A Reinforcement Learning Algorithm for Continuous State Spaces using Multiple Fuzzy-ART Networks

Abstract

This paper describes a new reinforcement learning system for unknown continuous state space environments. The purpose of our study is to divide the continuous state space to enable a reinforcement learning agent to perform a task well. Our method uses multiple fuzzy-ART (adaptive resonance theory) networks to divide a continuous state space. In our method, multiple reinforcement learning modules that use the fuzzy-ART networks as state recognizers learn concurrently, and the agent changes the state spaces for action selection from low resolution to high resolution in order to realize a good balance between the speed of the learning and its optimality. The results of the mobile robot simulation show the usefulness and efficiency of our learning system