Quality-oriented study on mapping island model genetic algorithm onto CUDA GPU

Genetic algorithm (GA), a global search method, has widespread applications in various fields. One very promising variant model of GA is the island model GA (IMGA) that introduces the key idea of migration to explore a wider search space. Migration will exchange chromosomes between islands, resulting in better-quality solutions. However, IMGA takes a long time to solve the large-scale NP-hard problems. In order to shorten the computation time, modern graphic process unit (GPU), as highly-parallel architecture, has been widely adopted in order to accelerate the execution of NP-hard algorithms. However, most previous studies on GPUs are focused on performance only, because the found solution qualities of the CPU and the GPU implementation of the same method are exactly the same. Therefore, it is usually previous work that did not report on quality. In this paper, we investigate how to find a better solution within a reasonable time when parallelizing IMGA on GPU, and we take the UA-FLP as a study example. Firstly, we propose an efficient approach of parallel tournament selection operator on GPU to achieve a better solution quality in a shorter amount of time. Secondly, we focus on how to tune three important parameters of IMGA to obtain a better solution efficiently, including the number of islands, the number of generations, and the number of chromosomes. In particular, different parameters have a different impact on solution quality improvement and execution time increment. We address the challenge of how to trade off between solution quality and execution time for these parameters. Finally, experiments and statistics are conducted to help researchers set parameters more efficiently to obtain better solutions when GPUs are used to accelerate IMGA. It has been observed that the order of influence on solution quality is: The number of chromosomes, the number of generations, and the number of islands, which can guide users to obtain better solutions efficiently with moderate increment of execution time. Furthermore, if we give higher priority on reducing execution time on GPU, the quality of the best solution can be improved by about 3%, with an acceleration that is 29 times faster than the CPU counterpart, after applying our suggested parameter settings. However, if we give solution quality a higher priority, i.e., the GPU execution time is close to the CPU's, the solution quality can be improved up to 8%.