Implementation Techniques for Massively Parallel Multi-objective Optimization

Most real-world search and optimization problems involve more than one goal. The task of finding multiple optimal solutions for such problems is known as multi-objective optimization (MOO). It has been a challenge for researchers and practitioners to find solutions for MOO problems. Many techniques have been developed in operations research and other related disciplines, but the complexity of MOO problems such as large search spaces, uncertainty, noise, and disjoint Pareto curves may prevent the use of these techniques. At this stage, evolutionary algorithms (EAs) are preferred and have been extensively used for the last two decades. The main reason is that EAs are population-based techniques which can evolve a Pareto front in one run. But EAs require a relatively large computation time. A remedy is to perform function evaluations in parallel. But another bottleneck is Pareto ranking in multi-objective evolutionary algorithms. In this chapter, a brief introduction to MOO and techniques used to solve MOO problems is given. Thereafter, a GPGPU-based archive stochastic ranking evolutionary algorithm is discussed that can perform function evaluations and ranking on a massively parallel GPU card. Results are compared with CPU and GPU versions of NSGA-II.