FCM-Induced Switching Reinforcement Learning for Collaborative Learning

Switching data analysis is a method for simultaneously preforming local model estimation and data partition of mixed datasets drawn from multiple environments. This paper proposes a novel switching multi-agent reinforcement learning model, which simultaneously analyzes the clustering of agents and cluster-wise Q-learning under the Fuzzy c-means (FCM) concept. Multi-agents are assumed to be solving their problems in several different environments while we do not know in which environment each agent is solving the problem. The advantages of switching Q-learning are demonstrated through numerical experiments with a mixed bandit problem such that parallel multi-agent learning is useful for improving the efficiency while agent clustering contributes to handling the difference characteristics of multiple environments.