数学物理学报 ›› 2023, Vol. 43 ›› Issue (2): 570-580.

• • 上一篇    下一篇

两个带重启方向的改进 HS 型共轭梯度法

刘鹏杰(),吴彦强*(),邵枫,张艳,邵虎   

  1. 中国矿业大学数学学院 江苏徐州 221116
  • 收稿日期:2022-02-11 修回日期:2022-10-17 出版日期:2023-04-26 发布日期:2023-04-17
  • 通讯作者: 吴彦强,E-mail: wyq1976819@126.com
  • 作者简介:刘鹏杰,E-mail: liupengjie2019@163.com
  • 基金资助:
    国家自然科学基金(72071202);中央高校基本科研业务费专项资金(2017XKQY090);江苏省研究生科研与实践创新计划(KYCX22_2491);中国矿业大学研究生创新计划(2022WLKXJ021);和中国矿业大学数学研究项目(2022DLZD04-203)

Two Extended HS-type Conjugate Gradient Methods with Restart Directions

Liu Pengjie(),Wu Yanqiang(),Shao Feng,Zhang Yan,Shao Hu   

  1. School of Mathematics, China University of Mining and Technology, Jiangsu Xuzhou 221116
  • Received:2022-02-11 Revised:2022-10-17 Online:2023-04-26 Published:2023-04-17
  • Supported by:
    National Natural Science Foundation of China(72071202);Fundamental Research Funds for the Central Universities(2017XKQY090);Postgraduate Research & Practice Innovation Program of Jiangsu Province(KYCX22_2491);Graduate Innovation Program of China University of Mining and Technology(2022WLKXJ021);Teaching and Research Project of CUMT(2022DLZD04-203)

摘要:

共轭梯度法是求解大规模无约束优化的有效方法之一. 该文首先对 Hestenes-Stiefel (HS) 共轭参数改进,再通过引入重启条件及重启方向, 建立两个带重启方向的改进 HS 型共轭梯度法. 第一个方法在弱 Wolfe 线搜索下产生下降方向, 第二个方法独立于任何线搜索得到充分下降性. 常规假设下, 分析并获得两个新方法的全局收敛性. 最后, 数值比对试验结果及性能图显示新方法是有效的.

关键词: 无约束优化, 共轭梯度法, 重启方向, 弱 Wolfe 线搜索, 全局收敛性

Abstract:

The conjugate gradient method is one of the effective methods to solve large-scale unconstrained optimization. In this paper, the Hestenes-Stiefel (HS) conjugate parameter is improved, and then two extended HS-type conjugate gradient methods with restart directions are established by introducing restart conditions and restart directions. The first method produces descent direction under the weak Wolfe line search, and the second one obtains sufficient descent independent of any line search. Under conventional assumptions, the global convergence results of the two proposed methods are analyzed and obtained. Finally, the numerical comparison results and performance graphs show the effectiveness of the new methods.

Key words: Unconstrained optimization, Conjugate gradient method, Restart direction, Weak Wolfe line search, Global convergence

中图分类号: 

  • O221.2