Evolutionary computation in the design of optimum neural controllers

This paper presents a novel technique in which fuzzy and evolutionary techniques are fused for the design of a class of optimum neural controllers. In the proposed technique the attributes of the performance of the closed system, i.e. overshoot, rise time and settling time in response to a step demand are related to the suitability of the controller through fuzzy linguistic rules. De-fuzzification of the resultant fuzzy suitability membership function yields the measure of suitability of the design. This measure is subsequently used in a genetic algorithm, which performs a stochastic search for the optimum parameters of the neural controller in a bounded parameter space. The genetic algorithm spawns a set of controller candidates at every iteration and through successive use of genetic operators systematically eliminates those candidates which yield inferior closed system performance. The procedure ultimately converges to an optimum neural controller that satisfies multiple criteria, which are specified qualitatively. The technique is applied to the design of a neural controller for a mechatronic system.