All issues

The article considers minimization of the expectation of convex function. Problems of this type often arise in machine learning and a variety of other applications. In practice, stochastic gradient descent (SGD) and similar procedures are usually used to solve such problems. We propose to use the ellipsoid method with mini-batching, which converges linearly and can be more efficient than SGD for a class of problems. This is verified by our experiments, which are publicly available. The algorithm does not require neither smoothness nor strong convexity of the objective to achieve linear convergence. Thus, its complexity does not depend on the conditional number of the problem. We prove that the method arrives at an approximate solution with given probability when using mini-batches of size proportional to the desired accuracy to the power −2. This enables efficient parallel execution of the algorithm, whereas possibilities for batch parallelization of SGD are rather limited. Despite fast convergence, ellipsoid method can result in a greater total number of calls to oracle than SGD, which works decently with small batches. Complexity is quadratic in dimension of the problem, hence the method is suitable for relatively small dimensionalities.

Finding the global minimum of a nonconvex function is one of the key and most difficult problems of the modern optimization. In this paper we consider special classes of nonconvex problems which have a clear and distinct global minimum.

In the first part of the paper we consider two classes of «good» nonconvex functions, which can be bounded below and above by a parabolic function. This class of problems has not been widely studied in the literature, although it is rather interesting from an applied point of view. Moreover, for such problems first-order and higher-order methods may be completely ineffective in finding a global minimum. This is due to the fact that the function may oscillate heavily or may be very noisy. Therefore, our new methods use only zero-order information and are based on grid search. The size and fineness of this grid, and hence the guarantee of convergence speed and oracle complexity, depend on the «goodness» of the problem. In particular, we show that if the function is bounded by fairly close parabolic functions, then the complexity is independent of the dimension of the problem. We show that our new methods converge with a linear convergence rate $\log(1/\varepsilon)$ to a global minimum on the cube.

In the second part of the paper, we consider the nonconvex optimization problem from a different angle. We assume that the target minimizing function is the sum of the convex quadratic problem and a nonconvex «noise» function proportional to the distance to the global solution. Considering functions with such noise assumptions for zero-order methods is new in the literature. For such a problem, we use the classical gradient-free approach with gradient approximation through finite differences. We show how the convergence analysis for our problems can be reduced to the standard analysis for convex optimization problems. In particular, we achieve a linear convergence rate for such problems as well.

Experimental results confirm the efficiency and practical applicability of all the obtained methods.

In this paper, we introduce a new version of Gauss–Newton method for solving a system of nonlinear equations based on ideas of the residual upper bound for a system of nonlinear equations and a quadratic regularization term. The introduced Gauss–Newton method in practice virtually forms the whole parameterized family of the methods solving systems of nonlinear equations and regression problems. The developed family of Gauss–Newton methods completely consists of iterative methods with generalization for cases of non-euclidean normed spaces, including special forms of Levenberg–Marquardt algorithms. The developed methods use the local model based on a parameterized proximal mapping allowing us to use an inexact oracle of «black–box» form with restrictions for the computational precision and computational complexity. We perform an efficiency analysis including global and local convergence for the developed family of methods with an arbitrary oracle in terms of iteration complexity, precision and complexity of both local model and oracle, problem dimensionality. We present global sublinear convergence rates for methods of the proposed family for solving a system of nonlinear equations, consisting of Lipschitz smooth functions. We prove local superlinear convergence under extra natural non-degeneracy assumptions for system of nonlinear functions. We prove both local and global linear convergence for a system of nonlinear equations under Polyak–Lojasiewicz condition for proposed Gauss– Newton methods. Besides theoretical justifications of methods we also consider practical implementation issues. In particular, for conducted experiments we present effective computational schemes for the exact oracle regarding to the dimensionality of a problem. The proposed family of methods unites several existing and frequent in practice Gauss–Newton method modifications, allowing us to construct a flexible and convenient method implementable using standard convex optimization and computational linear algebra techniques.

The paper is devoted to convex-concave saddle point problems where the objective is a sum of a large number of functions. Such problems attract considerable attention of the mathematical community due to the variety of applications in machine learning, including adversarial learning, adversarial attacks and robust reinforcement learning, to name a few. The individual functions in the sum usually represent losses related to examples from a data set. Additionally, the formulation admits a possibly nonsmooth composite term. Such terms often reflect regularization in machine learning problems. We assume that the dimension of one of the variable groups is relatively small (about a hundred or less), and the other one is large. This case arises, for example, when one considers the dual formulation for a minimization problem with a moderate number of constraints. The proposed approach is based on using Vaidya’s cutting plane method to minimize with respect to the outer block of variables. This optimization algorithm is especially effective when the dimension of the problem is not very large. An inexact oracle for Vaidya’s method is calculated via an approximate solution of the inner maximization problem, which is solved by the accelerated variance reduced algorithm Katyusha. Thus, we leverage the structure of the problem to achieve fast convergence. Separate complexity bounds for gradients of different components with respect to different variables are obtained in the study. The proposed approach is imposing very mild assumptions about the objective. In particular, neither strong convexity nor smoothness is required with respect to the low-dimensional variable group. The number of steps of the proposed algorithm as well as the arithmetic complexity of each step explicitly depend on the dimensionality of the outer variable, hence the assumption that it is relatively small.

Authors describe a two-stage traffic assignment model. It contains of two blocks. The first block consists of a model for calculating a correspondence (demand) matrix, whereas the second block is a traffic assignment model. The first model calculates a matrix of correspondences using a matrix of transport costs (it characterizes the required volumes of movement from one area to another, it is time in this case). To solve this problem, authors propose to use one of the most popular methods of calculating the correspondence matrix in urban studies — the entropy model. The second model describes exactly how the needs for displacement specified by the correspondence matrix are distributed along the possible paths. Knowing the ways of the flows distribution along the paths, it is possible to calculate the cost matrix. Equilibrium in a two-stage model is a fixed point in the sequence of these two models. In practice the problem of finding a fixed point can be solved by the fixed-point iteration method. Unfortunately, at the moment the issue of convergence and estimations of the convergence rate for this method has not been studied quite thoroughly. In addition, the numerical implementation of the algorithm results in many problems. In particular, if the starting point is incorrect, situations may arise where the algorithm requires extremely large numbers to be computed and exceeds the available memory even on the most modern computers. Therefore the article proposes a method for reducing the problem of finding the equilibrium to the problem of the convex non-smooth optimization. Also a numerical method for solving the obtained optimization problem is proposed. Numerical experiments were carried out for both methods of solving the problem. The authors used data for Vladivostok (for this city information from various sources was processed and collected in a new dataset) and two smaller cities in the USA. It was not possible to achieve convergence by the method of fixed-point iteration, whereas the second model for the same dataset demonstrated convergence rate $k^{-1.67}$.

The work is devoted to the construction of efficient and applicable to real tasks first-order methods of convex optimization, that is, using only values of the target function and its derivatives. Construction uses OGMG, fast gradient method which is optimal by complexity, but requires to know the Lipschitz constant for gradient and the strong convexity constant to determine the number of steps and step length. This requirement makes practical usage very hard. An adaptive on the constant for strong convexity algorithm ACGM is proposed, based on restarts of the OGM-G with update of the strong convexity constant estimate, and an adaptive on the Lipschitz constant for gradient ALGM, in which the use of OGM-G restarts is supplemented by the selection of the Lipschitz constant with verification of the smoothness conditions used in the universal gradient descent method. This eliminates the disadvantages of the original method associated with the need to know these constants, which makes practical usage possible. Optimality of estimates for the complexity of the constructed algorithms is proved. To verify the results obtained, experiments on model functions and real tasks from machine learning are carried out.

First-order optimization methods are workhorses in a wide range of modern applications in economics, physics, biology, machine learning, control, and other fields. Among other first-order methods accelerated and momentum ones obtain special attention because of their practical efficiency. The heavy-ball method (HB) is one of the first momentum methods. The method was proposed in 1964 and the first analysis was conducted for quadratic strongly convex functions. Since then a number of variations of HB have been proposed and analyzed. In particular, HB is known for its simplicity in implementation and its performance on nonconvex problems. However, as other momentum methods, it has nonmonotone behavior, and for optimal parameters, the method suffers from the so-called peak effect. To address this issue, in this paper, we consider an averaged version of the heavy-ball method (AHB). We show that for quadratic problems AHB has a smaller maximal deviation from the solution than HB. Moreover, for general convex and strongly convex functions, we prove non-accelerated rates of global convergence of AHB, its weighted version WAHB, and for AHB with restarts R-AHB. To the best of our knowledge, such guarantees for HB with averaging were not explicitly proven for strongly convex problems in the existing works. Finally, we conduct several numerical experiments on minimizing quadratic and nonquadratic functions to demonstrate the advantages of using averaging for HB. Moreover, we also tested one more modification of AHB called the tail-averaged heavy-ball method (TAHB). In the experiments, we observed that HB with a properly adjusted averaging scheme converges faster than HB without averaging and has smaller oscillations.

The article is devoted to solving ill-posed problems of mathematical physics for elliptic and parabolic equations, such as the Cauchy problem for the Helmholtz equation and the retrospective Cauchy problem for the heat equation with constant coefficients. These problems are reduced to problems of convex optimization in Hilbert space. The gradients of the corresponding functionals are calculated approximately by solving two well-posed problems. A new method is proposed for solving the optimization problems under study, it is component-by-component descent in the basis of eigenfunctions of a self-adjoint operator associated with the problem. If it was possible to calculate the gradient exactly, this method would give an arbitrarily exact solution of the problem, depending on the number of considered elements of the basis. In real cases, the inaccuracy of calculations leads to a violation of monotonicity, which requires the use of restarts and limits the achievable quality. The paper presents the results of experiments confirming the effectiveness of the constructed method. It is determined that the new approach is superior to approaches based on the use of gradient optimization methods: it allows to achieve better quality of solution with significantly less computational resources. It is assumed that the constructed method can be generalized to other problems.

We consider one of the problems of transport modelling — searching the equilibrium distribution of traffic flows in the network. We use the classic Beckman’s model to describe time costs and flow distribution in the network represented by directed graph. Meanwhile agents’ behavior is not completely rational, what is described by the introduction of Markov logit dynamics: any driver selects a route randomly according to the Gibbs’ distribution taking into account current time costs on the edges of the graph. Thus, the problem is reduced to searching of the stationary distribution for this dynamics which is a stochastic Nash – Wardrope equilibrium in the corresponding population congestion game in the transport network. Since the game is potential, this problem is equivalent to the problem of minimization of some functional over flows distribution. The stochasticity is reflected in the appearance of the entropy regularization, in contrast to non-stochastic case. The dual problem is constructed to obtain a solution of the optimization problem. The universal primal-dual gradient method is applied. A major specificity of this method lies in an adaptive adjustment to the local smoothness of the problem, what is most important in case of the complex structure of the objective function and an inability to obtain a prior smoothness bound with acceptable accuracy. Such a situation occurs in the considered problem since the properties of the function strongly depend on the transport graph, on which we do not impose strong restrictions. The article describes the algorithm including the numerical differentiation for calculation of the objective function value and gradient. In addition, the paper represents a theoretical estimate of time complexity of the algorithm and the results of numerical experiments conducted on a small American town.

In this paper, we consider convex stochastic optimization problems arising in machine learning applications (e. g., risk minimization) and mathematical statistics (e. g., maximum likelihood estimation). There are two main approaches to solve such kinds of problems, namely the Stochastic Approximation approach (online approach) and the Sample Average Approximation approach, also known as the Monte Carlo approach, (offline approach). In the offline approach, the problem is replaced by its empirical counterpart (the empirical risk minimization problem). The natural question is how to define the problem sample size, i. e., how many realizations should be sampled so that the quite accurate solution of the empirical problem be the solution of the original problem with the desired precision. This issue is one of the main issues in modern machine learning and optimization. In the last decade, a lot of significant advances were made in these areas to solve convex stochastic optimization problems on the Euclidean balls (or the whole space). In this work, we are based on these advances and study the case of arbitrary balls in the $p$-norms. We also explore the question of how the parameter $p$ affects the estimates of the required number of terms as a function of empirical risk.

In this paper, both convex and saddle point optimization problems are considered. For strongly convex problems, the existing results on the same sample sizes in both approaches (online and offline) were generalized to arbitrary norms. Moreover, it was shown that the strong convexity condition can be weakened: the obtained results are valid for functions satisfying the quadratic growth condition. In the case when this condition is not met, it is proposed to use the regularization of the original problem in an arbitrary norm. In contradistinction to convex problems, saddle point problems are much less studied. For saddle point problems, the sample size was obtained under the condition of $\gamma$-growth of the objective function. When $\gamma = 1$, this condition is the condition of sharp minimum in convex problems. In this article, it was shown that the sample size in the case of a sharp minimum is almost independent of the desired accuracy of the solution of the original problem.