Genetic Optimization of Fuzzy Networks

A novel fuzzy network controller is introduced which is interesting from both a theoretical and a practical viewpoint. It is similar to a radial basis function neural network, contains structured information and may be characterized by a few parameters only. For training of these networks with experiments or by examples, a nonstandard genetic algorithm is applied, using a real-valued parameter encoding scheme and an appropriate cross-over. The adaptation of a direct fuzzy controller for a simple system illustrates the procedure. In a second example the integrated design and optimization approach is shown for a typical industrial controller stabilizing a laboratory size magnetic levitation system. It includes nonlinear components for fuzzy anti-windup.