高级检索

    王学武, 谢祖洪, 周昕, 顾幸生. 基于统计信息反馈的分步多目标优化[J]. 华东理工大学学报(自然科学版), 2022, 48(5): 665-676. DOI: 10.14135/j.cnki.1006-3080.20210427004
    引用本文: 王学武, 谢祖洪, 周昕, 顾幸生. 基于统计信息反馈的分步多目标优化[J]. 华东理工大学学报(自然科学版), 2022, 48(5): 665-676. DOI: 10.14135/j.cnki.1006-3080.20210427004
    WANG Xuewu, XIE Zuhong, ZHOU Xin, GU Xingsheng. A Stepwise Multi-Objective Evolutionary Optimization Algorithm Based on Statistical Feedback Information[J]. Journal of East China University of Science and Technology, 2022, 48(5): 665-676. DOI: 10.14135/j.cnki.1006-3080.20210427004
    Citation: WANG Xuewu, XIE Zuhong, ZHOU Xin, GU Xingsheng. A Stepwise Multi-Objective Evolutionary Optimization Algorithm Based on Statistical Feedback Information[J]. Journal of East China University of Science and Technology, 2022, 48(5): 665-676. DOI: 10.14135/j.cnki.1006-3080.20210427004

    基于统计信息反馈的分步多目标优化

    A Stepwise Multi-Objective Evolutionary Optimization Algorithm Based on Statistical Feedback Information

    • 摘要: 传统的多目标进化算法通常协同考虑解的分布性和收敛性,在搜索初期会生成大量的支配解,造成计算资源的浪费,甚至导致算法不收敛。本文提出了一种基于统计信息反馈的分步多目标优化算法。将算法分为单目标探索阶段、单目标到多目标的过渡阶段、群体划分局部优化阶段3个阶段,根据每个阶段的性质设计任务和策略,以增强算法的收敛性和分布性。在第二、三阶段中,根据目标函数值将解划分为不同群体,分别对不同区域解的信息进行统计分析,再根据反馈统计信息指导亲本选择过程,改善解的分布性和收敛性。在DTLZ和WFG系列问题上进行测试,并与其他多目标进化算法进行了比较,实验结果验证了本文算法在复杂、难收敛问题上的优势。

       

      Abstract: Since convergence and diversity are taken into consideration cooperatively in the whole iteration process, the traditional multi-objective optimization algorithms will generate a large number of dominated solutions in the early stage of search, which will inevitably result in the waste of computing resources, even the non-convergence of solving procedure. Aiming at the above shortcoming, this paper proposes a stepwise multi-objective optimization algorithm based on statistical information feedback (SFI-SMOEA), which is composed of three steps, i.e., the optimal value exploration for each objective, rough search for Pareto front, local optimization stage with group division. According to the feature of each step, the corresponding task and strategy is assigned to promote convergence and diversity. In the 2nd and 3rd steps, the solutions are divided into different groups according to the objective function, the information of different regional solutions is statistically analyzed, and then, the statistical feedback information (SFI) of different regional solutions is utilized to guide the parent-selection process. Thus, the distribution and convergence of the solutions can be controlled more precisely. Finally, the proposed algorithm is tested on DTLZ and WFG series issues, and compared with other multi-objective optimization algorithms, from which it is illustrated that the proposed algorithm has the advantages of solving complex and difficult convergence problems.

       

    /

    返回文章
    返回