Social interaction of cooperative communication and group generation in multi-agent reinforcement learning systems

Recently, researches on multi-agent systems (MAS) which autonomous agents are able to learn cooperative behavior are actively performed. It is necessary for social agents to interact each other in order to have excellent cooperative performance in MAS. But it is difficult to give appropriate coordination at the right time, not to mention the generation of group behavior. We have aimed at the group behavior generation of social agents who have excellent autonomous learning ability like human through cooperative communication between agents to acquire cooperative behavior. Social agents are able to change the environment states to individual states and communicate their cooperative behavior by sending a special signal, which can not be understood originally but could be understood by other agents through reinforcement learning (RL) after some learning processes. Furthermore, if the agents could communicate their cooperative behavior successfully, their group identity will be strengthened by exchanging rewards among them. As the learning process, social agents can get the better cooperative ability than normal reinforcement learning method through group identity. Lastly, social agents not only adjust themselves to environment but also affect other agents to generate intelligent group behavior.