Policy Gradient Reinforcement Learning for Membership Functions in Policy Represented by Fuzzy Rules: Application to Simulations on Speed Control of an Automobile

Abstract

A method of a fusion of fuzzy inference and policy gradient reinforcement learning has been proposed that directly learns, as maximizes the expected value of the reward per episode, parameters in a policy function represented by fuzzy rules with weights and membership functions. A study has applied this method to a task of speed control of an automobile and has obtained correct policies with learned weights of rules, some of which control speed of the automobile appropriately. However, membership functions that quantify fuzzy concepts were designed based on human knowledge. Therefore, in this research, we show the result of experiments that the fusion method can learn the membership functions represented by a layered neural network.