GPU based High-efficiency PSO Algorithm with Initialization strategy and Thread Slef-adaption

Abstract

Abstract Particle Swarm Optimization (PSO) is one of the most commonly heuristics-based methods that has been used to solve various optimization problems due to its simplicity and robustness. However, when comes to practical applications, it requires a huge computational cost. With the development of parallel computing and Graphics Processing Unit (GPU) calculating, many researchers have tried taking these techniques to break down the obstacle of computational efficiency. It is a challenging problem for the long-term application of PSO. In this paper, we propose a HEPSO algorithm that focuses on the procedure optimization of PSO in GPU-based architecture. It optimizes the GPU computation process from two following aspects: 1) Migrate the data initialization procedure from CPUs to GPUs to reduce the huge IO loss caused by repeating migration while the computing process. 2) Employ a self-adaptive thread management strategy to improve the algorithm execution efficiency. Moreover, we use four benchmark optimization functions to test the efficiency of our HEPSO. The experiment results show that the time speedup ratio between HEPSO and GPU-PSO can exceed 6 times. Meanwhile, when we evaluate the performance of HEPSO with the time consumption for functions converge, HEPSO only needs 1/3 time of GPU-PSO in most cases.