Interactive self-self-reflection based multiagent reinforcement learning for coordination

One of the main goals of artificial intelligence is to realize an intelligent agent that behaves autonomously by its sense of values. Reinforcement learning (RL) is the major learning mechanism for the agent to adapt itself to various situations of an unknown environment flexibly. The merit of RL is that only giving the agent a goal state as setting a reward, a set of optimal behavior sequences toward the goal state from each state can be learned by trial and error. However, in a multiagent system environment that has mutual dependency among agents, it is difficult for a human to setup suitable learning goals for each agent, besides the existent framework of RL that aims for objective and egoistic optimality is inadequate. Therefore, it requires the active and interactive learning function that treats how to coordinate the interaction among other learning agents. The paper presents a framework of multiagent reinforcement learning to generate and coordinate each learning goal interactively among agents. To realize this, it treats each learning goal as a reinforcement signal that can be communicated among agents. Then the issues of the self-generation of goals and evaluation criteria are discussed.