Digital Pheromone Implementation of PSO with Velocity Vector Accelerated by Commodity Graphics Hardware

Abstract

Listen

Translate article

In this paper, a model for Graphics Processing Unit (GPU) implementation of Particle Swarm Optimization (PSO) using digital pheromones to coordinate swarms within ndimensional design spaces is presented. Particularly, the velocity vector computations are carried out on graphics hardware. Previous work by the authors demonstrated the capability of digital pheromones within PSO for searching n-dimensional design spaces with improved accuracy, efficiency and reliability in serial, parallel and GPU computing environments. The GPU implementation was limited to computing the objective function values alone. Modern GPUs have proven to outperform the number of floating point operations when compared to CPUs through inherent data parallel architecture and higher bandwidth capabilities. This paper presents a method to implement velocity vector computations on a GPU along with objective function evaluations. Three different modes of implementation are studied and presented - First, CPU-CPU where objective function and velocity vector are calculated on CPU alone. Second, GPU-CPU where objective function is computed on the GPU and velocity vector is computed on GPU. Third, GPU-GPU where objective function and velocity vector are both evaluated on the GPU. The results from these three implementations are presented followed by conclusions and recommendations on the best approach for utilizing the full potential of GPUs for PSO.