Genetic-based auto-design of fuzzy controllers

Abstract

This dissertation represents a departure from the conventional design of fuzzy controllers. Two different design approaches are proposed. One is a full-optimization for applications where high performance is critical. The other involves an efficient design approach where fast development is of primary concern. A genetic algorithm, as an optimization technique, is employed to automate and at the same time to optimize the fuzzy controller design process. This optimization requires a predefined performance index. An overview of fuzzy controllers is first presented in which the novel concept of characteristic points is developed. This concept allows one to appreciate the role of each set of fuzzy controller parameters, and leads to the main motivation for automating the design process. An insight into the nature of the problem leads to the suitability of a genetic algorithm, as an appropiate search technique for this automation/optimization. A particular genetic algorithm is coded for the concurrent optimization of controller parameters. This is contrasted with the alternative approach, where controller parameters are optimized sequentially. As an application example, electrical drive systems are considered. A novel perspective on the field oriented control of induction motors is first presented, followed by several possible designs of the fuzzy controllers for such a drive system. In each case, the fuzzy controller is designed using one of the proposed genetic algorithms, and results are compared with those of a conventional counterpart. Also in this dissertation, a novel perspective on the robustness of a fuzzy controller is presented which suggests designing a fuzzy controller based on sliding mode control--a well established robust control scheme. Based on this view, an efficient near-optimal design technique of a fuzzy controller is proposed. For instance, given a 7 x 7 decision table a search space of 84 dimensions collapses into a search space of 7 dimensions. While this is achieved at the expense of decreasing the performance index slightly, it can be employed for a large class of systems where fast tuning of the controller is the primary concern. Furthermore, this approach is not restricted to the genetic-based auto-design of a fuzzy controller.