Fuzzy Reinforcement Learning System Based on Genetic Algorithms

Fuzzy inference systems are increasingly being used in development of intelligent systems. However, these systems use knowledge bases which are refined manualy. Fuzzy Reinforcement Learning has been shown to provide automatic fine-tuning capability for several experimental systems. In particular, the GARIC architecture developed at NASA Ames, is an experimental system which demonstrates the process of automatic rule refinement. Here, we propose to extend this theory with Genetic Algorithms. The main innovation in this research is the idea of replacing the Stochastic Action Modification method, as currently used in fuzzy reinforcement learning, with the more powerful search techniques of genetic algorithms. We will explore the feasibility of developing a commercial software tool based on extended fuzzy reinforcement learning and Genetic Algorithms. The development of this method and its software shell will significantly contribute to the use of knowledge-based systems for aerospace applications as stated in the topic 06.05 of the 1993 SBIR program solicitation. This research offers a great potential for all fuzzy inference systems for automatically improving their performances by using machine learning.