"An Investigation into the Application of Genetic Algorithms to the Design of Fuzzy Classification Systems"
A review of the current literature on the use of genetic algorithms to develop fuzzy systems is presented. A number of problems and limitations of the techniques currently used are identified. Desirable properties of techniques to automatically develop fuzzy systems are proposed, and areas of interest for research identified.
Several techniques for the automatic design of membership functions and fuzzy rules using genetic algorithms are developed. The first group of techniques focus on methods for automatically obtaining the membership functions for fuzzy systems. The second group of techniques investigate different fuzzy rule representations for enhancing the performance of systems.
The techniques developed are tested on three well known classification problems, to determine their relative performance, and to select the best technique. The results for the best technique are compared with other classification methods such as C4.5 and backpropagation neural networks. The technique developed consistently performed well compared to other techniques, demonstrating the potential of genetic algorithms for designing fuzzy classification systems.
The results indicate techniques that use more flexible and expressive rule representations are more likely to result in better performances. The results for the automated development of membership functions were unexpected, finding little, if any improvement over subjectively developed membership functions. More research is necessary to determine if this is also a property of other classification problems.