A fuzzy controller that learn from past performance

Machine learning in control systems is a very important problem that has been investigated by many researches. This paper presents a method to implement learning in a fuzzy controller where the learning is based on the past performance of the controller against a disturbance. A second fuzzy system is used to evaluate the control performance in terms of rise time and overshoot. The performance is quantitatively expressed as a fuzzy performance index.. The reinforcement type of learning mechanism employed in the controller uses the performance index as a feedback. After a disturbance, consequence values of the fuzzy rules which were activated during the transient are adjusted based on the fuzzy performance index and other information such as truth values of the rules and process errors. The working of the learning process is illustrated through a simulation example of a chemical reactor.