An efficient gradient method with approximate optimal stepsize for large-scale unconstrained optimization

In this paper, we introduce a new concept of approximate optimal stepsize for gradient method, use it to interpret the Barzilai-Borwein (BB) method, and present an efficient gradient method with approximate optimal stepsize for large unconstrained optimization. If the objective function f is not close to a quadratic on a line segment between the current iterate x _k and the latest iterate x _k?1, we construct a conic model to generate the approximate optimal stepsize for gradient method if the conic model is suitable to be used. Otherwise, we construct a new quadratic model or two other new approximation models to generate the approximate optimal stepsize for gradient method. We analyze the convergence of the proposed method under some suitable conditions. Numerical results show the proposed method is very promising.     

A new adaptive Barzilai and Borwein method for unconstrained optimization

In this paper we view the Barzilai and Borwein (BB) method from a new angle, and present a new adaptive Barzilai and Borwein (NABB) method with a nonmonotone line search for general unconstrained optimization. In the proposed method, the scalar approximation to the Hessian matrix is updated by the Broyden class formula to generate an adaptive stepsize. It is remarkable that the new stepsize is chosen adaptively in the interval which contains the two well-known BB stepsizes. Moreover, for the negative curvature direction, a strategy for the choice of the stepsize is designed to accelerate the convergence rate of the NABB method. Furthermore, we apply the NABB method without any line search to strictly convex quadratic minimization. The numerical experiments show the NABB method is very promising.     

解大型无约束优化的新的非单调简单模型BB-TR方法（英文）

The trust region(TR) method for optimization is a class of effective methods.The conic model can be regarded as a generalized quadratic model and it possesses the good convergence properties of the quadratic model near the minimizer.The Barzilai and Borwein(BB) gradient method is also an effective method,it can be used for solving large scale optimization problems to avoid the expensive computation and storage of matrices.In addition,the BB stepsize is easy to determine without large computational efforts.In this paper,based on the conic trust region framework,we employ the generalized BB stepsize,and propose a new nonmonotone adaptive trust region method based on simple conic model for large scale unconstrained optimization.Unlike traditional conic model,the Hessian approximation is an scalar matrix based on the generalized BB stepsize,which resulting a simple conic model.By adding the nonmonotone technique and adaptive technique to the simple conic model,the new method needs less storage location and converges faster.The global convergence of the algorithm is established under certain conditions.Numerical results indicate that the new method is effective and attractive for large scale unconstrained optimization problems.     

A new gradient method via quasi-Cauchy relation which guarantees descent

We propose a new monotone algorithm for unconstrained optimization in the frame of Barzilai and Borwein (BB) method and analyze the convergence properties of this new descent method. Motivated by the fact that BB method does not guarantee descent in the objective function at each iteration, but performs better than the steepest descent method, we therefore attempt to find stepsize formula which enables us to approximate the Hessian based on the Quasi-Cauchy equation and possess monotone property in each iteration. Practical insights on the effectiveness of the proposed techniques are given by a numerical comparison with the BB method.     

Alternate step gradient method*

The Barzilai and Borwein (BB) gradient method does not guarantee a descent in the objective function at each iteration, but performs better than the classical steepest descent (SD) method in practice. So far, the BB method has found many successful applications and generalizations in linear systems, unconstrained optimization, convex-constrained optimization, stochastic optimization, etc. In this article, we propose a new gradient method that uses the SD and the BB steps alternately. Hence the name “alternate step (AS) gradient method.” Our theoretical and numerical analyses show that the AS method is a promising alternative to the BB method for linear systems. Unconstrained optimization algorithms related to the AS method are also discussed. Particularly, a more efficient gradient algorithm is provided by exploring the idea of the AS method in the GBB algorithm by Raydan (1997).

To establish a general R-linear convergence result for gradient methods, an important property of the stepsize is drawn in this article. Consequently, R-linear convergence result is established for a large collection of gradient methods, including the AS method. Some interesting insights into gradient methods and discussion about monotonicity and nonmonotonicity are also given.     

Flow Search Approach and New Bounds for the m-Step Linear Conjugate Gradient Algorithm

A flow search approach is presented in this paper. In the approach, each iterative process involves a subproblem, whose variables are the stepsize parameters. Every feasible solution of the subproblem corresponds to some serial search stages, the stepsize parameters in different search stages may interact mutually, and their optimal values are determined by evaluating the total effect of the interaction. The main idea of the flow search approach is illustrated via the minimization of a convex quadratic function. Based on the flow search approach, some properties of the m-step linear conjugate gradient algorithm are analyzed and new bounds on its convergence rate are also presented. Theoretical and numerical results indicate that the new bounds are better than the well-known ones.     

Convergence of Conjugate Gradient Methods with a Closed-Form Stepsize Formula

Conjugate gradient methods are efficient methods for minimizing differentiable objective functions in large dimension spaces. However, converging line search strategies are usually not easy to choose, nor to implement. Sun and colleagues (Ann. Oper. Res. 103:161–173, 2001; J. Comput. Appl. Math. 146:37–45, 2002) introduced a simple stepsize formula. However, the associated convergence domain happens to be overrestrictive, since it precludes the optimal stepsize in the convex quadratic case. Here, we identify this stepsize formula with one iteration of the Weiszfeld algorithm in the scalar case. More generally, we propose to make use of a finite number of iterates of such an algorithm to compute the stepsize. In this framework, we establish a new convergence domain, that incorporates the optimal stepsize in the convex quadratic case. The authors thank the associate editor and the reviewer for helpful comments and suggestions. C. Labat is now in postdoctoral position, Johns Hopkins University, Baltimore, MD, United States.     

A NEW STEPSIZE FOR THE STEEPEST DESCENT METHOD

The steepest descent method is the simplest gradient method for optimization. It is well known that exact line searches along each steepest descent direction may converge very slowly. An important result was given by Barzilar and Borwein, which is proved to be superlinearly convergent for convex quadratic in two dimensional space, and performs quite well for high dimensional problems. The BB method is not monotone, thus it is not easy to be generalized for general nonlinear functions unless certain non-monotone techniques being applied. Therefore, it is very desirable to find stepsize formulae which enable fast convergence and possess the monotone property. Such a stepsize αk for the steepest descent method is suggested in this paper. An algorithm with this new stepsize in even iterations and exact line search in odd iterations is proposed. Numerical results are presented, which confirm that the new method can find the exact solution within 3 iteration for two dimensional problems. The new method is very efficient for small scale problems. A modified version of the new method is also presented, where the new technique for selecting the stepsize is used after every two exact line searches. The modified algorithm is comparable to the Barzilar-Borwein method for large scale problems and better for small scale problems.     

R-linear convergence of the Barzilai and Borwein gradient method

Combined with non-monotone line search, the Barzilai and Borwein(BB) gradient method has been successfully extended for solvingunconstrained optimization problems and is competitive withconjugate gradient methods. In this paper, we establish theR-linear convergence of the BB method for any-dimensional stronglyconvex quadratics. One corollary of this result is that theBB method is also locally R-linear convergent for general objectivefunctions, and hence the stepsize in the BB method will alwaysbe accepted by the non-monotone line search when the iterateis close to the solution.     

A new subspace minimization conjugate gradient method with nonmonotone line search for unconstrained optimization

A new subspace minimization conjugate gradient algorithm with a nonmonotone Wolfe line search is proposed and analyzed. In the scheme, we propose two choices of the search direction by minimizing a quadratic approximation of the objective function in special subspaces, and state criterions on how to choose the direction. Under given conditions, we obtain the significant conclusion that each choice of the direction satisfies the sufficient descent property. Based on the idea on how the function is close to a quadratic function, a new strategy for choosing the initial stepsize is presented for the line search. With the used nonmonotone Wolfe line search, we prove the global convergence of the proposed method for general nonlinear functions under mild assumptions. Numerical comparisons are given with well-known CGOPT and CG_DESCENT and show that the proposed algorithm is very promising.     

Readily implementable conjugate gradient methods

Conjugate gradient methods have been extensively used to locate unconstrained minimum points of real-valued functions. At present, there are several readily implementable conjugate gradient algorithms that do not require exact line search and yet are shown to be superlinearly convergent. However, these existing algorithms usually require several trials to find an acceptable stepsize at each iteration, and their inexact line search can be very timeconsuming.In this paper we present new readily implementable conjugate gradient algorithms that will eventually require only one trial stepsize to find an acceptable stepsize at each iteration.Making usual continuity assumptions on the function being minimized, we have established the following properties of the proposed algorithms. Without any convexity assumptions on the function being minimized, the algorithms are globally convergent in the sense that every accumulation point of the generated sequences is a stationary point. Furthermore, when the generated sequences converge to local minimum points satisfying second-order sufficient conditions for optimality, the algorithms eventually demand only one trial stepsize at each iteration, and their rate of convergence isn-step superlinear andn-step quadratic.This research was supported in part by the National Science Foundation under Grant No. ENG 76-09913.     

从常步长梯度方法的视角看不可微凸优化增广Lagrange方法的收敛性

增广Lagrange方法是求解非线性规划的一种有效方法.从一新的角度证明不等式约束非线性非光滑凸优化问题的增广Lagrange方法的收敛性.用常步长梯度法的收敛性定理证明基于增广Lagrange函数的对偶问题的常步长梯度方法的收敛性,由此得到增广Lagrange方法乘子迭代的全局收敛性.     

关于外梯度法的步长规则

１．引言 设为Ｒｎ中的一个非空闭凸集,Ｆ（ｘ）为Ｒｎ　Ｒｎ中的一个连续向量函数．变分不等式问题（Ｆ,）就是：找一向量ｘ　使得当 ＝Ｒ时,（１．１）退化成非线性互补问题。在这篇文章中总假定：（Ｈ１） ,这里表示（１．１）的解集;（Ｈ２）Ｆ（ｘ）是单调的,即对,（ｘ－ｙ）（Ｆ（ｘ）－Ｆ（ｘ）－Ｆ（ｙ））． 这类问题出现在工程物理、经济管理等领域,有着极为广泛的应用．因此,其数值解近年来受到重视,提出许多有效算法,见综述［１, ２］．在现有的算法中, Ｋｏｒｐｅｌｅｖｉｃｈ的外梯度法［３］（何炳生称它为投影…     

求解稀疏逻辑回归问题的嵌套BB算法的分裂增广拉格朗日算法

逻辑回归是经典的分类方法,广泛应用于数据挖掘、机器学习和计算机视觉.现研究带有程。模约束的逻辑回归问题.这类问题广泛用于分类问题中的特征提取,且一般是NP-难的.为了求解这类问题,提出了嵌套BB(Barzilai and Borwein)算法的分裂增广拉格朗日算法(SALM-BB).该算法在迭代中交替地求解一个无约束凸优化问题和一个带程。模约束的二次优化问题.然后借助BB算法求解无约束凸优化问题.通过简单的等价变形直接得到带程。模约束二次优化问题的精确解,并且给出了算法的收敛性定理.最后通过数值实验来测试SALM-BB算法对稀疏逻辑回归问题的计算精确性.数据来源包括真实的UCI数据和模拟数据.数值实验表明,相对于一阶算法SLEP,SALM-BB能够得到更低的平均逻辑损失和错分率.     

Numerical experiments on dual matrix algorithms for function minimization

In Ref. 2, four algorithms of dual matrices for function minimization were introduced. These algorithms are characterized by the simultaneous use of two matrices and by the property that the one-dimensional search for the optimal stepsize is not needed for convergence. For a quadratic function, these algorithms lead to the solution in at mostn+1 iterations, wheren is the number of variables in the function. Since the one-dimensional search is not needed, the total number of gradient evaluations for convergence is at mostn+2. In this paper, the above-mentioned algorithms are tested numerically by using five nonquadratic functions. In order to investigate the effects of the stepsize on the performances of these algorithms, four schemes for the stepsize factor are employed, two corresponding to small-step processes and two corresponding to large-step processes. The numerical results show that, in spite of the wide range employed in the choice of the stepsize factor, all algorithms exhibit satisfactory convergence properties and compare favorably with the corresponding quadratically convergent algorithms using one-dimensional searches for optimal stepsizes.     

New algorithms for singly linearly constrained quadratic programs subject to lower and upper bounds

There are many applications related to singly linearly constrained quadratic programs subjected to upper and lower bounds. In this paper, a new algorithm based on secant approximation is provided for the case in which the Hessian matrix is diagonal and positive definite. To deal with the general case where the Hessian is not diagonal, a new efficient projected gradient algorithm is proposed. The basic features of the projected gradient algorithm are: 1) a new formula is used for the stepsize; 2) a recently-established adaptive non-monotone line search is incorporated; and 3) the optimal stepsize is determined by quadratic interpolation if the non-monotone line search criterion fails to be satisfied. Numerical experiments on large-scale random test problems and some medium-scale quadratic programs arising in the training of Support Vector Machines demonstrate the usefulness of these algorithms. This work was supported by the EPRSC in UK (no. GR/R87208/01) and the Chinese NSF grants (no. 10171104 and 40233029).     

A new reduced gradient method for solving linearly constrained multiobjective optimization problems

In this paper, we consider the linearly constrained multiobjective minimization, and we propose a new reduced gradient method for solving this problem. Our approach solves iteratively a convex quadratic optimization subproblem to calculate a suitable descent direction for all the objective functions, and then use a bisection algorithm to find an optimal stepsize along this direction. We prove, under natural assumptions, that the proposed algorithm is well-defined and converges globally to Pareto critical points of the problem. Finally, this algorithm is implemented in the MATLAB environment and comparative results of numerical experiments are reported.     

A Barzilai–Borwein type method for stochastic linear complementarity problems

We consider the expected residual minimization (ERM) formulation of stochastic linear complementarity problem (SLCP). By employing the Barzilai–Borwein (BB) stepsize and active set strategy, we present a BB type method for solving the ERM problem. The global convergence of the proposed method is proved under mild conditions. Preliminary numerical results show that the method is promising.     

The Incremental Gauss-Newton Algorithm with Adaptive Stepsize Rule

In this paper, we consider the Extended Kalman Filter (EKF) for solving nonlinear least squares problems. EKF is an incremental iterative method based on Gauss-Newton method that has nice convergence properties. Although EKF has the global convergence property under some conditions, the convergence rate is only sublinear under the same conditions. One of the reasons why EKF shows slow convergence is the lack of explicit stepsize. In the paper, we propose a stepsize rule for EKF and establish global convergence of the algorithm under the boundedness of the generated sequence and appropriate assumptions on the objective function. A notable feature of the stepsize rule is that the stepsize is kept greater than or equal to 1 at each iteration, and increases at a linear rate of k under an additional condition. Therefore, we can expect that the proposed method converges faster than the original EKF. We report some numerical results, which demonstrate that the proposed method is promising.     

Method of dual matrices for function minimization

In this paper, the method of dual matrices for the minimization of functions is introduced. The method, which is developed on the model of a quadratic function, is characterized by two matrices at each iteration. One matrix is such that a linearly independent set of directions can be generated, regardless of the stepsize employed. The other matrix is such that, at the point where the first matrix fails to yield a gradient linearly independent of all the previous gradients, it generates a displacement leading to the minimal point. Thus, the one-dimensional search is bypassed. For a quadratic function, it is proved that the minimal point is obtained in at mostn + 1 iterations, wheren is the number of variables in the function. Since the one-dimensional search is not needed, the total number of gradient evaluations for convergence is at mostn + 2.Three algorithms of the method are presented. A reverse algorithm, which permits the use of only one matrix, is also given. Considerations pertaining to the applications of this method to the minimization of a quadratic function and a nonquadratic function are given. It is believed that, since the one-dimensional search can be bypassed, a considerable amount of computational saving can be achieved.This paper, supported by the National Science Foundation, Grant No. GP-32453, is based on Ref. 1.