A Novel Multi-population Particle Swarm Optimization with Learning Patterns Evolved by Genetic Algorithm

Abstract

In recent years, particle swarm optimization (PSO) and genetic algorithm (GA) have been applied to solve various real-world problems. However, the original PSO is based on single population whose learning patterns (inertia weights, learning factors) has no potentials in evolution. All particles in the population interact and search according to a fixed pattern, which leads to the reduction of population diversity in the later iterations and premature convergence on complex and multi-modal problems. Therefore, a novel multi-population PSO with learning patterns evolved by GA is proposed to improve diversity and exploration capabilities of populations. Meanwhile, the local search of PSO particles which start in the same position also evolved by GA independently maintains exploitation ability inside each sub population. Experimental results show that the accuracy is comparable and our method improves the convergence speed.