Retaining diversity of search point distribution through a breeder genetic algorithm for neural network learning

Genetic algorithms (GA) have been used for training of fixed structure neural networks and for optimisation of network structure. The crucial issue of algorithms is their premature convergence that deteriorates the diversity of individual search points. Several techniques have being applied to retain the diversity of the search point distribution. In this paper the application of a breeder genetic algorithm (BGA) for neural network learning is considered as well as the problem of retaining diversity. Truncation selection, extended intermediate recombination, and variable mutation range are proposed. It is shown that the performance of BGA is superior to GA in retaining diversity.