摘要
传统的多目标进化算法多是基于Pareto最优概念的类随机搜索算法,求解速度较慢,特别是当问题维度变高,需要群体规模较大时,上述问题更加凸显。这一问题已经获得越来越多研究人员以及从业人员的关注。实验仿真中可以发现,构造非支配集和保持群体多样性这两部分工作占用了算法99%以上的执行时间。解决上述问题的一个有效方法就是对这一部分算法进行并行化改造。本文提出了一种基于CUDA平台的并行化解决方案,采用小生境技术实现共享适应度来维持候选解集的多样性,将多目标进化算法的实现全部置于GPU端,区别于以往研究中非支配排序的部分工作以及群体多样性保持的全部工作仍在CPU上执行。通过对ZDT系列函数的仿真结果,可以看出本文算法性能远远优于NSGA-Ⅱ和NPGA。最后通过求解油品调和过程这一有约束多目标优化问题,可以看出在解决化工应用中的有约束多目标优化问题时,该算法依然表现出优异的加速效果。
Most of the basic multi-objective evolutionary algorithm （MOEA） is one kind of similar random search algorithm base on the concept of the Pareto Optimization, which with the slowly speed. Especially the dimension of the problem becomes larger which needs a larger population, the effectiveness of MOEA become more prominent. This problem has been gaining increasing attention among researchers and practitioners. From the experiments, the procedure to build the non-dominated set and to maintain diversity are time-consuming, more than 99 % of the execution time is used in performing the two procedures. A promising approach to overcome this limitation is to paraUelize these algorithms. In this paper, we propose a parallel MOEA based on Compute Unified Device Architecture （CUDA）. Fitness sharing which was based on niche technology is introduced to enhance population diversity. All steps of the parallel MOEA are performed on Graphics Processing Units （GPU）. That＇s the difference between the previous studies which part of the procedure of building the non-dominated set and maintaining diversity is still executed on the CPU. Through the simulation of ZDT series test function and the analysis results with the NSGA-II and NPGA, it can be seen that the performance of the proposed algorithm is much better than the NSGA-II and NPGA with less run times. Finally, it is applied to constrained multi-objective optimization of gasoline blending, the algorithm shows better performance.
出处
《计算机与应用化学》
CAS
2015年第1期1-8,共8页
Computers and Applied Chemistry
基金
国家自然科学基金项目(U1162202
61222303)
上海市自然科学基金(15ZR1408900)
上海市"科技创新行动计划"研发平台建设项目(13DZ2295300)
上海市重点学科建设项目(B504)资助
关键词
多目标
进化算法
CUDA
GPU
并行计算
Multi-objective
evolutionary algorithms
CUDA
GPU
parallel programming
作者简介
作者简介：胡宾宾（1988-），男，硕士，研究生，E-mail：hubinbinheda@126．com
联系人：祁荣宾，女，博士，副研究员．E-mail：qirongbin@ecust．edu．cn
