Research On PSO With Adaptation Strategy Mixed

The function optimization problem is one of the basic researches in optimization problems and it is also a research hotpot in the search field in recent decades. The traditional algorithms typically use gradient information or sub-gradient information to solve the problems. When solving problems with the feature of high-dimensional, non-convex or having many local extreme points, its effect is not good. Swarm intelligence optimization algorithms use the transition probability to select and search randomly. It has strong global search ability, fast convergence, high search rate and robustness. It reflects good performance in function optimization problems and it has become a hotspot in the study of optimization methods.Particle Swarm algorithm is a global random search algorithm based on cooperation between swarms. This thesis improves the preference from multiple angles based on PSO, which is reflecting mainly in the following aspects:(1) In order to enhance the convergence ability of Comprehensive Learning Particle Swarm Optimization of the latter part, this thesis proposed an hybrid optimization algorithm based on tabu strategy, named CLPSO+Tabu(CMA-ES). CLPSO+Tabu(CMA-ES) takes Tabu search algorithm as subsequent search operation to improve CLPSO. This thesis introduces the Covariance Matrix Adaptation Evolution Strategy into the Tabu algorithm based on Gaussian distribution and guide the distribution of neighborhood structure to construct new adaptive neighborhood structure and guide the selection of candidate solution to solve the problem of low convergence accuracy and achieve better convergence effect. Experiment results show that, compared with CLPSO, CLPSO + Tabu(CMA-ES) algorithm has better convergence effect on the vast majority of functions.(2) Aiming at solveing the disadvantages that PSO is easy to fall into local optimum and is only suitable for partial functions, the thesis improves particle swarm velocity updating formula from multiple angles based on DE. This thesis introduces different mutation strategies like DE/rand/1, DE/current_to_rand/1, DE/ current_to_best/1 and DE/best/2 into PSO to improve particle swarm velocity updating formula. Experiment results show that DE/best/2 is suitable for unimodal functions, DE/rand/1 is fit for a part of multimodal functions, DE/ current_to_best/1 is fit for rotation functions, DE/current_to_rand/1 is moderate. Different improvements on PSO lead to optimization of performance in different directions.(3) In order to make the algorithm can adaptively select appropriate strategies for different stages of different problems, this thesis constructs velocity updating formula strategies pool with CLPSO and three DBV algorithms. Considering the different impact of the strategies, this thesis introduces adaptive framework to select appropriate strategy adaptively. The algorithm mixes advantages of different strategies and makes the algorithms suitable for solving different optimization problems. Experiments show that using the adaptive framework to select the PSO update formula and selecting suitable algorithm at every stages of the search are good to integrating the advantages of various algorithms and making the performance further improved.