Distributed co-evolutionary particle swarm optimization using adaptive migration strategy

Abstract

The performance of conventional particle swarm optimization (PSO) depends on the topology when solving different kinds of problems. A global version PSO (GPSO) may be suitable for unimodal problems but may easily fall into local optima in multimodal problems. While a local version PSO (LPSO) may be good at dealing with multimodal problems but may converge slowly in unimodal problems. In this paper, we propose a co-evolutionary particle swarm optimization (CEPSO) that combines the advantages of both GPSO and LPSO. In CEPSO, two distributed populations are adopted to run GPSO and LPSO respectively. The two populations are with the same population size in the beginning. During the evolutionary process, their performance on the problem being solved will be evaluated and compared, and then an adaptive migration strategy (AMS) is adopted to dynamically control the population size of each population. That is, the worst particle in the poorly-performed population will migrate to the well-performed population. This way, CEPSO can put more computational effort to the population that is more suitable for the problem. We compared CEPSO with conventional GPSO, LPSO, and two other state-of-the-are PSO variants to verify its performance. Experimental results show that CEPSO has the ability to combine advantages of both GPSO and LPSO to well solve both unimodal and multi-modal problems.