Abstract:

The main motivation for adding learning capabilities to fuzzy rule-based expert systems is the desire to reduce the cost and time of knowledge acquisition. Particularly attractive are in this context learning algorithms with the ability to synthesize rules from past cases already available in the data-base. With the aim of automating knowledge acquisition we present in this article fuzzy classifier systems which integrate a fuzzy rule base, a genetic algorithm and an apportionment of credit function. As a first result we give a variant of the mutation operator which allows us to derive a global convergence result for the genetic algorithm under rather weak assumptions. With this approach a combination of the advantages of genetic algorithms and simulated annealing algorithms is achieved.