半监督距离度量学习内蕴加速投影梯度算法

考虑求解一类半监督距离度量学习问题. 由于样本集(数据库)的规模与复杂性的激增, 在考虑距离度量学习问题时, 必须考虑学习来的距离度量矩阵具有稀疏性的特点. 因此, 在现有的距离度量学习模型中, 增加了学习矩阵的稀疏约束. 为了便于模型求解, 稀疏约束应用了Frobenius 范数约束. 进一步, 通过罚函数方法将Frobenius范数约束罚到目标函数, 使得具有稀疏约束的模型转化成无约束优化问题. 为了求解问题, 提出了正定矩阵群上加速投影梯度算法, 克服了矩阵群上不能直接进行线性组合的困难, 并分析了算法的收敛性. 最后通过UCI数据库的分类问题的例子, 进行了数值实验, 数值实验的结果说明了学习矩阵的稀疏性以及加速投影梯度算法的有效性.     

混合遗传算法收敛性分析

针对遗传算法爬山能力弱但合局搜索能力强的特点 ,本文将遗传算法嵌入到基入传统优化的拟下降算法中 ,并对算法的拟下降步骤做了一定的改进 ,使得整个算法具有全局收敛性 .本文采用马尔可夫的观点进一步证明了算法的全局收敛性 ,并用极难优化的测试函数给出了数值算例 ,证明了本文算法为一种可行的全局优化算法 .     

D.C.优化的最速下降法的收敛性

Some properties of a class of quasi-differentiable functions(the difference of two finite convex functions) are considered in this paper. And the convergence of the steepest descent algorithm for unconstrained and constrained quasi-differentiable programming is proved.     

最速下降法的一次迭代收敛性的一点注记

推广了最速下降法经过一次迭代到达严格凸二次规划问题的最优解的充分必要条件：初始点可以表示为最优解和Hesse矩阵的一个特征向量之和.证明此条件也是最速下降法经过有限次迭代后到达最优解的充要条件.丰富了最速下降法的理论,有助于更好地认识和理解最速下降法,对相关算法的教学有一定的启发意义.     

拟增生算子方程广义最速下降法的收敛性特征条件

本文给出了广义最速下降法强收敛于定义在一致光滑实Banach空间的真子集上的局部有界拟增生算子的零点的一特征条件.所得的结果推广和统一如徐宗本和蒋耀林等人的相应结果.     

拟增生算子方程的广义最速下降逼近的收敛性

证明了广义最速下降逼近强收敛于定义在一致光滑实Banach空间的真子集上的局部有界拟增生算子的零点的一充要条件,相关的结果处理含ψ-强拟增生算子的非线性方程迭代解的收敛性.所得的结果推广和统一如Xu和Roach,Xu、Zhang和Roach,Chidume,Zegeye和Ntatin,徐宗本和蒋耀林,Chidume,Zhou等人的相应结果.     

方向导数的定义及其在最速下降法中的应用

本文从基础理论拓展到实践,从方向导数讲到最速下降法,意在培养学生利用理论知识解决实际问题的能力.     

半监督距离度量学习的内蕴加速投影梯度算法

考虑求解一类半监督距离度量学习问题.由于样本集(数据库)的规模与复杂性的激增,在考虑距离度量学习问题时,必须考虑学习来的距离度量矩阵具有稀疏性的特点.因此,在现有的距离度量学习模型中,增加了学习矩阵的稀疏约束.为了便于模型求解,稀疏约束应用了Frobenius范数约束.进一步,通过罚函数方法将Frobenius范数约束罚到目标函数,使得具有稀疏约束的模型转化成无约束优化问题.为了求解问题,提出了正定矩阵群上加速投影梯度算法,克服了矩阵群上不能直接进行线性组合的困难,并分析了算法的收敛性.最后通过UCI数据库的分类问题的例子,进行了数值实验,数值实验的结果说明了学习矩阵的稀疏性以及加速投影梯度算法的有效性.     

广义最速下降法逼近拟增生算子方程解的一个特征条件

倪仁兴最近的文章研究了广义最速下降法强收敛于拟增生算子方程解的一特征条件．本文对此进行了修正和改进，给出了一个新的特征条件．所得结果同时改进和推广了一些已有的结果．     

关于广义Φ-增生算子的最速下降法的收敛定理

设E为一致光滑Banach空间,A:E→E为有界次连续广义Φ-增生算子满足,对任意x0∈E,选取m≥1,使得‖x0-x*‖≤m且lim/r→∞Φ(r)>m‖Ax0‖.设{Cn}为[0,1]中数列满足控制条件:i)Cn→0(n→∞);ii)∞∑(n=0)=∞.设{xn}n≥0由下式产生xn+1=xn-CnAxn, n≥0, (@)则存在常数a>0,当Cn相似文献   

Convergence of the steepest descent method for minimizing quasiconvex functions

To minimize a continuously differentiable quasiconvex functionf: ⁿ , Armijo's steepest descent method generates a sequencex ^k+1 =x ^k –t _k f(x ^k ), wheret _k >0. We establish strong convergence properties of this classic method: either , s.t. ; or arg minf = , x ^k andf(x ^k ) inff. We also discuss extensions to other line searches.The research of the first author was supported by the Polish Academy of Sciences. The second author acknowledges the support of the Department of Industrial Engineering, Hong Kong University of Science and Technology.We wish to thank two anonymous referees for their valuable comments. In particular, one referee has suggested the use of quasiconvexity instead of convexity off.     

关于正定矩阵一不等式的简单证明

设A=(aij)是一n阶正定实对称矩阵,本文用代数方法证明了|A|≤a11a22…ann,当且仅当A是对角矩阵时等号成立.且证法简单.     

Convergence analysis of an optimally accurate frozen multi-level projected steepest descent iteration for solving inverse problems

In this paper, we introduce a novel projected steepest descent iterative method with frozen derivative. The classical projected steepest descent iterative method involves the computation of derivative of the nonlinear operator at each iterate. The method of this paper requires the computation of derivative of the nonlinear operator only at an initial point. We exhibit the convergence analysis of our method by assuming the conditional stability of the inverse problem on a convex and compact set. Further, by assuming the conditional stability on a nested family of convex and compact subsets, we develop a multi-level method. In order to enhance the accuracy of approximation between neighboring levels, we couple it with the growth of stability constants. This along with a suitable discrepancy criterion ensures that the algorithm proceeds from level to level and terminates within finite steps. Finally, we discuss an inverse problem on which our methods are applicable.     

High quality preconditioning of a general symmetric positive definite matrix based on its UTU + UTR + RTU-decomposition

A new matrix decomposition of the form A = U^TU + U^TR + R^TU is proposed and investigated, where U is an upper triangular matrix (an approximation to the exact Cholesky factor U₀), and R is a strictly upper triangular error matrix (with small elements and the fill-in limited by that of U₀). For an arbitrary symmetric positive matrix A such a decomposition always exists and can be efficiently constructed; however it is not unique, and is determined by the choice of an involved truncation rule. An analysis of both spectral and K-condition numbers is given for the preconditioned matrix M = U^−T AU⁻¹ and a comparison is made with the RIC preconditioning proposed by Ajiz and Jennings. A concept of approximation order of an incomplete factorization is introduced and it is shown that RIC is the first order method, whereas the proposed method is of second order. The idea underlying the proposed method is also applicable to the analysis of CGNE-type methods for general non-singular matrices and approximate LU factorizations of non-symmetric positive definite matrices. Practical use of the preconditioning techniques developed is discussed and illustrated by an extensive set of numerical examples. Copyright © 1999 John Wiley & Sons, Ltd.