Retaining diversity of genetic algorithms for multivariable optimization and neural network learning

Methods to retain diversity of the allele distribution in the search for genetic algorithms (GAs) are presented. The authors seek a technique to prevent premature convergence and refine the performance of GA for use in multivariable optimization and unsupervised learning of neural networks. An integer string representation for chromosomes is defined which is well fitted to this usage. The diversity of each locus and rareness of a chromosome are evaluated based on the distribution of alleles in a population. The fitness of a chromosome is adjusted with the rareness so that rare chromosomes will be likely to survive. Mutation width is introduced to adjust the effect of mutation which can generate rare chromosomes. By dynamically changing mutation width at each locus according to the diversity, prematurity can be avoided while conserving effective convergence. Case studies with problems of neural network pattern matching and unsupervised learning of a neural network which controls an inverted pendulum are discussed