Cluster Identification Techniques in Genetic Algorithms for Multimodal Optimization

This work proposes techniques in sharing- enhanced genetic algorithms to identify regions where designs cluster in a multimodal design space. A crowdedness function is defined to represent the degree of crowding in the neighborhood of a design. Crowdedness function values of designs in a population are used as criteria for the identification of a design cluster if it has formed. An automatically determined radius of the hyperspherical cluster is then used as both the sharing radius and the radius of mating restriction. Cluster identification techniques in genetic algorithms not only will increase the possibilities of locating more relative optima but also will speed up the convergence rate of located optima. Two illustrative multimodal function minimization problems are used as benchmarks to test the proposed techniques.