FCAPS: Fuzzy Controller with Approximated Policy Search Approach

Abstract

A fuzzy controller requires an engineer to tune its rules for controlling a given plant. To reduce the burden, we develop a gradient-based tuning method for the fuzzy controller. The developed method is closely related to a theory of reinforcement learning, but takes advantages of a practical assumption made for faster learning. In reinforcement learning, values of problem states need to be acquired through lots of trial-and-error interactions between the controller and the plant. And the plant dynamics should also be learned by the controller. In this research, we assume that an approximated value function of the problem states can be represented as a function of a Euclidean distance from a goal state and an action executed at the state.And, we propose to use it for the gradient search as an evaluation function. Our experimental results on a pole-balancing problem show that the proposed method can tune the fuzzy controller to have an optimal policy for reaching the goal state despite an unknown plant dynamics in not only a set-point problem but also a tracking problem.