GPU-Based Parallel Multi-objective Particle Swarm Optimization

In the recent years, multi-objective particle swarm optimization (MOPSO) has become quite popular in the field of multi-objective optimization. However, due to a large amount of fitness evaluations as well as the task of archive maintaining, the running time of MOPSO for optimizing some difficult problems may be quite long. This paper proposes a parallel MOPSO based on consumer-level Graphics Processing Units (GPU), which, to our knowledge, is the first approach of optimizing multi-objective problems via PSO on the platform of GPU. Experiments on 4 two-objective benchmark test functions are conducted. Compared with the CPU based sequential MOPSO, our GPU based parallel MOPSO is much more efficient in terms of running time, and the speedups range from 3.74 to 7.92 times. When the problem is large-scale, i.e. the dimension of the decision vector is large, the speedups can be bigger than 10 times. Furthermore, the experimental results show that the larger the size of the swarm is, the more nondominated solutions are found, the higher the quality of solutions are, and the bigger the speedup is.