A GPU Implementation of MOEA/D-ACO for the Multiobjective Traveling Salesman Problem

Several Ant Colony Optimization (ACO) algorithms have been proposed to solve multiobjective problems. MOEA/D-ACO is a multiobjective algorithm based on ACO and the decomposition approach and presents best results when compared with MOEA/D and BiCriterion on discrete optimization problems. This paper proposes a new parallel implementation of MOEA/D-ACO for execution on the Graphics Processing Unit (GPU) using NVIDIA CUDA in order to improve the efficiency, reaching high quality results in a reasonable execution time. Based on recent researches, both the solution construction and pheromone update phases are implemented using a data parallel approach. We use the multiobjective Travelling Salesman Problem (MTSP) as application. We report speedups of up to 11x from the sequential implementation. Moreover, the results point out that: the number of objectives and the number of sub problems affect directly the speedup.