A Framework for Multiagent Reinforcement Learning Using Fuzzy Sets

摘要

This paper discusses the framework of a collaborative learning method using adaptive fuzzy sets in the category of multiagent reinforcement learning. In this article, the existence of coordination problems leads to an action evaluating function through taking all of the rewards of agents into account. We will give a proof that our proposed method converges due to adjusting learning rates by introducing fuzzy membership function over time for all agents. Based on the Q-learning of Markov decision processes (MDPs), we propose that application of an adaptive fuzzy membership function to obtaining optimal learning rates, for the purpose of gaining robustness and swiftness of the Q-learning method via tuned parameter of the fuzzy membership function actively.