Action learning to single robot using MAS — A proposal of agents action decision method based repeated consultation

Robots can employ a multi-agent system (MAS) as a technique to adapt to complex environments. In a MAS, numerous agents operate autonomously, but each agent is required to make decisions by considering other agents. Thus, agent cooperation is an important feature of a MAS. In this study, we focus on a MAS where the agents make connections by reinforcement learning. We propose a method that allows agents to learn and cooperate via communication. The actions of other agents are added to the state of each agent. Each agent performs virtual action selection and communicates with other agents to produce each action output.