• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2017, Vol. 53 ›› Issue (13): 170-178.doi: 10.3901/JME.2017.13.170

• 数字化设计与制造 • 上一篇    下一篇

集成最小化置信下限和信赖域的动态代理模型优化策略

曾锋1,2, 周金柱2,3   

  1. 1. 中国船舶重工集团公司第七二二研究所 武汉 430079;
    2. 西安电子科技大学电子装备结构设计教育部重点实验室 西安 710071;
    3. 大连理工大学工业装备结构分析国家重点实验室 大连 116024
  • 出版日期:2017-07-05 发布日期:2017-07-05
  • 作者简介:

    曾锋,男,1992年出生。主要研究方向为基于代理模型的多学科优化。

    E-mail:zf576039783@163.com

    周金柱(通信作者),男,1979年出生,博士,副教授,硕士研究生导师。主要研究方向为智能蒙皮天线的设计与制造和基于自适应代理模型的多学科优化。

    E-mail:xidian_ jzzhou@126.com

  • 基金资助:
    * 国家自然科学基金(51305323,51490660,11403089)、陕西省自然科学基金(2016JM5002)、中央高校基本科研业务费专项资金(JB140404)和工业装备结构分析国家重点实验室开放基金(GZ15110)资助项目; 20160511收到初稿,20161203收到修改稿;

Optimization Strategy for Dynamic Metamodel Integrating Minimize Lower Confidence Bound and Trust Region

ZENG Feng1,2, ZHOU Jinzhu2,3   

  1. 1. The 722th Research Institute, China Shipbuilding Industry Corporation, Wuhan 430079
    , 2. Key Laboratory of Electronic Equipment Structure Design of Ministry of Education, Xidian University, Xi’an 710071
    , 3. State Key Laboratory of Structural Analysis for Industrial Equipment, Dalian University of Technology, Dalian 116024
  • Online:2017-07-05 Published:2017-07-05

摘要:

代理模型广泛应用于工程优化领域中。提出一种集成最小化置信下限和信赖域的动态Kriging代理模型优化策略,以提高全局收敛性和优化效率。在该优化策略中,利用Maximin拉丁超立方体试验设计方法选取初始样本点建立Kriging代理模型,令置信下限公式中的平衡常数等于已有样本点间的最小距离,并采用遗传算法对置信下限公式进行优化。根据已知信息更新信赖域,在新的信赖域内选取样本点更新代理模型,直至收敛。将该策略应用于数学测试算例和工字梁设计优化实例中,试验对比结果表明该优化策略不仅可以获得最优解,而且能够显著地提高优化效率。

关键词: 动态代理模型, 信赖域, 最小化置信下限, Kriging

Abstract: The metamodel model is widely used in engineering optimization. an optimization strategy for dynamic metamodel by integrating minimize lower confidence bound and trust region into Kriging metamodel optimization is proposed, in order to enhance global convergence and optimization efficiency. In this strategy, the initial sampling points are firstly selected by maximin Latin hypercube design method and the Kriging metamodel is constructed. During the optimization process, the equilibrium constant is equal to the minimal Euclidean distance between current sampling points, and then genetic algorithm is employed to optimize current equation of lower confidence bound. Subsequently, the trust region is updated according to the current known information, and the new sampling point in the trust region is added to update the metamodel until the potential optimum is satisfied the convergence conditions. Finally, the optimization strategy is validated by using several numerical benchmark problems and the I-beam design optimization problem. Comparing with other optimization strategies, the proposed optimization strategy can not only obtain the optimal solution, but also improve significantly the optimization efficiency.

Key words: dynamic metamodel, Kriging, minimize lower confidence bound, trust region