高级检索
    李密青 郑金华 罗 彪. 一种基于最小生成树的多目标进化算法[J]. 计算机研究与发展, 2009, 46(5): 803-813.
    引用本文: 李密青 郑金华 罗 彪. 一种基于最小生成树的多目标进化算法[J]. 计算机研究与发展, 2009, 46(5): 803-813.
    Li Miqing, Zheng Jinhua, and Luo Biao. A Multi-Objective Evolutionary Algorithm Based on Minimum Spanning Tree[J]. Journal of Computer Research and Development, 2009, 46(5): 803-813.
    Citation: Li Miqing, Zheng Jinhua, and Luo Biao. A Multi-Objective Evolutionary Algorithm Based on Minimum Spanning Tree[J]. Journal of Computer Research and Development, 2009, 46(5): 803-813.

    一种基于最小生成树的多目标进化算法

    A Multi-Objective Evolutionary Algorithm Based on Minimum Spanning Tree

    • 摘要: 怎样保证朝Pareto最优解的方向搜索和如何获得均匀分布且范围广泛的非支配解是多目标进化算法(MOEA)设计时的两个关键问题,它们很大程度上取决于适应度赋值和外部种群维护这两个重要部分.提出了一种基于最小生成树的多目标进化算法(MST_MOEA).在考虑了个体间支配关系的基础上,利用个体与非支配集的距离和不同等级个体的树聚集密度来对适应度赋值;在外部种群的非支配解个数超过规定的种群规模时,用树的度数和树聚集密度对其进行修剪.将其应用于不同维数下9个测试函数,并与NSGA-II,SPEA2进行对比,结果证实了算法良好的收敛性和分布性.

       

      Abstract: Fitness assignment and external population maintenance are the two critical parts of multi-objective evolutionary algorithms (MOEA). They affect the two basic objectives (convergence and distribution) of MOEA design to a great extent. It is difficult for many existing MOEAs to obtain satisfying results. In this paper, a multi-objective evolutionary algorithm based on minimum spanning tree (MST_MOEA) is proposed. In the algorithm, a density estimation metric, tree crowding density, is defined. And on the basis of Pareto dominance relationship, the algorithm assigns fitness using two factors, the distance between the current individual and non-dominated solutions set and the tree crowding density of different rank individual. Moreover, if the size of non-dominated solutions set exceeds that of population, the degree information of solution combined with tree crowding density is employed to truncate population. Particularly, the solutions which have higher degree will be eliminated, for the degree in minimum spanning tree represents not only the density of solution, but the situation in population to a certain extent. Finally, the proposed algorithm is applied to nine test functions on different dimensions and has an extensively comparative study with NSGA-II and SPEA2. Experimental results demonstrate that the proposed algorithm indicates good performance in convergence and distribution.

       

    /

    返回文章
    返回