The Optimized Implementation Of Island-Based Genetic Algorithm On CUDA

With the GPU's (Graphics Processing Unit, GPU) rapid development, the image processing, computer simulation and other areas of development has been a strong advance. The same time, people have started to use the power of GPU's parallel computing ability to solve some practical problems, more and more people began to participate as well as GPU-based general-purpose GPU computing to the study of the subject. Among these, the parallel algorithm of research is a hot topic.In this paper, the goal is learning and research general-purpose large-scale parallel computing based on CUDA, and we chose the island-based genetic algorithm to study. Genetic Algorithm is an imitation of the law of survival of the fittest in the biosphere, and it is a randomized search method. The algorithm's characters:operating on the structure of objects, but restricted by the continuity of the function and the derivative; good internal implicit parallelism and excellent global optimization; using the probability of the parade and guidance on ways to automatically optimize the search space, do not specifically required to adjust the search direction, it has been widely used in combinatorial optimization, function optimization, genetic coding, machine learning and artificial life and other fields is the calculation of the moment the smart key technology. Although the genetic algorithm can efficiently solve many areas of many practical problems, but its execution time has become a major factor limiting its use, especially in the task of solving some huge amount of problems, genetic algorithms have running time long. However, through parallelization the process of fitness evaluation, the problem that the algorithm running time is too long has been resolved so that you can greatly reduce the running time and improve the efficiency of genetic algorithm.In this paper, we studied and analyzed island-based genetic algorithm on the CUDA architecture, and implement the algorithm:first, the entire population is divided into several sub-populations, each sub-population is isolated from each processor on the GPU; these distributed sub-population can evolve themselves to achieve the optimal state; then use the transport operator to mix the good characters showed by each sub-population. In this paper, we tested the algorithm on CPU and GPU, the island-based genetic algorithm accelerated by CUDA has an advance on speed and quality results.