Hybrid Differential Evolution Algorithm Based On Population Reset And Double Subpopulation

Nonlinear optimizations often appear in science and engineering fields, have a high degree of nonlinearity and multimodality in general. Designing effec-tive algorithm to solve this kind of problem can promote the progress of science and technology effectively. However, traditional optimization methods require the continuity or differentiability of objective function, but not satisfy in prac-tice. Thus they couldn't meet the requirement of practical problems. Since the fifty's of last century, many heuristic algorithms based on swarm optimization are proposed, those include differential evolution(DE), particle swarm optimiza-tion(PSO), firefly algorithm(FA) and so on. These algorithms are free to analytic property of objection function and has been successfully applied to solve a va-riety of optimization problems due to search efficiency, less parameters settings and easy operation.As a simple and effective heuristic algorithm, differential evolution has the advantage of simple structure and less control parameters. But it still has some shortcomings such as slow convergence and falling into local optimum easily when applied in solving complex optimization problems. In order to improve the performance of differential evolution, two improved algorithms are proposed in this thesis.1. Based on one step K-means clusters and particle swarm optimization, this thesis presents an hybrid differential evolution algorithm to improve low speed and help individuals avoid trapping into local optimum. The velocity updating formula of particle swarm optimization is modified by one step K-means cluster-ing algorithm to enhance the ability of escaping from local optimum. To balance the ability of global and local search ability, then the improved particle swarm optimization algorithm are combined with differential evolution algorithm by means of a linear decreasing selective probability. Finally, some individuals with poor performance are reset to maintain population diversity and help individuals escape from local optimum under mild condition.2. To improve further, a dynamic subpopulation differential evolution algo-rithm with adaptive control parameters is designed. The population is divided into elite subpopulation and ordinary subpopulation according to their fitness value. For the individuals belong to elite subpopulation, they are perturbed by d-ifference vectors selecting randomly to increase the probability of searching glob-al optimum, while the others are guided by elite individuals to prevent the al-gorithm from trapping into local optimum. Moreover, to balance the exploration and exploitation ability of the algorithm, different scaling factor and crossover probability are used for each individuals in the two subpopulation, and the con-trol parameters are adjusted adaptively according to evolution situation.Numerical experiments on several benchmark functions show that the pro-posed algorithms have rapid convergence speed, strong search ability and good robustness.