FUZZY INFERENCE SYSTEMS OPTIMIZATION BY REINFORCEMENT LEARNING

Abstract

Fuzzy rules for control can be effectively tuned via reinforcement learning. Reinforcement learning is a weak learning method wich only requires information on the succes or failure of the control application. In this paper a reinforcement learning method is used to tune on line the conclusion part of fuzzy inference system rules. The fuzzy rules are tuned in order to maximize the return function . To illustrate its effectivness, the learning method is applied to the well known Cart-Pole balancing system problem. The results obtained show significant improvements of the speed of learning.