LI Juwen,WU Zezhong.Comparison of BFGS and DFP Quasi-Newton Method based on Armijo Search Step[J].Journal of Chengdu University of Information Technology,2021,36(05):558-563.[doi:10.16836/j.cnki.jcuit.2021.05.014]
基于Armijo搜索步长的BFGS与DFP拟牛顿法的比较研究
- Title:
- Comparison of BFGS and DFP Quasi-Newton Method based on Armijo Search Step
- 文章编号:
- 2096-1618(2021)05-0558-06
- Keywords:
- unconstrained optimization; BFGS Quasi-Newton method; DFP Quasi-Newton method; Armijo search
- 分类号:
- O221.2
- 文献标志码:
- A
- 摘要:
- 拟牛顿法是求解无约束优化问题的重要方法,采用非精确Armijo准则确认搜索步长,其中初始点的选取采用两种不同的方法:利用MATLAB工具箱中的rand命令对BFGS和DFP两种算法的初始点进行随机选取; 固定选择两个不同的初始点。讨论不同的初始点选取方法对两种算法收敛效率及结果的影响,最后对两种算法收敛效果进行比较研究。结果表明:在多项式函数中,初始点的选取方法对DFP法的收敛效率有一定影响,在低次函数中,DFP法收敛效率更好,在高次函数中,使用BFGS法的收敛效果更好; 在非多项式函数中,随机取点对计算结果有一定影响,选择离极小点近的点作为初始点得到的最小值更好,并且使用BFGS法的收敛速度更快。
- Abstract:
- Quasi-Newton method is an important method to solve unconstrained optimization problems. In this paper, an inexact Armijo criterion is used to confirm the search step size, and two different methods are used to select the initial points: one is to use the rand command in MATLAB toolbox to randomly select the initial points of BFGS and DFP algorithms; the other is to select two different initial points fixedly. The influence of different initial point selection methods on the convergence efficiency and results of the two algorithms is discussed. Finally, the convergence effect of the two algorithms is compared. The results show that in polynomial function, the selection method of initial point has certain influence on the convergence efficiency of DFP method. In lower order function, DFP method has better convergence efficiency. And in higher order function, BFGS method has better convergence effect. In the non-polynomial function, the random selection of points has a certain influence on the calculation results. It is better to select the point near the minimum point as the initial point, and the BFGS method has faster convergence speed.
参考文献/References:
[1] 马昌凤,柯艺芬,谢亚君.最优化计算方法及其MATLAB程序实现[M].北京:国防工业出版社,2015.
[2] 王宜举.非线性最优化理论与方法[M].北京:科学出版社,2012.
[3] Broyden C G.The convergence of a class of double-rank minimization algorithms[J].J.Institute of Mathematics and Its Applications,1970,6(1):76-90.
[4] 刘庆吉,张长海.一种修正的DFP方法[J].大庆石油学院学报,1990(1):100-104.
[5] 张恒,何伟.基于新拟牛顿方程的改进BFGS方法[J].淮海工学院学报(自然科学版),2007,16(3):10-12.
[6] 潘和平,孟庆鑫.地-井时域激电拟牛顿法反演问题研究[J].电波科学学报,2013,28(5):184-191.
[7] 徐耀松,谭亮.基于改进拟牛顿-K近邻的TDOA定位算法[J].传感技术学报,2018,31(10):122-127.
[8] 时伟.基于新拟牛顿法的自适应均衡技术[J].北京邮电大学学报,2012,35(3):86-89.
[9] 张静.关于线搜索的Armijo型方法[C].第十届中国青年信息与管理学者大会论文集.2008.
[10] 张丽,周伟军.Armijo线性搜索下Hager-Zhang共轭梯度法的全局收敛性[J].数学物理学报,2008,28(5):840-845.
[11] McCormick G P.A modification of Armijo’s step-size rule for negative curvature[J].Mathematical Programming,1977,13(1):111-115.
[12] 黄飞,吴泽忠.基于Armijo搜索步长的几种共轭梯度法的分析对比[J].成都信息工程大学学报,2019,34(2):209-215.
相似文献/References:
[1]杨 茜,吴泽忠,贺盛瑜.一类改进的BFGS拟牛顿法及与其他几种拟牛顿法的比较研究[J].成都信息工程大学学报,2023,38(02):227.[doi:10.16836/j.cnki.jcuit.2023.02.015]
YANG Qian,WU Zezhong,HE Shengyu.A Kind of Improved BFGS Quasi-Newton Method and its Comparison with other Quasi-Newton Methods[J].Journal of Chengdu University of Information Technology,2023,38(05):227.[doi:10.16836/j.cnki.jcuit.2023.02.015]
备注/Memo
收稿日期:2021-01-04
基金项目:国家自然科学基金资助项目(71672013)