Search

We design and analyze a novel accelerated gradient-based algorithm for a class of bilevel optimization problems. These problems have various applications arising from machine learning and image processing, where optimal solutions of the two levels are interdependent. That is, achieving the optimal solution of an upper-level problem depends on the solution set of a lower-level optimization problem. We significantly improve existing iteration complexity to $\mathcal{O}(\epsilon^{-0.5})$ for both suboptimality and infeasibility error metrics, where $\epsilon>0$ denotes an arbitrary scalar. In addition, contrary to existing methods that require solving the optimization problem sequentially (initially solving an optimization problem to approximate the solution of the lower-level problem followed by a second algorithm), our algorithm concurrently solves the optimization problem. To the best of our knowledge, the proposed algorithm has the fastest known iteration complexity, which matches the optimal complexity for single-level optimization. We conduct numerical experiments on sparse linear regression problems to demonstrate the efficacy of our approach.

We consider a class of hierarchical multi-agent optimization problems over networks where agents seek to compute an approximate solution to a single-stage stochastic mathematical program with equilibrium constraints (MPEC). MPECs subsume several important problem classes including Stackelberg games, bilevel programs, and traffic equilibrium problems, to name a few. Our goal in this work is to provably resolve stochastic MPECs in distributed regimes where the agents only have access to their local objectives and an inexact best-response to the lower-level equilibrium problem. To this end, we devise a new method called randomized smoothed distributed zeroth-order gradient tracking (rs-DZGT). This is a novel gradient tracking scheme where agents employ a zeroth-order implicit scheme to approximate their (unavailable) local gradients. Leveraging the properties of a randomized smoothing technique, we establish the convergence of the method and derive complexity guarantees for computing a stationary point of an optimization problem with a smoothed implicit global objective. We also provide preliminary numerical experiments where we compare the performance of rs-DZGT on networks under different settings with that of its centralized counterpart.

Motivated by the emergence of federated learning (FL), we design and analyze federated methods for addressing: (i) Nondifferentiable nonconvex optimization; (ii) Bilevel optimization; (iii) Minimax problems; and (iv) Two-stage stochastic mathematical programs with equilibrium constraints (2s-SMPEC). Research on these problems has been limited and afflicted by reliance on strong assumptions, including the need for differentiability of the implicit function and the absence of constraints in the lower-level problem, among others. We make the following contributions. In (i), by leveraging convolution-based smoothing and Clarke's subdifferential calculus, we devise a randomized smoothing-enabled zeroth-order FL method and derive communication and iteration complexity guarantees for computing an approximate Clarke stationary point. To contend with (ii) and (iii), we devise a unifying randomized implicit zeroth-order FL framework, equipped with explicit communication and iteration complexities. Importantly, our method utilizes delays during local steps to skip calls to the inexact lower-level FL oracle. This results in significant reduction in communication overhead. In (iv), we devise an inexact implicit variant of the method in (i). Remarkably, this method achieves a total communication complexity matching that of single-level nonsmooth nonconvex optimization in FL. We empirically validate the theoretical findings on instances of federated nonsmooth and hierarchical problems., Comment: A preliminary version of this article has been accepted at The 37th Annual Conference on Neural Information Processing Systems (NeurIPS 2023)

We consider a class of hierarchical variational inequality (VI) problems that subsumes VI-constrained optimization and several other problem classes including the optimal solution selection problem and the optimal Nash equilibrium (NE) seeking problem. Our main contributions are threefold. (i) We consider bilevel VIs with monotone and Lipschitz continuous mappings and devise a single-timescale iteratively regularized extragradient method, named IR-EG$_{{\texttt{m,m}}}$. We improve the existing iteration complexity results for addressing both bilevel VI and VI-constrained convex optimization problems. (ii) Under the strong monotonicity of the outer level mapping, we develop a method named IR-EG$_{{\texttt{s,m}}}$ and derive faster guarantees than those in (i). We also study the iteration complexity of this method under a constant regularization parameter. These results appear to be new for both bilevel VIs and VI-constrained optimization. (iii) To our knowledge, complexity guarantees for computing the optimal NE in nonconvex settings do not exist. Motivated by this lacuna, we consider VI-constrained nonconvex optimization problems and devise an inexactly-projected gradient method, named IPR-EG, where the projection onto the unknown set of equilibria is performed using IR-EG$_{{\texttt{s,m}}}$ with a prescribed termination criterion and an adaptive regularization parameter. We obtain new complexity guarantees in terms of a residual map and an infeasibility metric for computing a stationary point. We validate the theoretical findings using preliminary numerical experiments for computing the best and the worst Nash equilibria.

In this paper, we consider stochastic monotone Nash games where each player's strategy set is characterized by possibly a large number of explicit convex constraint inequalities. Notably, the functional constraints of each player may depend on the strategies of other players, allowing for capturing a subclass of generalized Nash equilibrium problems (GNEP). While there is limited work that provide guarantees for this class of stochastic GNEPs, even when the functional constraints of the players are independent of each other, the majority of the existing methods rely on employing projected stochastic approximation (SA) methods. However, the projected SA methods perform poorly when the constraint set is afflicted by the presence of a large number of possibly nonlinear functional inequalities. Motivated by the absence of performance guarantees for computing the Nash equilibrium in constrained stochastic monotone Nash games, we develop a single timescale randomized Lagrangian multiplier stochastic approximation method where in the primal space, we employ an SA scheme, and in the dual space, we employ a randomized block-coordinate scheme where only a randomly selected Lagrangian multiplier is updated. We show that our method achieves a convergence rate of $\mathcal{O}\left(\frac{\log(k)}{\sqrt{k}}\right)$ for suitably defined suboptimality and infeasibility metrics in a mean sense., Comment: The result of this paper has been presented at International Conference on Continuous Optimization (ICCOPT) 2022 and East Coast Optimization Meeting (ECOM) 2023

The goal in this paper is to approximate the Price of Stability (PoS) in stochastic Nash games using stochastic approximation (SA) schemes. PoS is amongst the most popular metrics in game theory and provides an avenue for estimating the efficiency of Nash games. In particular, knowing the value of PoS can help with designing efficient networked systems, including transportation networks and power market mechanisms. Motivated by the lack of efficient methods for computing the PoS, first we consider stochastic optimization problems with a nonsmooth and merely convex objective function and a merely monotone stochastic variational inequality (SVI) constraint. This problem appears in the numerator of the PoS ratio. We develop a randomized block-coordinate stochastic extra-(sub)gradient method where we employ a novel iterative penalization scheme to account for the mapping of the SVI in each of the two gradient updates of the algorithm. We obtain an iteration complexity of the order $\epsilon^{-4}$ that appears to be best known result for this class of constrained stochastic optimization problems, where $\epsilon$ denotes an arbitrary bound on suitably defined infeasibility and suboptimality metrics. Second, we develop an SA-based scheme for approximating the PoS and derive lower and upper bounds on the approximation error. To validate the theoretical findings, we provide preliminary simulation results on a networked stochastic Nash Cournot competition.

We consider distributed optimization over networks where each agent is associated with a smooth and strongly convex local objective function. We assume that the agents only have access to unbiased estimators of the gradient of their objective functions. Motivated by big data applications, our goal lies in addressing this problem when the dimensionality of the solution space is possibly large and consequently, the computation of the local gradient mappings may become expensive. We develop a randomized block-coordinate variant of the recently developed distributed stochastic gradient tracking (DSGT) method. We derive non-asymptotic convergence rates of the order $1/k$ and $1/k^2$ in terms of an optimality metric and a consensus violation metric, respectively. Importantly, while block-coordinate schemes have been studied for distributed optimization problems before, the proposed algorithm appears to be the first randomized block-coordinate gradient tracking method that is equipped with the aforementioned convergence rate statements. We validate the performance of the proposed method on the MNIST and a synthetic data set under different network settings.

We consider the minimization of an $L_0$-Lipschitz continuous and expectation-valued function, denoted by $f$ and defined as $f(x)\triangleq \mathbb{E}[\tilde{f}(x,\omega)]$, over a Cartesian product of closed and convex sets with a view towards obtaining both asymptotics as well as rate and complexity guarantees for computing an approximate stationary point (in a Clarke sense). We adopt a smoothing-based approach reliant on minimizing $f_{\eta}$ where $f_{\eta}(x) \triangleq \mathbb{E}_{u}[f(x+\eta u)]$, $u$ is a random variable defined on a unit sphere, and $\eta > 0$. In fact, it is observed that a stationary point of the $\eta$-smoothed problem is a $2\eta$-stationary point for the original problem in the Clarke sense. In such a setting, we derive a suitable residual function that provides a metric for stationarity for the smoothed problem. By leveraging a zeroth-order framework reliant on utilizing sampled function evaluations implemented in a block-structured regime, we make two sets of contributions for the sequence generated by the proposed scheme. (i) The residual function of the smoothed problem tends to zero almost surely along the generated sequence; (ii) To compute an $x$ that ensures that the expected norm of the residual of the $\eta$-smoothed problem is within $\epsilon$ requires no greater than $\mathcal{O}(\tfrac{1}{\eta \epsilon^2})$ projection steps and $\mathcal{O}\left(\tfrac{1}{\eta^2 \epsilon^4}\right)$ function evaluations. These statements appear to be novel and there appear to be few results to contend with general nonsmooth, nonconvex, and stochastic regimes via zeroth-order approaches.

We consider minimizing a sum of agent-specific nondifferentiable merely convex functions over the solution set of a variational inequality (VI) problem in that each agent is associated with a local monotone mapping. This problem finds an application in computation of the best equilibrium in nonlinear complementarity problems arising in transportation networks. We develop an iteratively regularized incremental gradient method where at each iteration, agents communicate over a cycle graph to update their solution iterates using their local information about the objective and the mapping. The proposed method is single-timescale in the sense that it does not involve any excessive hard-to-project computation per iteration. We derive non-asymptotic agent-wise convergence rates for the suboptimality of the global objective function and infeasibility of the VI constraints measured by a suitably defined dual gap function. The proposed method appears to be the first fully iterative scheme equipped with iteration complexity that can address distributed optimization problems with VI constraints over cycle graphs. Preliminary numerical experiments for a transportation network problem and a support vector machine model are presented.

Stochastic MPECs have found increasing relevance for modeling a broad range of settings in engineering and statistics. Yet, there seem to be no efficient first/zeroth-order schemes equipped with non-asymptotic rate guarantees for resolving even deterministic variants of such problems. We consider SMPECs where the parametrized lower-level equilibrium problem is given by a deterministic/stochastic VI problem whose mapping is strongly monotone. We develop a zeroth-order implicit algorithmic framework by leveraging a locally randomized spherical smoothing scheme. We present schemes for single-stage and two-stage stochastic MPECs when the upper-level problem is either convex or nonconvex. (I). Single-stage SMPECs: In convex regimes, our proposed inexact schemes are characterized by a complexity in upper-level projections, upper-level samples, and lower-level projections of $\mathcal{O}(\tfrac{1}{\epsilon^2})$, $\mathcal{O}(\tfrac{1}{\epsilon^2})$, and $\mathcal{O}(\tfrac{1}{\epsilon^2}\ln(\tfrac{1}{\epsilon}))$ , respectively. Analogous bounds for the nonconvex regime are $\mathcal{O}(\tfrac{1}{\epsilon})$, $\mathcal{O}(\tfrac{1}{\epsilon^2})$, and $\mathcal{O}(\tfrac{1}{\epsilon^3})$, respectively . (II). Two-stage SMPECs: In convex regimes, our proposed inexact schemes have a complexity in upper-level projections, upper-level samples, and lower-level projections of $\mathcal{O}(\tfrac{1}{\epsilon^2}),\mathcal{O}(\tfrac{1}{\epsilon^2})$, and $\mathcal{O}(\tfrac{1}{\epsilon^2}\ln(\tfrac{1}{\epsilon}))$ while the corresponding bounds in the nonconvex regime are $\mathcal{O}(\tfrac{1}{\epsilon})$, $\mathcal{O}(\tfrac{1}{\epsilon^2})$, and $\mathcal{O}(\tfrac{1}{\epsilon^2}\ln(\tfrac{1}{\epsilon}))$ , respectively . In addition, we derive statements for exact as well as accelerated counterparts. We also provide a comprehensive set of numerical results for validating the theoretical findings.

We consider a class of optimization problems with Cartesian variational inequality (CVI) constraints, where the objective function is convex and the CVI is associated with a monotone mapping and a convex Cartesian product set. This mathematical formulation captures a wide range of optimization problems including those complicated by the presence of equilibrium constraints, complementarity constraints, or an inner-level large scale optimization problem. In particular, an important motivating application arises from the notion of efficiency estimation of equilibria in multi-agent networks, e.g., communication networks and power systems. In the literature, the iteration complexity of the existing solution methods for optimization problems with CVI constraints appears to be unknown. To address this shortcoming, we develop a first-order method called averaging randomized block iteratively regularized gradient (aRB-IRG). The main contributions include: (i) In the case where the associated set of the CVI is bounded and the objective function is nondifferentiable and convex, we derive new non-asymptotic suboptimality and infeasibility convergence rate statements in a mean sense. We also obtain deterministic variants of the convergence rates when we suppress the randomized block-coordinate scheme. Importantly, this paper appears to be the first work to provide these rate guarantees for this class of problems. (ii) In the case where the CVI set is unbounded and the objective function is smooth and strongly convex, utilizing the properties of the Tikhonov trajectory, we establish the global convergence of aRB-IRG in an almost sure and a mean sense. We provide the numerical experiments for computing the best Nash equilibrium in a networked Cournot competition model.

Motivated by applications arising from sensor networks and machine learning, we consider the problem of minimizing a finite sum of nondifferentiable convex functions where each component function is associated with an agent and a hard-to-project constraint set. Among well-known avenues to address finite sum problems is the class of incremental gradient (IG) methods where a single component function is selected at each iteration in a cyclic or randomized manner. When the problem is constrained, the existing IG schemes (including projected IG, proximal IAG, and SAGA) require a projection step onto the feasible set at each iteration. Consequently, the performance of these schemes is afflicted with costly projections when the problem includes: (1) nonlinear constraints, or (2) a large number of linear constraints. Our focus in this paper lies in addressing both of these challenges. We develop an algorithm called averaged iteratively regularized incremental gradient (aIR-IG) that does not involve any hard-to-project computation. Under mild assumptions, we derive non-asymptotic rates of convergence for both suboptimality and infeasibility metrics. Numerically, we show that the proposed scheme outperforms the standard projected IG methods on distributed soft-margin support vector machine problems.

Motivated by emerging applications in wireless sensor networks and large-scale data processing, we consider distributed optimization over directed networks where the agents communicate their information locally to their neighbors to cooperatively minimize a global cost function. We introduce a new unifying distributed constrained optimization model that is characterized as a bilevel optimization problem. This model captures a wide range of existing problems over directed networks including: (i) Distributed optimization with linear constraints; (ii) Distributed unconstrained nonstrongly convex optimization over directed networks. Employing a novel regularization-based relaxation approach and gradient-tracking schemes, we develop an iteratively regularized push-pull gradient algorithm. We establish the consensus and derive new convergence rate statements for suboptimality and infeasibility of the generated iterates for solving the bilevel model. The proposed algorithm and the complexity analysis obtained in this work appear to be new for addressing the bilevel model and also for the two sub-classes of problems. The numerical performance of the proposed algorithm is presented.

The goal in this paper is to develop first-order methods equipped with convergence rates for multi-agent optimization problems on semidefinite matrix spaces. These problems include cooperative optimization problems and non-cooperative Nash games. Accordingly, first we consider a multi-agent system where the agents cooperatively minimize the summation of their local convex objectives, and second, we consider Cartesian stochastic variational inequality (CSVI) problems with monotone mappings for addressing stochastic Nash games on semidefinite matrix spaces. Despite the recent advancements in first-order methods addressing problems over vector spaces, there seems to be a major gap in the theory of the first-order methods for optimization problems and equilibriums on semidefinite matrix spaces. In particular, to the best of our knowledge, there exists no method with provable convergence rate for solving the two classes of problems under mild assumptions. Most existing methods either rely on strong assumptions, or require a two-loop framework where at each iteration a semidefinite optimization problem needs to be solved. Motivated by this gap, in the first part of the paper, we develop a mirror descent incremental subgradient method for minimizing a finite-sum function. We show that the iterates generated by the algorithm converge asymptotically to an optimal solution and derive a non-asymptotic convergence rate. In the second part, we consider semidefinite CSVI problems. We develop a stochastic mirror descent method that only requires monotonicity of the mapping. We show that the iterates generated by the algorithm converge to a solution of the CSVI almost surely. Using a suitably defined gap function, we derive a convergence rate statement. This work appears to be the first that provides a convergence speed guarantee for monotone CSVIs on semidefinite matrix spaces., Comment: arXiv admin note: text overlap with arXiv:1809.09155

A wide range of applications arising in machine learning and signal processing can be cast as convex optimization problems. These problems are often ill-posed, i.e., the optimal solution lacks a desired property such as uniqueness or sparsity. In the literature, to address ill-posedness, a bilevel optimization problem is considered where the goal is to find among optimal solutions of the inner level optimization problem, a solution that minimizes a secondary metric, i.e., the outer level objective function. In addressing the resulting bilevel model, the convergence analysis of most existing methods is limited to the case where both inner and outer level objectives are differentiable deterministic functions. While these assumptions may not hold in big data applications, to the best of our knowledge, no solution method equipped with complexity analysis exists to address presence of uncertainty and nondifferentiability in both levels in this class of problems. Motivated by this gap, we develop a first-order method called Iterative Regularized Stochastic Mirror Descent (IR-SMD). We establish the global convergence of the iterate generated by the algorithm to the optimal solution of the bilevel problem in an almost sure and a mean sense. We derive a convergence rate of ${\cal O}\left(1/N^{0.5-\delta}\right)$ for the inner level problem, where $\delta>0$ is an arbitrary small scalar. Numerical experiments for solving two classes of bilevel problems, including a large scale binary text classification application, are presented.

We study a class of bilevel convex optimization problems where the goal is to find the minimizer of an objective function in the upper level, among the set of all optimal solutions of an optimization problem in the lower level. A wide range of problems in convex optimization can be formulated using this class. An important example is the case where an optimization problem is ill-posed. In this paper, our interest lies in addressing the bilevel problems, where the lower level objective is given as a finite sum of separate nondifferentiable convex component functions. This is the case in a variety of applications in distributed optimization, such as large-scale data processing in machine learning and neural networks. To the best of our knowledge, this class of bilevel problems, with a finite sum in the lower level, has not been addressed before. Motivated by this gap, we develop an iterative regularized incremental subgradient method, where the agents update their iterates in a cyclic manner using a regularized subgradient. Under a suitable choice of the regularization parameter sequence, we establish the convergence of the proposed algorithm and derive a rate of $\mathcal{O} \left({1}/k^{0.5-\epsilon}\right)$ in terms of the lower level objective function for an arbitrary small $\epsilon>0$. We present the performance of the algorithm on a binary text classification problem., Comment: 8 pages, 1 figure

Motivated by high-dimensional nonlinear optimization problems as well as ill-posed optimization problems arising in image processing, we consider a bilevel optimization model where we seek among the optimal solutions of the inner level problem, a solution that minimizes a secondary metric. Our goal is to address the high-dimensionality of the bilevel problem, and the nondifferentiability of the objective function. Minimal norm gradient, sequential averaging, and iterative regularization are some of the recent schemes developed for addressing the bilevel problem. But none of them address the high-dimensional structure and nondifferentiability. With this gap in the literature, we develop a randomized block coordinate iterative regularized gradient descent scheme (RB-IRG). We establish the convergence of the sequence generated by RB-IRG to the unique solution of the bilevel problem of interest. Furthermore, we derive a rate of convergence $\mathcal{O} \left(\frac{1}{{k}^{0.5-\delta}}\right)$, with respect to the inner level objective function. We demonstrate the performance of RB-IRG in solving the ill-posed problems arising in image processing., Comment: 8 pages, 2 figures, American Control Conference

Motivated by multi-user optimization problems and non-cooperative Nash games in stochastic regimes, we consider stochastic variational inequality (SVI) problems on matrix spaces where the variables are positive semidefinite matrices and the mapping is merely monotone. Much of the interest in the theory of variational inequality (VI) has focused on addressing VIs on vector spaces.Yet, most existing methods either rely on strong assumptions, or require a two-loop framework where at each iteration, a projection problem, i.e., a semidefinite optimization problem needs to be solved. Motivated by this gap, we develop a stochastic mirror descent method where we choose the distance generating function to be defined as the quantum entropy. This method is a single-loop first-order method in the sense that it only requires a gradient-type of update at each iteration. The novelty of this work lies in the convergence analysis that is carried out through employing an auxiliary sequence of stochastic matrices. Our contribution is three-fold: (i) under this setting and employing averaging techniques, we show that the iterate generated by the algorithm converges to a weak solution of the SVI; (ii) moreover, we derive a convergence rate in terms of the expected value of a suitably defined gap function; (iii) we implement the developed method for solving a multiple-input multiple-output multi-cell cellular wireless network composed of seven hexagonal cells and present the numerical experiments supporting the convergence of the proposed method.

Classical theory for quasi-Newton schemes has focused on smooth deterministic unconstrained optimization while recent forays into stochastic convex optimization have largely resided in smooth, unconstrained, and strongly convex regimes. Naturally, there is a compelling need to address nonsmoothness, the lack of strong convexity, and the presence of constraints. Accordingly, this paper presents a quasi-Newton framework that can process merely convex and possibly nonsmooth (but smoothable) stochastic convex problems. We propose a framework that combines iterative smoothing and regularization with a variance-reduced scheme reliant on using increasing sample-sizes of gradients. We make the following contributions. (i) We develop a regularized and smoothed variable sample-size BFGS update (rsL-BFGS) that generates a sequence of Hessian approximations and can accommodate nonsmooth convex objectives by utilizing iterative regularization and smoothing. (ii) In strongly convex regimes with state-dependent noise, the proposed variable sample-size stochastic quasi-Newton scheme admits a non-asymptotic linear rate of convergence while the oracle complexity of computing an $\epsilon$-solution is $\mathcal{O}(\kappa^{m+1}/\epsilon)$ where $\kappa$ is the condition number and $m\geq 1$. In nonsmooth (but smoothable) regimes, using Moreau smoothing retains the linear convergence rate. To contend with the possible unavailability of Lipschitzian and strong convexity parameters, we also provide sublinear rates; (iii) In merely convex but smooth settings, the regularized VS-SQN scheme rVS-SQN displays a rate of $\mathcal{O}(1/k^{(1-\varepsilon)})$. When the smoothness requirements are weakened, the rate for the regularized and smoothed VS-SQN scheme worsens to $\mathcal{O}(k^{-1/3})$. Such statements allow for a state-dependent noise assumption under a quadratic growth property.

Motivated by applications arising from large scale optimization and machine learning, we consider stochastic quasi-Newton (SQN) methods for solving unconstrained convex optimization problems. The convergence analysis of the SQN methods, both full and limited-memory variants, require the objective function to be strongly convex. However, this assumption is fairly restrictive and does not hold for applications such as minimizing the logistic regression loss function. To the best of our knowledge, no rate statements currently exist for SQN methods in the absence of such an assumption. Also, among the existing first-order methods for addressing stochastic optimization problems with merely convex objectives, those equipped with provable convergence rates employ averaging. However, this averaging technique has a detrimental impact on inducing sparsity. Motivated by these gaps, the main contributions of the paper are as follows: (i) Addressing large scale stochastic optimization problems, we develop an iteratively regularized stochastic limited-memory BFGS (IRS-LBFGS) algorithm, where the stepsize, regularization parameter, and the Hessian inverse approximation matrix are updated iteratively. We establish the convergence to an optimal solution of the original problem both in an almost-sure and mean senses. We derive the convergence rate in terms of the objective function's values and show that it is of the order $\mathcal{O}\left(k^{-\left(\frac{1}{3}-\epsilon\right)}\right)$, where $\epsilon$ is an arbitrary small positive scalar; (ii) In deterministic regime, we show that the regularized limited-memory BFGS algorithm displays a rate of the order $\mathcal{O}\left(\frac{1}{k^{1 -\epsilon'}}\right)$, where $\epsilon'$ is an arbitrary small positive scalar. We present our numerical experiments performed on a large scale text classification problem.

We consider randomized block coordinate stochastic mirror descent (RBSMD) methods for solving high-dimensional stochastic optimization problems with strongly convex objective functions. Our goal is to develop RBSMD schemes that achieve a rate of convergence with a minimum constant factor with respect to the choice of the stepsize sequence. To this end, we consider both subgradient and gradient RBSMD methods addressing nonsmooth and smooth problems, respectively. For each scheme, (i) we develop self-tuned stepsize rules characterized in terms of problem parameters and algorithm settings; (ii) we show that the non-averaging iterate generated by the underlying RBSMD method converges to the optimal solution both in an almost sure and a mean sense; (iii) we show that the mean squared error is minimized. When problem parameters are unknown, we develop a unifying self-tuned update rule that can be applied in both subgradient and gradient SMD methods, and show that for any arbitrary and small enough initial stepsize, a suitably defined error bound is minimized. We provide constant factor comparisons with standard SMD and RBSMD methods. Our numerical experiments performed on an SVM model display that the self-tuned schemes are significantly robust with respect to the choice of problem parameters, and the initial stepsize.

Motivated by multi-user optimization problems and non-cooperative Nash games in uncertain regimes, we consider stochastic Cartesian variational inequalities (SCVI) where the set is given as the Cartesian product of a collection of component sets. First, we consider the case where the number of the component sets is large. For solving this type of problems, the classical stochastic approximation methods and their prox generalizations are computationally inefficient as each iteration becomes very costly. To address this challenge, we develop a randomized block stochastic mirror-prox (B-SMP) algorithm, where at each iteration only a randomly selected block coordinate of the solution is updated through implementing two consecutive projection steps. Under standard assumptions on the problem and settings of the algorithm, we show that when the mapping is strictly pseudo-monotone, the algorithm generates a sequence of iterates that converges to the solution of the problem almost surely. To derive rate statements, we assume that the maps are strongly pseudo-monotone and obtain {a non-asymptotic mean squared error $\mathcal{O}\left(\frac{d}{k}\right)$, where $k$ is the iteration number and $d$ is the number of component sets. Second, we consider large-scale stochastic optimization problems with convex objectives. For this class of problems, we develop a new averaging scheme for the B-SMP algorithm. Unlike the classical averaging stochastic mirror-prox (SMP) method where a decreasing set of weights for the averaging sequence is used, here we consider a different set of weights that are characterized in terms of the stepsizes and a {parameter}. We show that using such weights, the objective values of the averaged sequence converges to the optimal value in the mean sense with the rate $\mathcal{O}\left(\frac{\sqrt{d}}{\sqrt{k}}\right)$.

Motivated by applications in optimization and machine learning, we consider stochastic quasi-Newton (SQN) methods for solving stochastic optimization problems. In the literature, the convergence analysis of these algorithms relies on strong convexity of the objective function. To our knowledge, no theoretical analysis is provided for the rate statements in the absence of this assumption. Motivated by this gap, we allow the objective function to be merely convex and we develop a cyclic regularized SQN method where the gradient mapping and the Hessian approximation matrix are both regularized at each iteration and are updated in a cyclic manner. We show that, under suitable assumptions on the stepsize and regularization parameters, the objective function value converges to the optimal objective function of the original problem in both almost sure and the expected senses. For each case, a class of feasible sequences that guarantees the convergence is provided. Moreover, the rate of convergence in terms of the objective function value is derived. Our empirical analysis on a binary classification problem shows that the proposed scheme performs well compared to both classic regularization SQN schemes and stochastic approximation method., Comment: 8 pages

Traditionally, stochastic approximation schemes for SVIs have relied on strong monotonicity and Lipschitzian properties of the underlying map. In contrast, we consider monotone stochastic variational inequality (SVI) problems where the strong monotonicity and Lipschitzian assumptions on the mappings are weakened. In the first part of the paper, to address such shortcomings, a regularized smoothed SA (RSSA) scheme is developed wherein the stepsize, smoothing, and regularization parameters are diminishing sequences updated after every iteration. Under suitable assumptions on the sequences, we show that the algorithm generates iterates that converge to a solution in an almost sure sense, extending the results in [16] to the non-Lipschitzian regime. Motivated by the need to develop non-asymptotic rate statements, in the second part of the paper, we develop a variant of the RSSA scheme, denoted by aRSSA$_r$, in which we employ a weighted iterate-averaging, parametrized by a scalar $r$ where $r = 1$ provides us with the standard averaging scheme. We make several contributions in this context: First, we show that the gap function associated with the sequences by the aRSSA$_r$ scheme tends to zero when the parameter sequences are chosen appropriately. Second, we show that the gap function associated with the averaged sequence diminishes to zero at the optimal rate $\cal{O}(1/\sqrt{K})$ after $K$ steps when smoothing and regularization are suppressed and $r < 1$, thus improving the rate statement for the standard averaging which admits a rate of $\cal{O}(\ln(K)/\sqrt{K})$. Third, we develop a window-based variant of this scheme that also displays the optimal rate for $r < 1$. Notably, we prove the superiority of the scheme with $r < 1$ with its counterpart with $r=1$ in terms of the constant factor of the error bound when the size of the averaging window is sufficiently large.

We consider stochastic variational inequality problems where the mapping is monotone over a compact convex set. We present two robust variants of stochastic extragradient algorithms for solving such problems. Of these, the first scheme employs an iterative averaging technique where we consider a generalized choice for the weights in the averaged sequence. Our first contribution is to show that using an appropriate choice for these weights, a suitably defined gap function attains the optimal rate of convergence ${\cal O}\left(\frac{1}{\sqrt{k}}\right)$. In the second part of the paper, under an additional assumption of weak-sharpness, we update the stepsize sequence using a recursive rule that leverages problem parameters. The second contribution lies in showing that employing such a sequence, the extragradient algorithm possesses almost-sure convergence to the solution as well as convergence in a mean-squared sense to the solution of the problem at the rate ${\cal O}\left(\frac{1}{k}\right)$. Motivated by the absence of a Lipschitzian parameter, in both schemes we utilize a locally randomized smoothing scheme. Importantly, by approximating a smooth mapping, this scheme enables us to estimate the Lipschitzian parameter. The smoothing parameter is updated per iteration and we show convergence to the solution of the original problem in both algorithms.

We consider a distributed stochastic approximation (SA) scheme for computing an equilibrium of a stochastic Nash game. Standard SA schemes employ diminishing steplength sequences that are square summable but not summable. Such requirements provide a little or no guidance for how to leverage Lipschitzian and monotonicity properties of the problem and naive choices generally do not preform uniformly well on a breadth of problems. While a centralized adaptive stepsize SA scheme is proposed in [1] for the optimization framework, such a scheme provides no freedom for the agents in choosing their own stepsizes. Thus, a direct application of centralized stepsize schemes is impractical in solving Nash games. Furthermore, extensions to game-theoretic regimes where players may independently choose steplength sequences are limited to recent work by Koshal et al. [2]. Motivated by these shortcomings, we present a distributed algorithm in which each player updates his steplength based on the previous steplength and some problem parameters. The steplength rules are derived from minimizing an upper bound of the errors associated with players' decisions. It is shown that these rules generate sequences that converge almost surely to an equilibrium of the stochastic Nash game. Importantly, variants of this rule are suggested where players independently select steplength sequences while abiding by an overall coordination requirement. Preliminary numerical results are seen to be promising., Comment: 8 pages, Proceedings of the American Control Conference, Washington, 2013

Motivated by problems arising in decentralized control problems and non-cooperative Nash games, we consider a class of strongly monotone Cartesian variational inequality (VI) problems, where the mappings either contain expectations or their evaluations are corrupted by error. Such complications are captured under the umbrella of Cartesian stochastic variational inequality problems and we consider solving such problems via stochastic approximation (SA) schemes. Specifically, we propose a scheme wherein the steplength sequence is derived by a rule that depends on problem parameters such as monotonicity and Lipschitz constants. The proposed scheme is seen to produce sequences that are guaranteed to converge almost surely to the unique solution of the problem. To cope with networked multi-agent generalizations, we provide requirements under which independently chosen steplength rules still possess desirable almost-sure convergence properties. In the second part of this paper, we consider a regime where Lipschitz constants on the map are either unavailable or difficult to derive. Here, we present a local randomization technique that allows for deriving an approximation of the original mapping, which is then shown to be Lipschitz continuous with a prescribed constant. Using this technique, we introduce a locally randomized SA algorithm and provide almost-sure convergence theory for the resulting sequence of iterates to an approximate solution of the original variational inequality problem. Finally, the paper concludes with some preliminary numerical results on a stochastic rate allocation problem and a stochastic Nash-Cournot game., Comment: 36 pages, submitted to Mathematical Programming

The performance of standard stochastic approximation implementations can vary significantly based on the choice of the steplength sequence, and in general, little guidance is provided about good choices. Motivated by this gap, in the first part of the paper, we present two adaptive steplength schemes for strongly convex differentiable stochastic optimization problems, equipped with convergence theory. The first scheme, referred to as a recursive steplength stochastic approximation scheme, optimizes the error bounds to derive a rule that expresses the steplength at a given iteration as a simple function of the steplength at the previous iteration and certain problem parameters. This rule is seen to lead to the optimal steplength sequence over a prescribed set of choices. The second scheme, termed as a cascading steplength stochastic approximation scheme, maintains the steplength sequence as a piecewise-constant decreasing function with the reduction in the steplength occurring when a suitable error threshold is met. In the second part of the paper, we allow for nondifferentiable objective and we propose a local smoothing technique that leads to a differentiable approximation of the function. Assuming a uniform distribution on the local randomness, we establish a Lipschitzian property for the gradient of the approximation and prove that the obtained Lipschitz bound grows at a modest rate with problem size. This facilitates the development of an adaptive steplength stochastic approximation framework, which now requires sampling in the product space of the original measure and the artificially introduced distribution. The resulting adaptive steplength schemes are applied to three stochastic optimization problems. We observe that both schemes perform well in practice and display markedly less reliance on user-defined parameters.