquire a rich variation of actions. It is known that imitation learning is powerful to learn behavior efficiently; however, existing studies have not considered the diversity of group members such as physical properties and learning capabilities. We proposed the adversarial imitation learning method, which employs an auxiliary task reward for multi-agent behavior learning. The proposed model was evaluated in the multi-agent tagging task under different group property situations. Experimental results showed that the proposed model with specific parameters could acquire actions variously than the model without the auxiliary reward.