GPU based High-Efficiency PSO Algorithm with Initialization strategy and Thread Self-adaption

Abstract

Abstract Particle Swarm Optimization (PSO) is a widely employed heuristic-based method that effectively addresses a variety of optimization problems owing to its simplicity and robustness. Despite its advantages, PSO's extensive computational demands can hinder its practical applications. As parallel computing and Graphics Processing Units (GPUs) have advanced, researchers have explored these technologies to overcome the computational efficiency limitations of PSO. This paper presents a High-Efficiency PSO (HEPSO) algorithm, which optimizes the PSO process within a GPU-based architecture. The HEPSO algorithm enhances GPU computational performance by: 1) transferring the data initialization process from CPUs to GPUs, minimizing the I/O overhead resulting from repetitive data migration during computation. 2) implementing a self-adaptive thread management strategy to improve algorithm execution efficiency. To evaluate the efficacy of HEPSO, we conducted experiments using four benchmark optimization functions. Our findings indicate that the time speedup ratio of HEPSO compared to GPU-PSO exceeds sixfold. Furthermore, when assessing the time required for function convergence, HEPSO outperforms GPU-PSO, necessitating only 1/3 of the time in most instances.