引用本文:刘晓彤,孙超利,王浩,谢刚.两阶段模型协同搜索的昂贵多目标进化优化[J].控制理论与应用,2024,41(9):1676~1684.[点击复制]
LIU Xiao-tong,SUN Chao-li,WANG Hao,XIE Gang.Expensive multi-objective evolutionary optimization with cooperative search of two-stage surrogate models[J].Control Theory and Technology,2024,41(9):1676~1684.[点击复制]
两阶段模型协同搜索的昂贵多目标进化优化
Expensive multi-objective evolutionary optimization with cooperative search of two-stage surrogate models
摘要点击 2800  全文点击 28  投稿时间：2022-12-20  修订日期：2024-07-05
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2023.21095
  2024,41(9):1676-1684
中文关键词  多目标优化  昂贵优化问题  集成模型  协同搜索  填充采样策略
英文关键词  multi-objective optimization  expensive optimization problem  ensemble model  cooperative search  infill sampling strategy
基金项目  国家自然科学基金面上项目(61876123), 山西省重点研发计划项目(202102020101002)资助.
作者单位E-mail
刘晓彤 太原科技大学 xiaotong.liu@stu.tyust.edu.cn 
孙超利* 太原科技大学 chaoli.sun@tyust.edu.cn 
王浩 太原科技大学  
谢刚 太原科技大学  
中文摘要
      近年来, 昂贵多目标优化问题的求解获得了越来越多的关注. 然而, 随着决策空间维度的升高, 模型的有效性和准确性很难保证. 因此, 本文提出了一种两阶段模型协同搜索的昂贵多目标进化优化. 在该方法中, 每轮种群进化前构建全局模型, 以辅助加快对最优解集的搜索. 随后, 利用搜索到的种群选择其邻域样本训练局部模型, 对二者集成辅助算法进行进一步搜索. 最后, 提出基于不确定度的填充采样策略选点, 进行真实评价. 为了验证算法的有效性, 将本文算法与4个算法分别在DTLZ和MaF测试集以及两个实际问题上进行比较, 实验结果表明其具有良好的性能.
英文摘要
      It has been paid more and more attention in recent years to solve expensive multi-objective optimization problems. However, it is challenging to train accurate and efficient models when the dimension of the decision space increases. Thus, expensive multi-objective evolutionary optimization with cooperative search of two-stage surrogate models (EMO-CS) is proposed in this paper for solving expensive problems. In the proposed method, a global model will be trained, before each iteration starts, to assist in speeding up the search for optimal solutions. Then a set of samples in the archive will be found and used to train a local model. The global and local models are used as an ensemble model, whose optimal solutions will be searched for and used to be selected for expensive objective evaluation based on the proposed uncertaintybased sampling criterion. Experimental results show that the proposed method performs better than four state-of-the-art algorithms on DTLZ and MaF test suites and two real-world optimization problems.