A Genetic Algorithm for Expert System Rule Generation

We applied genetic algorithms to fuzzy rule generation to compute expert system rules from data. We have attempted to improve on existing techniques for the automatic generation of fuzzy logic expert system rules with a method we call genetic data clustering (GDC). A genetic algorithm groups training data points by their degree of similarity, and fuzzy logic expert system rules are formed from statistics over the members of these groups. We designed mutation and crossover operators, and an objective function specifically for the task of data clustering. We employed variable mutation probabilities and variable chromosome lengths. We tested the algorithm against two data sets, both large and small, including the Anderson and Fisher Iris data, and a LANDSAT satellite image data set. Overall, the results indicate that the GDC algorithm achieved average performance compared to a suite of classification algorithms found in the literature. The rule representation format was the primary limitation to performance.