A CUDA Implementation of an Improved Decomposition Based Evolutionary Algorithm for Multi-Objective Optimization

Abstract

In last few years, the concept of decomposition has been extensively used in a number of evolutionary algorithms, wherein a multiobjective problem is solved as a set of single objective sub-problems. Such algorithms have demonstrated significant break-through for solving problems with four or more objectives (also referred to as many-objective optimization problems). Along these lines, authors have previously proposed a decomposition based evolutionary algorithm (DBEA). While DBEA is amenable to parallelization, existing implementations of DBEA (and a number of other such algorithms) use several serial components which are designed for single CPU applications. Recently, parallel computing infrastructure has become increasingly affordable, e.g. graphic processing units (GPUs) and application programming interfaces such as compute unified device architecture (CUDA). Hence, there is a significant interest in the research community to redesign such algorithms to exploit the benefits of parallel computing infrastructure. This work presents an improved CUDA based DBEA. The algorithm aims to offer computational savings via parallelization, while maintaining its performance close to existing state-of-the-art sequential implementations. Parallel structure is deployed for population initialization, evaluation, and selection with preemptive association strategies. Performance of the parallel implementation is presented and compared with its sequential counterpart on a number of well established benchmarks to highlight its benefits.