Reinforcement Learning Based on Dynamic Construction of the Fuzzy State Space -Adjustment of Fuzzy Sets of States-

Abstract

In the previous paper, we proposed a method with dynamic construction facility of the state space, where we initially have no state and gradually add a new state of fuzzy set with removing unnecessary actions. We adjusted Q values for actions but not fuzzy sets for states. In this paper, therefore, we propose a method to adjust fuzzy sets, the central value and width of its membership functions, by TD (Temporal Difference) error. Then, we apply this method to the pursuit problem in real number environment.