A Communication Approach for Cooperation of Multi-Agent Reinforcement Learning Based on Value Decomposition

Le résumé fourni par la source

In cooperative multi-agent reinforcement learning, communication can improve coordination between agents and form effective cooperative strategies. However, the additional communication structure increases the learning cost for agents and brings challenges for many value decomposition methods. Excessive message transfer also leads to significant communication overhead, hindering its practicality in many real-world environments. In this paper, a value decomposition-based multi-agent communication method is proposed. Agents will consider the state of teammates when making decisions, decide the content of communication messages and the communication target according to the attention level. This method satisfies the Individual-Global-Max condition and evaluates the influence of communication messages on teammates' decision-making to quantify their contribution to the team reward, thereby achieving an efficient and reasonable value decomposition. To minimize communication overhead, agents decide whether to communicate with the target based on their level of attention to teammates and cut off low attention communication connections during the execution phase. The empirical results show that this framework helps to promote cooperation between agents while reducing a certain amount of communication overhead without sacrificing too much performance.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.