维普资讯 http://www.cqvip.com 第32卷增刊 2007年9月 广西大学学报(自然科学版) Journal of Guangxi University(Nat Sci Ed) Vo1.32,SuP Sept..200 7 文章编号:1001—7445(2007)增一0239・05 A hybrid conj ugate gradient method for unconstrained optimization WEI Zeng—xin,DENG Xiao—hong,ZHOU Ya—qun (Co11ege of Mathematics and Information Science。Guangxi University・Nanning 530004・Chi“ ) Abstract:A new hybrid conjugate method based on PRP method is proposed・The con g “ prop y studied. Numerica1 results show that it is effcient. Key w。rds:unc。nstra ned。ptimizati。n;c。njugate gradientemeth。d;w。Ife line search c。nditi。ns;suffic descent property;global convergence・ CLC number:O221.2 Document code:A 1 Introduction The conjugate gradient method is designed to so!Ve the following unconstrained no l;“。 optimization problem min厂( ), e (]) where厂:R一 R is a smooth,nonlinear function whose gradient will be denoted by g ・The iter嚣÷ Ve forn1u1a of the conjugate gradient method is given by z +1一z 十 , (2) where }is a step—length,and d^is the search direction defined by “ 』--gk, 一一l一 士 d 一。, 一U‘ ≥1, (3) , where is a sca1ar and g^denotes g( ^). rhere are s。me ram。us formu1as f。r ilk,such as p^ , p^D ,p 。,p^ ,the detail analysis for these methods are in[3~8]・ ln paper[1],the author proposed an effcient formula for ,which is denoted by , it is iust the same as HS method if exact line search condition is used.For PRP method・though t performs very well with Wolfe line search,but the convergency property still cannot be proVed eVe ’ under the assumption of suffcient descent property㈨ . In[1 0],the author proposed a hybrid conjugate gradient method,in which —m x(0' n1in( , FR)).Another author in[2]modified this hybrid conjugate!nethod and obtain d b numerical result. 1n recent years,some other authors have done much WOtk。n studying COnjuga e gra ent type method and they obtained good resuhsL ・ In this paper,we propose a new hybrid conjugate method,、ve-】se the following formula 。obta in : Received date:2007—04—21 Fundation item:Supported by Guangxi NSF(0542043) Bi。graphy:WEI Zeng-xin(1962一),Male,Zhua“g People,Gua“gxi Province,Profess。r。f Guangxi University 维普资讯 http://www.cqvip.com 240 广西大学学报(自然科学版) 第32卷 -=max(0,min(flkp , )), (5) where +2 ’一 dk-, l lg 一 ll yk—gI—gI一1・ ’ (6) (7) We use both the weak Wolfe and strong line search conditions in this paper.weak Wolfe condition (WWP): 厂(zj+fId )--f(xI)≤ gTdI, (8) (9) g(xI+fIdI) dI≥ag[dI. strong Wolfe condition(SWP): f(xI+fIdI)--f(xI)≤ gTdI, (10) (11) lg(xI+fIdI) dI l≤一 gTdI. described as follows. The new algorithm and two assumptions required about厂(z)is Algorithm 1(New CG method) Step 0 Given z1∈R ,set 1一--gl,k一1.If g1—0,then stop. Step 1 Find ak>O satisfying the WWP. Step 2 Let Xk+I=zI+aMk and gk*!=:g(xk+1).If g¨1=0,then stop. Step 3 Compute by the formula(5)~(7)and generate d¨1 by(3). Step 4 Set k:一点+1,and go to step 1. Assumption A The level set n一{ ∈R l厂(z)≤f(x1)}is bounded. Assumption B The function g(z)is Lipschitz continuous in n,i.e.,there exists a constant L> 0 such that l lg(z)--g(y)ll≤ l1.z— l fIor any z,yE力. This paper is organized as follows.In section 2,we study the descent property of dI generated by new method;section 3 is sacrificed to the convergency property of new method;the numerical result is displayed in the last part. 2 Some property of the new method In[1 2],Gilbert and Nocedal say that some type of conjugate gradient method has a special property named property(*),it is described as follows. Property(*) Consider method of the form(2)~ (3),assume O<y≤l lgI ll≤ ,we say method of the form(2)~(3)has property(*),if there exist some constants 6>1 and y>0 such that the following two conditions are satisfied: II≤ l l≤去. This property is also owned tO the new method in which Lemma 1 Consider method of the form(2)~(3),if (*)holds. Proof Take6一 , = 'where LiS the Lipschits constant.Then is defined by(5)~(7). ll≤6, (12) (13) is defined by(5)~(7),then property l≤l 胩 l≤ If 十『 Ig 一1 )『I g l lg 一1 l 1/2 ===6 (14) Sk一1 ll≤ .We have,by using Assumption B,that: 维普资讯 http://www.cqvip.com 增刊 韦增欣等:求解无约束问题的一个杂交共轭梯度法 241 lf,I ̄<llf,…I≤ 3 Convergence property ≤ 一去. We study the convergence properties of Our new method and give a lemma at first. Lemma 2 Let厂(z)be the object function,suppose assumption A and B hold for厂( ),for the method of the form(2)~(3),if property(*)is owned,line search condition WWP is used,and there exists a positive constant c,such that g—Td^≤一f l lg^l l2, (15) then the method gives lim infl lg^ll一0.The proof of the above Lemma is given in[1 3J・ Theorem 1 Suppose assumption A and B hold for the object function (z),if sufficient descent condition(15)is satisfied,then Algorithm 1 gives lim inf II II一0. The above theorem comes directly from Lemma 1 and Lemma 2.Since strong Wolfe condition is stronger than weak Wolfe condition we can also get convergency conclusion if we replace WWP with SWP in algorithm 1.It is worth tO note that under strong Wolfe line search condition,only the descent assumption of dI is required[1 53 tO establish the convergence. In paper[1 33,Grippo and Lucidi proposed an Armijo type line search method which can ensure the sufficient descent property,it can be described as follows: —max{ r gTd^I d l ,l 一0,1,…}. (16) Grippo and Lucidi proved that if(1 6)is used tO find step size,there exist following relations: f(x +。)≤f(x )--3 。l ld ll。, 一c。(17) l lg…II。 ̄g*+lTd…≤一c。l lg…II。, (18) where r>O; ∈(O,1),O<f1<f2 are some constants.By using(18),we can also get the following relation: ≥ gTd^I d^I1 ’ where e is some constant. (19) Theorem 2 Suppose assumption A and B hold for the object function f( ),if line search condition WWP is replaced by(1 6),then Algorithm 1 gives: lim I Igk II=0. K—・∞ (20) Proof From(17)and(19) 川)≤ , (21) it indicates ≤。。. if the object function is bounded below.So (22) l g ll十I PR …dk一。ll≤ d,ll≤l Ig ll十I …d 一。 ≤lg ll(14- g^一 一1 I I11 d^ 旦)≤II g II+ gL1d^一1 踟II(1q_rL[g^一1 II。 )≤(1+f2rL)ll g^ll, I12 一)≤ (23) 维普资讯 http://www.cqvip.com 242 广两大学学报(自然科学版) 第32卷 it means ≤1 s。we ve 。。> ≥CI 2 (24) It means(20)is founded.Lipschitz continuous condition,(16)and(18)are used in the above proof. 4 Numerical experiment In this section we compare the numerical resuh of new method with both PRP and PRP methods.We chose the parameters as follows: 一0.01, 一0.1.The termination condition is I 1g^Il≤ 10一.In the following table,the columns‘Problem’and‘Dim’represent the name of the test problems and the dimension of the problems respectively.ni/n#ng are the numbers of iterations, function and gradient evaluations respectiveiy,N ,is calculated by using the following formula: N㈨,一nf十5* g, The new method is denoted by’HC—PRP’ Table 1 (25) 维普资讯 http://www.cqvip.com 增刊 韦增欣等:求解无约束问题的一个杂交共轭梯度法 243 Continued Table 1 From the table,we see the advantage of the new method to original ‘PRP’method is obvi0us if SWP is used:there are at least thirteen examples for any of which the tota1 calculation of our new method is some hundreds less than that of original‘PRP method,even more than one thousand times calculation is less for some examplesproposed in this paper is encouraging. . From above numerical analysis, we find our new method References [Ij Dai Y H-Liao L Z-New Conjugacy conditions and related Nonlinear conjugate gradient meth0ds[J].Applied Mathematics Optimization,2001,43:87。101. [2 MO J T・GU N Z,WEI Z X.Hybrid conjugate gradient methods for unconstrained optimizati0n[J]optimizati0n .Methods and Software,2005.1-,1 1. [3] Hestenese M R,Stiefel E.Method of conjugate gradient for solving linear equations[J].J Res Nat Bur Stand, 1952.1 952.(49):,109.436. [43 Fletcher R・Reeves C.Function minimization by conjugate gradients[J].Comput J,1964.(7):149.1 54[5] Polyak E,Ribi ̄ere G.Note sur la convergence de directions conjugees[J].Rev Francaise Inf0rmat Recherche . Operationelle,3eAnnee,1969,(16):35 43. [6] Polyak B T・The conjuga£e gradient methodin extreme problems[n UUSR Comput Math and Math Phvs,1 969, [7] [8] Dai Y H,Yuan Y X・A nonlinear conjugate gradient method with a strong global conve gence pr0perties[J]. SIAM Journal of Optimization,2000,(105:177—182. [9] Powell M J D・Restart pr0ce(1ures for the conjugate gradient method[J].Mathen1atical Pr。grammi“g,1985,(33 5: 241—254. 】n Touati D,Storey C.Efficient hybrid conjugate gradient techniques [J].Journa1 0f Optimizati0n The0rv and AI)plications.1990.(64):379。397. L 1 1 j Dai ’H・Analysis of conjugate gradient methods[D].Institute of Computationa1 Mathematics and Scientific/ Engineering Computing,Chinese Academy of Sciences( Chines 5(1997 5. [1 2 J Gilbe rt J C,Nocedal J・Global conve gence properties of conjugat gradient methods for optimizati0n[J]SIAM J .()ptimizntion,1992,2:21 4'2. L 1 3 Grippo L・Lucidi S,A globally convergent version of the Polyak—Ribiere conjugate gradient meth0d[J]Math .Prog,1 997.(78 5:375-391. (下转第252页) 维普资讯 http://www.cqvip.com 252 广西大学学报(自然科学版) 第32卷 Using complex network to study the evolution of tRNA sequences WEI Fang—ping 。LAN Zhen—xiong (1.College of Physics Science and Technology,Guangxi University,Nanning 530004,China; 2.Department of Information Technology,Guangxi Normal University,Nanning,530023.China) Abstract:We constructed tRNA sequences parallel network and antiparallel networds and investigated the feature of the networks;we found that the networks behave scale—free properties when the similar degree is large.On the other hand.we compared the degree distribution of the tRNA sequences of the whole tRNA group in two type networks,and found that the tRNA sequences are more consanguineous in antiparalle than in parallel network,which seemed consistent with the ideal perfectly that modern tRNA sequences evolved by the complementary duplication.Finally.we presented fl rough evolution model for modern tRNAs based on point mutation theory and the complementary method.The analysis indicated that the network of the tRNAs of the model has fl similar network behavior with the real tRNAs network while.So.the evolution mechanism of the tRNAs should be the mixture of the point mutation and complementary duplication. Key words:tRNA sequences;similar degree;network (责任编辑刘海涛) (I-接第243页) E14] Wei Z X,Li G Y,QI L Q.New nonlinear conjugate gradient formulas for large—scale unconstrained optimization problems EJ3.Applied Mathematics and Computation,2006,(179):407—430. [15] Dai Y H,Yuan Y X.Nonlinear conjugate gradient methods EM 3.Shanghai:Shanghai Scientific and Technical Publishers,(1997)44.45. 求解无约束问题的一个杂交共轭梯度法 韦增欣,邓小红,周亚群 (广西大学数学与信息科学学院,广西南宁530004) 摘要:给出了一个基于PRP方法的新的杂交共轭梯度法,并在适当的条件下,证明了新算法的全局收敛性.数 值结果表明提出的算法是有效的. 关键词:无约束优化;共轭梯度法;Wolfe线搜索;充分下降性;全局收敛性 中图分类号:0221.2 文献标识码:A (责任编辑刘海涛)