黃玲花
(廣西財(cái)經(jīng)學(xué)院 信息與統(tǒng)計(jì)學(xué)院,廣西 南寧 530003)
?
一個(gè)共軛梯度優(yōu)化方法及其在工程中的應(yīng)用*
黃玲花
(廣西財(cái)經(jīng)學(xué)院 信息與統(tǒng)計(jì)學(xué)院,廣西 南寧530003)
給出一個(gè)三項(xiàng)共軛梯度方法,該方法具有如下特點(diǎn):1)搜索方向在不需要任何線搜索的條件下具有充分下降性;2)搜索方向具有自動(dòng)屬于一個(gè)信賴(lài)域的特點(diǎn);3)新方法不但擁有梯度值信息還擁有函數(shù)值信息;4)方法對(duì)一般函數(shù)擁有全局收斂性.數(shù)值檢驗(yàn)結(jié)果表明新方法更具競(jìng)爭(zhēng)性.
共軛梯度;充分下降;收斂性
(1)
其中f(x):Rn→R是連續(xù)可微函數(shù).無(wú)約束最優(yōu)化問(wèn)題來(lái)源于眾多實(shí)際問(wèn)題,具有廣泛的應(yīng)用背景,此問(wèn)題的求解方法有很多種:牛頓法、擬牛頓法、信賴(lài)域方法和共軛梯度法等等,這些優(yōu)化方法能為求解其他優(yōu)化問(wèn)題提供強(qiáng)有力的理論支持.共軛梯度方法具有結(jié)構(gòu)簡(jiǎn)單、計(jì)算機(jī)存儲(chǔ)量小和高效率的特點(diǎn),因而被廣泛應(yīng)用.該方法的迭代公式是:xk+1=xk+αkdk,k=0, 1, 2,…
其中xk稱(chēng)為第k次迭代點(diǎn),αk>0是由線搜索產(chǎn)生的步長(zhǎng),dk是搜索方向,定義形式為
(2)
其中g(shù)k=▽f(xk)和gk+1=▽f(xk+1)是函數(shù)f(x)在xk和xk+1的梯度值,‖·‖表示歐式向量范數(shù).該方法數(shù)值表現(xiàn)優(yōu)越特別適合大規(guī)模優(yōu)化問(wèn)題,也常常被人們用于實(shí)際的問(wèn)題中,是研究最為熱門(mén)的共軛梯度公式之一.但是PRP方法的收斂性不理想,對(duì)于一般函數(shù)在弱Wolfe-Powell線搜索下的全局收斂性一直沒(méi)有得到證明,是一個(gè)開(kāi)放的問(wèn)題.鑒于此,許多學(xué)者希望發(fā)現(xiàn)數(shù)值表現(xiàn)能與PRP相媲美同時(shí)收斂性質(zhì)又比它優(yōu)越的方法,許多成果可參見(jiàn)文獻(xiàn)[8-15]等.研究發(fā)現(xiàn),該方法在非精確線搜索下對(duì)一般函數(shù)收斂性不好的一個(gè)主要原因在于它不能保證充分下降性,充分下降性是指對(duì)所有k,式(3)的不等式成立
(3)
研究發(fā)現(xiàn)該性質(zhì)在收斂性的理論分析中起著重要作用.基于PRP方法,為保證(3)關(guān)系式的成立,Zhang[16]等給出了一個(gè)三項(xiàng)共軛梯度方法,方向?yàn)?/p>
(4)
(5)
(6)
其中δ∈(0,1/2),σ∈(δ,1),受(6)的啟發(fā),Yuan[17]等給出了下面的共軛梯度方向
(7)
(8)
容易看出新公式不但擁有梯度值還擁有函數(shù)值信息.
算法1(修改的共軛梯度算法)
步驟0:給定x0∈Rn,c∈(0,1),δ∈(0,1/2),σ∈(δ,1)和終止參數(shù)ε>0.令d0=-g0=-▽f(x0),
置k:= 0.
步驟1:若‖gk‖≤ε,停止.
步驟2:尋找滿(mǎn)足(5)和(6)的步長(zhǎng)αk.
步驟3:令xk+1=xk+αkdk,如果‖gk+1‖≤ε,停止.
步驟4:利用公式(8)計(jì)算搜索方向.
步驟5:置k:=k+1,轉(zhuǎn)步驟2.
下面證明修改的三項(xiàng)公式方向具有充分下降性和信賴(lài)域的性質(zhì).
引理1 對(duì)k≥0,修改的三項(xiàng)公式的搜索方向滿(mǎn)足
(9)
(10)
(9)成立.對(duì)于關(guān)系式(10),根據(jù)(8)式,當(dāng)k=0,(10)顯然成立,當(dāng)k≥1時(shí),我們有
因此(10)成立.證畢.
假設(shè)條件(A):i) 水平集Ω={x∈Rn:f(x)≤f(x0)}有界.
ii)f在Ω上有下界且連續(xù)可微,它的梯度g滿(mǎn)足Lipschitz條件,即存在常數(shù)L>0滿(mǎn)足
‖g(x)-g(y)‖≤L‖x-y‖,?x,y∈Ω
(11)
證明:根據(jù)WWP線搜索的關(guān)系式(6)和Lipschitz條件(11),得到
(12)
為了驗(yàn)證算法的有效性,給出數(shù)值檢驗(yàn)結(jié)果,檢驗(yàn)函數(shù)是在工程領(lǐng)域經(jīng)常用到的Benchmark問(wèn)題,這些問(wèn)題列舉如下.
1)Spherefunction.
2)Schwefel'sfunction.
3)Rastriginfunction.
4)Schwefelfunction.
x*=(-420.9678,-420.9678,…,-420.9678),fSch(x*)=0
5)Griewankfunction.
上述的Benchmark問(wèn)題可從下述的網(wǎng)站中找到:http://www.cs.cmu.edu/afs/cs/project/jair/pub/volume24/ortizboyer05a-html/node6.html
實(shí)際計(jì)算中,參數(shù)選取如下:c=0.5,ε=10-5,δ=0.1,σ=0.9,停止準(zhǔn)則采用Himmeblau準(zhǔn)則:
表1 算法1的數(shù)值結(jié)果
從表1結(jié)果可以看出,對(duì)給定的終止條件,算法1能大部分成功求解,這說(shuō)明了方法的有效性.但是對(duì)一些問(wèn)題,數(shù)值表現(xiàn)不是很理想,這說(shuō)明方法具有待改進(jìn)的地方.特別是初始點(diǎn)的選擇方面,有待進(jìn)一步的研究.
上述Benchmark問(wèn)題是工程中的常用問(wèn)題,我們利用這些問(wèn)題進(jìn)行驗(yàn)證方法的有效性,目的是為了證實(shí)算法在實(shí)際領(lǐng)域中具有一定的應(yīng)用背景.當(dāng)然還有其他更多的應(yīng)用問(wèn)題有待進(jìn)一步研究和探討.
本文給出一個(gè)修改的三項(xiàng)PRP方法,證明了方法對(duì)一般函數(shù)在WWP線搜索下具有全局收斂性,并給出了數(shù)值檢驗(yàn)結(jié)果,相對(duì)于通常的PRP方法和三項(xiàng)PRP方法可得出如下結(jié)論:
1)與通常的PRP方法相比較,新方法不但具有充分下降性,還具有信賴(lài)域的特點(diǎn),同時(shí)能保證對(duì)一般函數(shù)在WWP線搜索下的全局收斂性,數(shù)值結(jié)果也具有競(jìng)爭(zhēng)性;
2)與通常的三項(xiàng)PRP方法相比較,新方法具有了信賴(lài)域的特點(diǎn)且對(duì)一般函數(shù)在WWP線搜索下的全局收斂性,同時(shí)擁有梯度值信息和函數(shù)值信息;
3)與文章Yuan[17]相比較,將方法推廣到一般的光滑優(yōu)化方法中,且兩種方法所具有的函數(shù)值信息不同.
[1] Y. Dai , Y. Yuan. A nonlinear conjugate gradient with a strong global convergence properties [J]. SIAM J. Optim, 2000 (10):177-182.
[2] R. Fletcher. Practical Method of Optimization, Vol I: Unconstrained Optimization [M]. 2nd edition, Wiley, New York, 1997.
[3] R. Fletcher , C. Reeves. Function minimization bu conjugate gradients[J].Compute. J, 1964 (7): 149-154.
[4] M. R. Hestenes , E. Stiefel. Method of conjugate gradient for solving linear equations, J, Res [J]. Nat. Bur. Stand, 1952 (49):409-436.
[5] Y. Liu , C. Storey. Effcient generalized conjugate gradient algorithms, part 1: theory[J].Journal of optimization theory and Application ,1992 (69).
[6] E. Polak , G. Ribiere.Note sur la xonvergence de directions conjugees[J]. Rev. Francaise informat Recherche Operatinelle , 3e Annee, 1969 (16).
[7] B. T. Polyak.The conjugate gradient method in extreme problems [J].USSR Comp Math Math Phys,1969 (9): 94-112.
[8] W. W. Hager , H. Zhang. A new conjugate gradient method with guaranteed descent and an e?cient line search [J]. SIAM Journal on Optimization, 2005 (16): 170-192.
[9] W. W. Hager , H. Zhang.Algorithm 851: CGDESCENT, A conjugate gradient method with guaranteed descent[J]. ACM Transactions on Mathematical Software, 2006 (32): 113-137.
[10] G. Li, C. Tang, Z. Wei.New conjugacy condition and related new conjugate gradient methods for unconstrained optimization problems[J]. Journal of Computational and Applied Mathematics, 2007(202):532-539.
[11] Z. Wei, G. Li, L. Qi. New nonlinear conjugate gradient formulas for large-scale unconstrained optimization problems[J].Applied Mathematics and Computation, 2006 (179):407-430.
[12] Z. Wei, G. Li, L. Qi.Global convergence of the PRP conjugate gradient methods with inexact line search for nonconvex unconstrained optimization problems[J].Mathematics of Computation, 2008 (77): 2173-2193.
[13] Z. Wei, S. Yao, L. Liu. The convergence properties of some new conjugate gradient methods[J]. Applied Mathematics and Computation, 2006 (183):1341-1350.
[14] G. L. Yuan. Modified nonlinear conjugate gradient methods with suffcient descent property for large-scale optimization problems[J]. Optimization Letters, 2009 (3): 11-21.
[15] G. L. Yuan , X. W. Lu. A modified PRP conjugate gradient method[J]. Annals of Operations Research, 2009 (166): 73-90.
[16] L. Zhang, W. Zhou, D. Li.A descent modified Polak-Ribi`ere-Polyak conjugate method and its global convergence[J]. IMA Journal on Numerical Analysis, 2006 (26) :629-649.
[17] G. Yuan, Z. Wei, G. Li. A modified Polak-Ribière-Polyak conjugate gradient algorithm for nonsmooth convex programs[J].Journal of Computational and Applied Mathematics, 2014 (255):86-96.
[18] J. Z. Zhang, N. Y. Deng, L. H. Chen, New quasi-Newton equation and related methods for unconstrained optimization[J]. Journal of Optimization Theory and Application, 1999 (102):147-167.
[責(zé)任編輯蘇琴][責(zé)任校對(duì)黃祖賓]
A Conjugate Gradient Optimization Method and Its Applications in Engineer
HUANG Ling-hua
(DepartmentofMathematicsandStatistics,GuangxiUniversityofFinanceandEconomics,Nanning530003,China)
In this paper, a conjugate gradient method is proposed. The given method possess the following features: 1) The search direction possesses the sufficient descent property; 2) The search direction belongs to a trust region;3) The new method has not only the gradient value but also function value;4) The presented method has the global convergence for general functions. Numerical results turns out the new method is more competitive to the normal method.
conjugate gradient;sufficient descent;trust region;convergence
2016-03-20.
廣西財(cái)經(jīng)學(xué)院數(shù)量經(jīng)濟(jì)學(xué)重點(diǎn)實(shí)驗(yàn)室項(xiàng)目開(kāi)放性課題(2014SYS05);國(guó)家自然科學(xué)基金項(xiàng)目資助(11161003).
黃玲花(1965-),女,廣西財(cái)經(jīng)學(xué)院信息與統(tǒng)計(jì)學(xué)院副教授,研究方向:應(yīng)用數(shù)學(xué).
O224
A
1673-8462(2016)02-0063-05