A Method to Design a Neural Network by the Genetic Algorithm with Partial Fitness

Abstract

In this paper, a method is described which improves a performance in learning by introducing partial fitness (PF) into the genetic algorithm (GA). The method divides a chromosome in the GA into several parts, the PFs of which are evaluated. Each part in chromosomes independently performs a selection and a crossover. Such a technique improves the perfomance in learning of the GA. This paper applies the method to a rotated coin recognition problem to examine its effectiveness. As a result, it is shown that it is better than the conventional GA on convergence in learning, makes a smaller network in size, and gives better generalization ability.