A new approach for multi-agent reinforcement learning

Abstract

The classical approach of reinforcement learning for single agent is based on the concept of reward that comes only from the environment. By trial-and-error, the agent has to learn to maximize its total accumulated reward. Several algorithms and techniques were developed for a single agent reinforcement learning. Our purpose is to benefit from all done work in reinforcement learning of an agent and extend it to multi-agent system. we have proposed a new approach that is based on the following idea: Communication or cooperation of a team can be achieved through mutual reinforcement of agents, that means agents can give and receive rewards from other agents and not only from the environment. The goal of each agent is to maximize the global accumulated reward received from environment and from other agents. We treat the multi-agent system as a unique entity. We define, state, action of the system and the value of each state with respect to the global policy of the system.