Reward Distribution of Multi-Agent Reinforcement Learning Based on Evaluation of Contribution Degree

Abstract

Reinforcement learning is a technique developed for a single agent. When we use it for the cooperative behavior in multi-agent environment as it is, the problem like how much reward should be distributed to each agent in order to learn correctly will be generated. Therefore, in this research we propose a reward distribution method called multi-agent profit sharing (MAPS) to solve the reward distribution problem occurred in multi-agent reinforcement learning. In this method, we evaluate all the behaviors of each agent by an individual behavior evaluation and a cooperative behavior evaluation using fuzzy rules. We compute the separate individual and cooperative contribution degree based on the behavior evaluation and distribute the reward according to the contribution degree. Using the cooperative contribution degree we evaluate all the behaviors from the overall system and construct the multi-agent system which agents are able to learn the cooperative behavior and group strategy efficiently.