Evolution of fuzzy nearest neighbor neural networks

Proposes a genetic algorithm-based approach to the design of compact fuzzy rule-based classification systems. In our approach, a fuzzy IF-THEN rule is generated by assigning a circular cone-type membership function and a certainty grade to each training pattern. Thus, each fuzzy IF-THEN rule can be viewed as a kind of nearest-neighbor classifier, which has its own certainty grade as well as its own localized receptive field specified by the radius of the circular cone-type membership function. A genetic algorithm is employed for selecting a small number of training patterns that are used for generating fuzzy IF-THEN rules. Our genetic algorithm has three objectives: to minimize the error rate, the rejection rate and the number of fuzzy IF-THEN rules. We also show that the fuzzy rule-based classification system constructed by the genetic algorithm can be represented by a neural network architecture that is similar to nearest-neighbor neural networks.