劉東毅,邵 琛
(1. 天津大學(xué)理學(xué)院,天津 300072;2. 天津大學(xué)電氣與自動化工程學(xué)院,天津 300072;3. 哈爾濱理工大學(xué)應(yīng)用科學(xué)學(xué)院,哈爾濱 150080)
不同參數(shù)βk對應(yīng)不同的算法,如常見的Hestenes-Stiefel(HS)算法[1],F(xiàn)letcher-Reeves(FR)算法[2],Polak-Ribiere-Polyak(PRP)算 法[3]和 Dai-Yuan(DY)算法[4]等.共軛梯度法還需要一維搜索,如廣泛使用的強Wolfe線性搜索算法,即步長kα滿足
為了保證共軛梯度法的全局收斂性,通常需要搜索方向kd滿足下降性質(zhì),即
式中: c>0 ;上標(biāo)T表示矩陣或向量轉(zhuǎn)置.與擬牛頓法不同,共軛梯度法對于不精確線搜索一般不滿足式(5)或式(6).因此,許多學(xué)者致力于研究對于任何線搜索都滿足下降性質(zhì)的算法.Shanno[7]基于 Perry[8]的思想,運用擬牛頓方程和自調(diào)比技術(shù),提出了無記憶擬牛頓法和自調(diào)比共軛梯度法.Touati-Ahmed和Storey[9]組合了 FR算法和 PRP算法,并提出了一些混合共軛梯度法.Liu和 Storey[10]提出了 LS算法.最近的研究進展參見文獻[11-14].
本文應(yīng)用對稱化技術(shù)[15]于 PRP算法,獲得一種對稱的 PRP算法,它們對于任何線搜索都滿足下降性質(zhì)式(5)和式(6).更重要的是,這種思想可以應(yīng)用到其他共軛梯度法中.
按照 Perry[8]的建議,經(jīng)典的 HS共軛梯度法[1]的搜索方向1k+d 寫成
證明 根據(jù) DSPRP算法構(gòu)造,由式(14)~式(16)可知,它產(chǎn)生的搜索方向1k+d 滿足充分下降條件式(6),再由強Wolfe線性搜索式(3)和式(5),可得到
通過一些試驗函數(shù)來驗證DSPRP算法的數(shù)值效果,并與PR+算法[6]作比較來說明DSPRP算法的優(yōu)越性和實用性.試驗函數(shù)見表1.
表1 試驗函數(shù)Tab.1 Test functions
自變量的個數(shù)n分別取為1,000和10,000(除函數(shù)4與6外).初始點由文獻提供.根據(jù)定理3,試驗使用強 Wolfe線性搜索,并且其子程序根據(jù)文獻[16]第3章中的算法編寫.記α0,1為線搜索第1次迭代初始試探步長,α0,k+1為線搜索隨后的第k+1次迭代初始試探步長.在DSPRP算法中,取γ= 10-4,δ= 1 0-4,μ= 0 .1,σk≡ 1.?dāng)?shù)值試驗分成如下4組.
表2 DSPRP算法的數(shù)值結(jié)果Tab.2 Numerical results of the DSPRP algorithm
表3 PR+算法的數(shù)值結(jié)果Tab.3 Numerical results of the PR+ algorithm
從表 2和表 3中可以看出,當(dāng)線搜比較精確時(20.1c= ),DSPRP算法與 PRP+算法不論從迭代次數(shù)還是 nfg的數(shù)值上比較,二者的性能均相差不多,但當(dāng)線搜比較粗糙時(20.9c= ),DSPRP算法要比 PRP+算法的效果好得多.例如:與 PR+1相比,G1和 G2只失敗了3次,而PR+1失敗了8次;與PR+2相比,G3和G4分別失敗了5次和3次,PR+2卻失敗了20次.因此,可以看出DSPRP算法更實用和穩(wěn)健.
筆者利用的對稱化思想也可以應(yīng)用到其他共軛梯度法中,如 HS法和 FR法等.在全局收斂性證明中,迭代矩陣譜的分布情況起了很重要的作用,由此得到的收斂性結(jié)論是=0.在數(shù)值試驗中γ和μ的值取得很小,這主要是為了考察對稱化的效果,而試驗結(jié)果正好驗證其良好的效果.
[1]Hestenes M R,Stiefel E. Methods of conjugate gradients for solving linear systems [J].Journal of Research of the National Bureau of Standards,1952,49(6):409-439.
[2]Fletcher R,Reeves C. Function minimization by conjugate gradients [J].Computer Journal,1964,7(2):149-154.
[3]Polyak B T. The conjugate gradient method in extreme problems [J]. USSRComputation Math and Math Physics,1969,9(4):94-112.
[4]Dai Yuhong,Yuan Yaxiang. A nonlinear conjugate gradient method with a strong global convergence property[J].SIAM Journal on Optimization,1999,10(1):177-182.
[5]Al-Baali M. Descent property and global convergence of the Fletcher-Reeves method with inexact line-search [J].IMA Journal of Numerical Analysis,1985,5(1):121-124.
[6]Gilbert J C,Nocedal J. Global convergence properties of conjugate gradient methods for optimization [J].SIAM Journal on Optimization,1992,2(1):21-42.
[7]Shanno D F. Conjugate gradient methods with inexact searches [J].Mathematics of Operations Research,1978,3(3):244-256.
[8]Perry A. A modified conjugate gradient algorithm [J].Operations Research Technical Notes,1978,26(6):1073-1078.
[9]Touati-Ahmed D,Storey C. Efficient hybrid conjugate gradient techniques [J].Journal of Optimization Theory and Applications,1990,64(2):379-397.
[10]Liu Y,Storey C. Efficient generalized conjugate gradient algorithms [J].Journal of Optimization Theory andApplications,1991,69(1):129-137.
[11]Hager W W,Zhang Hongchao. A new conjugate gradient method with guaranteed descent and an efficient line search [J].SIAM Journal on Optimization,2005,16(1):170-192.
[12]Yu Gaohang,Zhao Yanlin,Wei Zengxin. A descent nonlinear conjugate gradient method for large-scale unconstrained optimization [J].Applied Mathematics and Computation,2007,187(2):636-643.
[13]Zhang Li,Zhou Weijun,Li Donghui. Global convergence of a modified Fletcher-Reeves conjugate gradient method with Armijo-type line search [J].Numerische Mathematik,2006,104(4):561-572.
[14]Andrei N. Another hybrid conjugate gradient algorithm for unconstrained optimization [J].Numerical Algorithms,2008,47(2):143-156.
[15]Powell M J D. A new algorithm for unconstrained optimization [C]//Rosen J B,Mangasarian O L,Ritter K.Nonlinear Programming.New York:Academic Press,1970:31-66.
[16]Nocedal J,Wright S J.Numerical Optimization[M].2nd ed.New York:Springer Science +Business Media,Inc,2006.
[17]Buckley A,Lenir A. QN-like variable storage conjugate gradients[J].Mathematical Programming,1983,27(2):155-175.
[18]Moré J J,Garbow B S,Hillstrom K E. Testing unconstrained optimization software [J].ACM Transactions on Mathematical Software,1981,7(1):17-41.
[19]Grippo L,Lampariello F,Lucidi S A. Truncated Newton method with nonmonotone line search technique for unconstrained optimization [J].Journal of Optimization Theory and Applications,1989,60(3):401-419.
[20]Raydan M. The Barzilai and Borwein gradient method for the large scale unconstrained minimization problem [J].SIAM Journal on Optimization,1997,7(1):26-33.