A survey on multi-agent reinforcement learning: Coordination problems

Abstract

Learning in multiagent system needs to solve the complexity of the task, so multiagent reinforcement learning has been focused on theoretical research and various applications. In multiagent reinforcement learning, agents can be compete or cooperate to accomplish the goal. For cooperative multiagent reinforcement learning(CMRL), agents have to coordinate with other agents. Therefore, coordination problems in CMRL are getting more and more important because of increasing the number of agents and actions. There are several algorithms dealt with cooperative multiagent reinforcement learning using stochastic games, coordinated graph, and so on. These algorithms have some assumptions to coordinate each other, however assumptions are not consistent with characteristics of the multiagent system. In this paper, we provide a survey on coordination problems in cooperative multiagent reinforcement learning, and propose new approach to solve coordination problems.