A Heuristic Mutation Operator for Evolutionary Neural Network

Abstract

Weight adaptation, node deletion and node addition are three key types of mutation operations for the evolutionary neural network(ENN). The determination of mutation rate and the selection of mutation type are two important issues for evolution, and they have a crucial impact on the performance of ENN. In order to improve the convergence speed and classification accuracy of ENN, a heuristic mutation operator (HMO), which evolves connection weights and network structure simultaneously, was proposed. An adaptive mutation rate is applied in the mutation operator, and the mutation type is selected heuristically from weight adaptation, node deletion and node addition. When the population is not evolved continuously for many generations, in order to jump from the local optima and extend the search space, the mutation rate will be increased and the mutation type will be changed. The experimental results with three classification problems show that the HMO achieves better performance than the traditional mutation operator (TMO) in terms of convergence speed and classification accuracy.