Genetic algorithms with sharing for multimodal function optimization

Abstract

Many practical search and optimization problems require the investigation of multiple local optima. In this paper, the method of sharing functions is developed and investigated to permit the formation of stable subpopulations of different strings within a genetic algorithm (CA), thereby permitting the parallel investigation of many peaks. The theory and implementation of the method are investigated and two, one-dimensional test functions are considered. On a test function with five peaks of equal height, a GA without sharing loses strings at all but one peak; a GA with sharing maintains roughly equally sized subpopulations clustered about all five peaks. On a test function with five peaks of different sizes, a GA without sharing loses strings at all but the highest peak; a GA with sharing allocates decreasing numbers of strings to peaks of decreasing value as predicted by theory.