A Reinforcement Learning System Based on State Space Construction Using Fuzzy ART

摘要

A new reinforcement learning system using fuzzy ART (adaptive resonance theory) is proposed. In the proposed method, fuzzy ART is used to classify observed information and to construct effective state space. Then, profit sharing is employed as a reinforcement learning method. Furthermore, the proposed system is extended to the hierarchical structures for solving partially observable Markov decision process (POMDP) problems. Through various computer simulations using maze problems, it is confirmed that the proposed methods are effective to solve POMDP problems.