小批量随机块坐标下降算法

针对机器学习中广泛存在的一类问题：结构化随机优化问题（其中“结构化”是指问题的可行域具有块状结构,且目标函数的非光滑正则化部分在变量块之间是可分离的）,我们研究了小批量随机块坐标下降算法（mSBD）。按照求解非复合问题和复合问题分别给出了基本的mSBD和它的变体,对于非复合问题,分析了算法在没有一致有界梯度方差假设情况下的收敛性质。而对于复合问题,在不需要通常的Lipschitz梯度连续性假设条件下得到了算法的收敛性。最后通过数值实验验证了mSBD的有效性。     

求解界约束优化的一种新的非单调谱投影梯度法

本文给出求解界约束优化问题的一种新的非单调谱投影梯度算法. 该算法是将谱投影梯度算法与Zhang and Hager [SIAM Journal on Optimization,2004,4（4）：1043-1056]提出的非单调线搜索结合得到的方法. 在合理的假设条件下,证明了算法的全局收敛性.数值实验结果表明,与已有的界约束优化问题的谱投影梯度法比较,利用本文给出的算法求解界约束优化问题是有竞争力的.     

球面上$\ell_1$正则优化的随机临近梯度方法

本文研究球面上的$\ell_1$正则优化问题,其目标函数由一般光滑函数项和非光滑$\ell_1$正则项构成,且假设光滑函数的随机梯度可由随机一阶oracle估计.这类优化问题被广泛应用在机器学习,图像、信号处理和统计等领域.根据流形临近梯度法和随机梯度估计技术,提出一种球面随机临近梯度算法.基于非光滑函数的全局隐函数定理,分析了子问题解关于参数的Lipschtiz连续性,进而证明了算法的全局收敛性.在基于随机数据集和实际数据集的球面$\ell_1$正则二次规划问题、有限和SPCA问题和球面$\ell_1$正则逻辑回归问题上数值实验结果显示所提出的算法与流形临近梯度法、黎曼随机临近梯度法相比CPU时间上具有一定的优越性.     

求解log—最优组合投资问题的一个自适应算法及其理论分析

本文给出了一个求解log－最优组合投资问题的自适应算法,它是一个变型的随机逼近方法。该问题是一个约束优化问题,因此,采用基于约束流形的梯度上升方向替代常规梯度上升方向,在一些合理的假设下证明了算法的收敛性并进行了渐近稳定性分析。最后,本文将该算法应用于上海证券交易所提供的实际数据的log-最优组合投资问题求解,获得了理想的数值模拟结果。     

一个线性约束优化问题的广义梯度投影法

本文对线性约束优化问题提出了一个新的广义梯度投影法，该算法采用了非精确线性搜索，并在每次迭代运算中结合了广义投影矩阵和变尺度方法的思想确定其搜索方向．在通常的假设条件下，证明了该算法的整体收敛性和超线性收敛速度．     

非光滑凸优化问题的一个非精确梯度镜面下降算法

本文提出一个求解非光滑凸优化问题非精确梯度镜面下降算法.该算法是Allen-Zhu2016年提出求解光滑凸优化问题梯度镜面下降算法的推广,而且该算法允许目标函数中光滑部分梯度计算和非光滑部分邻近算子计算都存在误差,并且在适当条件下分析了该算法函数值序列的O(1/(k²))收敛速度,这里k表示迭代数.最后关于Lasso问题和Logistic问题的数值结果表明该算法是有效的.     

求解低秩矩阵恢复问题的非单调交替梯度方向法

低秩矩阵恢复问题作为一类在图像处理和信号数据分析等领域中都十分重要的问题已被广泛研究.本文在交替方向算法的框架下,应用非单调技术,提出一种求解低秩矩阵恢复问题的新算法.该算法在每一步迭代过程中,首先利用一步带有变步长梯度算法同时更新低秩部分的两块变量,然后采用非单调技术更新稀疏部分的变量.在一定的假设条件下,本文证明了...     

一类分层非结构化P2P系统的随机优化

对于一类利用中心式构架和分布式构架各自优点的分层非结构化P2P系统,通过定义一种Markov切换空间模型来描述其动态组划分切换行为,从而导出相应的系统性能优化问题.同时还在Markov决策过程理论的基础上给出了关于性能指标的参数化梯度优化算法,并通过相应的实例仿真验证了算法的有效性.     

随机Bregman ADMM及其在训练具有离散结构的支持向量机中的应用

针对具有多块可分结构的非凸优化问题提出了一类新的随机Bregman交替方向乘子法,在周期更新规则下, 证明了该算法的渐进收敛性; 在随机更新的规则下, 几乎确定的渐进收敛性得以证明。数值实验结果表明, 该算法可有效训练具有离散结构的支持向量机。     

三项记忆梯度法及其投影算法的收敛性分析

对于无约束优化问题,提出了一类新的三项记忆梯度算法．这类算法是在参数满足某些假设的条件下,确定它的取值范围,从而保证三项记忆梯度方向是使目标函数充分下降的方向．在非单调步长搜索下讨论了算法的全局收敛性．为了得到具有更好收敛性质的算法,结合Solodov and Svaiter(2000)中的部分技巧,提出了一种新的记忆梯度投影算法,并证明了该算法在函数伪凸的情况下具有整体收敛性．     

Minimization Algorithms Based on Supervisor and Searcher Cooperation

In the present work, we explore a general framework for the design of new minimization algorithms with desirable characteristics, namely, supervisor-searcher cooperation. We propose a class of algorithms within this framework and examine a gradient algorithm in the class. Global convergence is established for the deterministic case in the absence of noise and the convergence rate is studied. Both theoretical analysis and numerical tests show that the algorithm is efficient for the deterministic case. Furthermore, the fact that there is no line search procedure incorporated in the algorithm seems to strengthen its robustness so that it tackles effectively test problems with stronger stochastic noises. The numerical results for both deterministic and stochastic test problems illustrate the appealing attributes of the algorithm.     

Online Surrogate Problem Methodology for Stochastic Discrete Resource Allocation Problems

We consider stochastic discrete optimization problems where the decision variables are nonnegative integers. We propose and analyze an online control scheme which transforms the problem into a surrogate continuous optimization problem and proceeds to solve the latter using standard gradient-based approaches, while simultaneously updating both the actual and surrogate system states. It is shown that the solution of the original problem is recovered as an element of the discrete state neighborhood of the optimal surrogate state. For the special case of separable cost functions, we show that this methodology becomes particularly efficient. Finally, convergence of the proposed algorithm is established under standard technical conditions; numerical results are included in the paper to illustrate the fast convergence of this approach.     

A First-Order Stochastic Primal-Dual Algorithm with Correction Step

In this article, we investigate the convergence properties of a stochastic primal-dual splitting algorithm for solving structured monotone inclusions involving the sum of a cocoercive operator and a composite monotone operator. The proposed method is the stochastic extension to monotone inclusions of a proximal method studied in the literature for saddle point problems. It consists in a forward step determined by the stochastic evaluation of the cocoercive operator, a backward step in the dual variables involving the resolvent of the monotone operator, and an additional forward step using the stochastic evaluation of the cocoercive operator introduced in the first step. We prove weak almost sure convergence of the iterates by showing that the primal-dual sequence generated by the method is stochastic quasi-Fejér-monotone with respect to the set of zeros of the considered primal and dual inclusions. Additional results on ergodic convergence in expectation are considered for the special case of saddle point models.     

A stochastic inertial forward–backward splitting algorithm for multivariate monotone inclusions

We propose an inertial forward–backward splitting algorithm to compute a zero of a sum of two monotone operators allowing for stochastic errors in the computation of the operators. More precisely, we establish almost sure convergence in real Hilbert spaces of the sequence of iterates to an optimal solution. Then, based on this analysis, we introduce two new classes of stochastic inertial primal–dual splitting methods for solving structured systems of composite monotone inclusions and prove their convergence. Our results extend to the stochastic and inertial setting various types of structured monotone inclusion problems and corresponding algorithmic solutions. Application to minimization problems is discussed.     

Liquidity Costs: A New Numerical Methodology and an Empirical Study

We consider rate swaps which pay a fixed rate against a floating rate in the presence of bid-ask spread costs. Even for simple models of bid-ask spread costs, there is no explicit strategy optimizing an expected function of the hedging error. We here propose an efficient algorithm based on the stochastic gradient method to compute an approximate optimal strategy without solving a stochastic control problem. We validate our algorithm by numerical experiments. We also develop several variants of the algorithm and discuss their performances in terms of the numerical parameters and the liquidity cost.     

An Extended Ant Colony Algorithm and Its Convergence Analysis

In this work, we propose a stochastic algorithm for solving combinatorial optimization problems. The procedure is formulated within the Ant Colony Optimization (ACO) framework, and extends the so-called Graph-based Ant System with time-dependent evaporation factor, (GBAS/tdev) studied in Gutjahr (2002). In particular, we consider an ACO search procedure which also takes into account the objective function value. We provide a rigorous theoretical study on the convergence of the proposed algorithm. Further, for a toy example, we compare by simulation the rate of convergence of the proposed algorithm with those from the Random Search (RS) and from the corresponding procedure in Gutjahr (2002).AMS 2000 Subject Classification: 9OC15, 9OC27     

Efficient Management of Multiple Sets to Extract Complex Structures from Mathematical Programs

Most of the applied models written with an algebraic modeling language involve simultaneously several dimensions such as materials, location, time or uncertainty. The information about dimensions available in the algebraic formulation is usually sufficient to retrieve different block structures from mathematical programs. These structured problems can then be solved by adequate solution techniques. To illustrate this idea we focus on stochastic programming problems with recourse. Taking into account both time and uncertainty dimensions of these problems, we are able to retrieve different customized structures in their constraint matrices. We applied the Structure Exploiting Tool to retrieve the structure from models built with the GAMS modeling language. The underlying mathematical programs are solved with the decomposition algorithm that applies interior point methods. The optimization algorithm is run in a sequential and in a parallel computing environment.     

Uniform exponential convergence of sample average random functions under general sampling with applications in stochastic programming

Sample average approximation (SAA) is one of the most popular methods for solving stochastic optimization and equilibrium problems. Research on SAA has been mostly focused on the case when sampling is independent and identically distributed (iid) with exceptions (Dai et al. (2000) [9], Homem-de-Mello (2008) [16]). In this paper we study SAA with general sampling (including iid sampling and non-iid sampling) for solving nonsmooth stochastic optimization problems, stochastic Nash equilibrium problems and stochastic generalized equations. To this end, we first derive the uniform exponential convergence of the sample average of a class of lower semicontinuous random functions and then apply it to a nonsmooth stochastic minimization problem. Exponential convergence of estimators of both optimal solutions and M-stationary points (characterized by Mordukhovich limiting subgradients (Mordukhovich (2006) [23], Rockafellar and Wets (1998) [32])) are established under mild conditions. We also use the unform convergence result to establish the exponential rate of convergence of statistical estimators of a stochastic Nash equilibrium problem and estimators of the solutions to a stochastic generalized equation problem.     

On the use of the energy norm in trust-region and adaptive cubic regularization subproblems

We propose a multi-time scale quasi-Newton based smoothed functional (QN-SF) algorithm for stochastic optimization both with and without inequality constraints. The algorithm combines the smoothed functional (SF) scheme for estimating the gradient with the quasi-Newton method to solve the optimization problem. Newton algorithms typically update the Hessian at each instant and subsequently (a) project them to the space of positive definite and symmetric matrices, and (b) invert the projected Hessian. The latter operation is computationally expensive. In order to save computational effort, we propose in this paper a quasi-Newton SF (QN-SF) algorithm based on the Broyden-Fletcher-Goldfarb-Shanno (BFGS) update rule. In Bhatnagar (ACM TModel Comput S. 18(1): 27–62, 2007), a Jacobi variant of Newton SF (JN-SF) was proposed and implemented to save computational effort. We compare our QN-SF algorithm with gradient SF (G-SF) and JN-SF algorithms on two different problems – first on a simple stochastic function minimization problem and the other on a problem of optimal routing in a queueing network. We observe from the experiments that the QN-SF algorithm performs significantly better than both G-SF and JN-SF algorithms on both the problem settings. Next we extend the QN-SF algorithm to the case of constrained optimization. In this case too, the QN-SF algorithm performs much better than the JN-SF algorithm. Finally we present the proof of convergence for the QN-SF algorithm in both unconstrained and constrained settings.