Neighbor based Parents Selection for Real-coded Genetic Algorithms

Abstract

This paper proposes novel Real-coded Genetic Algorithms, where one search point and its neighbors are selected as crossover parents. This neighbor based parents selection provides Gaussian mixture models approximating underlying distribution of population and its accuracy is depend on the definition of the neighbor. Since from the viewpoint of Estimation of Distribution Algorithms, performance of Genetic Algorithms is depend on the accuracy of the approximation, this paper empirically compares four kinds of neighbor: (1)random neighbor, (2)nearest neighbor, (3)relative neighbor, (4)Gabriel neighbor. Intensive simulation results via multimodal function optimization have revealed that (1) relative neighbor provides the better approximation than other neighbors and (2) using neighbor based parents selection is better than the conventional GAs in terms of optimization performance.