Parameter and rule learning for fuzzy logic control systems using genetic algorithms
John A. Akec · 1996
The concept of reinforcement learning, or learning without a teacher or supervisor, is a challenging and interesting topic. Reinforcement learning, based on the method of temporal difference, has been applied with some success to neural network-based control systems and neural network modelling, and it is argued that it could therefore be fundamental to the design of control systems for plants of unknown or time varying parameters. However, neural network-based reinforcement has led to the design of algorithms which are as yet too complex to be readily exploitable in industrial applications. It therefore remains an active research area. Genetic algorithms have been applied to the problem of automatic rule selection and parameter learning for fuzzy logic based control systems. The question of formulating an effective cost function necessary for guiding the genetic search process, without external supervision or any training data, still presents many difficulties. This research paper presents a framework within which genetic algorithms can be complemented by ideas established in neural network-based reinforcement learning and classifier systems, for automatic rule generation and parameter learning for fuzzy logic based control systems. The ideas are illustrated by simulation studies for the control of a nonlinear, and inherently unstable dynamic system. (4 pages)