Three alternatives for parallel GPU-based implementations of high performance particle swarm optimization

Abstract

Particle Swarm Optimization (PSO) is heuristics-based method, in which the solution candidates of a problem go through a process that simulates a simplified model of social adaptation. In this paper, we propose three alternative algorithms to massively parallelize the PSO algorithm and implement them using a GPGPU-based architecture. We aim at improving the performance of computationally demanding optimizations of many-dimensional problems. The first algorithm parallelizes the particle's work. The second algorithm subdivides the search space into a grid of smaller domains and distributes the particles among them. The optimization subprocesses are performed in parallel. The third algorithm focuses on the work done with respect to each of the problem dimensions and does it in parallel. Note that in the second and third algorithms, all particles act in parallel too. We analyze and compare the speedups achieved by the GPU-based implementations of the proposed algorithms, showing the highlights and limitations imposed. © 2013 Springer-Verlag Berlin Heidelberg.