Search

We analyze nonlinearly preconditioned gradient methods for solving smooth minimization problems. We introduce a generalized smoothness property, based on the notion of abstract convexity, that is broader than Lipschitz smoothness and provide sufficient first- and second-order conditions. Notably, our framework encapsulates algorithms associated with the clipping gradient method and brings out novel insights for the class of $(L_0,L_1)$-smooth functions that has received widespread interest recently, thus allowing us to go beyond already established methods. We investigate the convergence of the proposed method in both the convex and nonconvex setting.

This paper presents an efficient Mixed-Integer Nonlinear Programming (MINLP) formulation for systems with discrete control inputs under dwell time constraints. By viewing such systems as a switched system, the problem is decomposed into a Sequence Optimization (SO) and a Switching Time Optimization (STO) -- the former providing the sequence of the switched system, and the latter calculating the optimal switching times. By limiting the feasible set of SO to subsequences of a master sequence, this formulation requires a small number of binary variables, independent of the number of time discretization nodes. This enables the proposed formulation to provide solutions efficiently, even for large numbers of time discretization nodes. To provide even faster solutions, an iterative algorithm is introduced to heuristically solve STO and SO. The proposed approaches are then showcased on four different switched systems and results demonstrate the efficiency of the MINLP formulation and the iterative algorithm.

The scaled relative graph (SRG) is a powerful graphical tool for analyzing the properties of operators, by mapping their graph onto the complex plane. In this work, we study the SRG of two classes of nonmonotone operators, namely the general class of semimonotone operators and a class of angle-bounded operators. In particular, we provide an analytical description of the SRG of these classes and show that membership of an operator to these classes can be verified through geometric containment of its SRG. To illustrate the importance of these results, we provide several examples in the context of electrical circuits. Most notably, we show that the Ebers-Moll transistor belongs to the class of angle-bounded operators and use this result to compute the response of a common-emitter amplifier using Chambolle-Pock, despite the underlying nonsmoothness and multi-valuedness, leveraging recent convergence results for this algorithm in the nonmonotone setting., Comment: 7 pages

This paper presents a novel approach to imitation learning from observations, where an autoregressive mixture of experts model is deployed to fit the underlying policy. The parameters of the model are learned via a two-stage framework. By leveraging the existing dynamics knowledge, the first stage of the framework estimates the control input sequences and hence reduces the problem complexity. At the second stage, the policy is learned by solving a regularized maximum-likelihood estimation problem using the estimated control input sequences. We further extend the learning procedure by incorporating a Lyapunov stability constraint to ensure asymptotic stability of the identified model, for accurate multi-step predictions. The effectiveness of the proposed framework is validated using two autonomous driving datasets collected from human demonstrations, demonstrating its practical applicability in modelling complex nonlinear dynamics.

This work introduces an unconventional inexact augmented Lagrangian method, where the augmenting term is a Euclidean norm raised to a power between one and two. The proposed algorithm is applicable to a broad class of constrained nonconvex minimization problems, that involve nonlinear equality constraints over a convex set under a mild regularity condition. First, we conduct a full complexity analysis of the method, leveraging an accelerated first-order algorithm for solving the H\"older-smooth subproblems. Next, we present an inexact proximal point method to tackle these subproblems, demonstrating that it achieves an improved convergence rate. Notably, this rate reduces to the best-known convergence rate for first-order methods when the augmenting term is a squared Euclidean norm. Our worst-case complexity results further show that using lower powers for the augmenting term leads to faster constraint satisfaction, albeit with a slower decrease in the dual residual. Numerical experiments support our theoretical findings, illustrating that this trade-off between constraint satisfaction and cost minimization is advantageous for certain practical problems.

We examine stability properties of primal-dual gradient flow dynamics for composite convex optimization problems with multiple, possibly nonsmooth, terms in the objective function under the generalized consensus constraint. The proposed dynamics are based on the proximal augmented Lagrangian and they provide a viable alternative to ADMM which faces significant challenges from both analysis and implementation viewpoints in large-scale multi-block scenarios. In contrast to customized algorithms with individualized convergence guarantees, we provide a systematic approach for solving a broad class of challenging composite optimization problems. We leverage various structural properties to establish global (exponential) convergence guarantees for the proposed dynamics. Our assumptions are much weaker than those required to prove (exponential) stability of various primal-dual dynamics as well as (linear) convergence of discrete-time methods, e.g., standard two-block and multi-block ADMM and EXTRA algorithms. Finally, we show necessity of some of our structural assumptions for exponential stability and provide computational experiments to demonstrate the convenience of the proposed dynamics for parallel and distributed computing applications., Comment: 31 pages; 4 figures

Lossy image compression is essential for efficient transmission and storage. Traditional compression methods mainly rely on discrete cosine transform (DCT) or singular value decomposition (SVD), both of which represent image data in continuous domains and therefore necessitate carefully designed quantizers. Notably, SVD-based methods are more sensitive to quantization errors than DCT-based methods like JPEG. To address this issue, we introduce a variant of integer matrix factorization (IMF) to develop a novel quantization-free lossy image compression method. IMF provides a low-rank representation of the image data as a product of two smaller factor matrices with bounded integer elements, thereby eliminating the need for quantization. We propose an efficient, provably convergent iterative algorithm for IMF using a block coordinate descent (BCD) scheme, with subproblems having closed-form solutions. Our experiments on the Kodak and CLIC 2024 datasets demonstrate that our IMF compression method consistently outperforms JPEG at low bit rates below 0.25 bits per pixel (bpp) and remains comparable at higher bit rates. We also assessed our method's capability to preserve visual semantics by evaluating an ImageNet pre-trained classifier on compressed images. Remarkably, our method improved top-1 accuracy by over 5 percentage points compared to JPEG at bit rates under 0.25 bpp. The project is available at https://github.com/pashtari/lrf ., Comment: 19 pages, 6 figures, 1 table, 1 algorithm

This paper proposes a general switching dynamical system model, and a custom majorization-minimization-based algorithm EM++ for identifying its parameters. For certain families of distributions, such as Gaussian distributions, this algorithm reduces to the well-known expectation-maximization method. We prove global convergence of the algorithm under suitable assumptions, thus addressing an important open issue in the switching system identification literature. The effectiveness of both the proposed model and algorithm is validated through extensive numerical experiments.

We consider gradient descent with constant stepsizes and derive exact worst-case convergence rates on the minimum gradient norm of the iterates. Our analysis covers all possible stepsizes and arbitrary upper/lower bounds on the curvature of the objective function, thus including convex, strongly convex and weakly convex (hypoconvex) objective functions. Unlike prior approaches, which rely solely on inequalities connecting consecutive iterations, our analysis employs inequalities involving an iterate and its two predecessors. While this complicates the proofs to some extent, it enables us to achieve, for the first time, an exact full-range analysis of gradient descent for any constant stepsizes (covering, in particular, normalized stepsizes greater than one), whereas the literature contained only conjectured rates of this type. In the nonconvex case, allowing arbitrary bounds on upper and lower curvatures extends existing partial results that are valid only for gradient Lipschitz functions (i.e., where lower and upper bounds on curvature are equal), leading to improved rates for weakly convex functions. From our exact rates, we deduce the optimal constant stepsize for gradient descent. Leveraging our analysis, we also introduce a new variant of gradient descent based on a unique, fixed sequence of variable stepsizes, demonstrating its superiority over any constant stepsize schedule., Comment: This paper replaces the arXiv submission of the authors: Tight convergence rates of the gradient method on smooth hypoconvex functions (arXiv:2203.00775)

In contrast with Mercer kernel-based approaches as used e.g., in Kernel Principal Component Analysis (KPCA), it was previously shown that Singular Value Decomposition (SVD) inherently relates to asymmetric kernels and Asymmetric Kernel Singular Value Decomposition (KSVD) has been proposed. However, the existing formulation to KSVD cannot work with infinite-dimensional feature mappings, the variational objective can be unbounded, and needs further numerical evaluation and exploration towards machine learning. In this work, i) we introduce a new asymmetric learning paradigm based on coupled covariance eigenproblem (CCE) through covariance operators, allowing infinite-dimensional feature maps. The solution to CCE is ultimately obtained from the SVD of the induced asymmetric kernel matrix, providing links to KSVD. ii) Starting from the integral equations corresponding to a pair of coupled adjoint eigenfunctions, we formalize the asymmetric Nystr\"om method through a finite sample approximation to speed up training. iii) We provide the first empirical evaluations verifying the practical utility and benefits of KSVD and compare with methods resorting to symmetrization or linear SVD across multiple tasks., Comment: 19 pages, 9 tables, 6 figures

We consider a difference-of-convex formulation where one of the terms is allowed to be hypoconvex (or weakly convex). We first examine the precise behavior of a single iteration of the Difference-of-Convex algorithm (DCA), giving a tight characterization of the objective function decrease. This requires distinguishing between eight distinct parameter regimes. Our proofs are inspired by the performance estimation framework, but are much simplified compared to similar previous work. We then derive sublinear DCA convergence rates towards critical points, distinguishing between cases where at least one of the functions is smooth and where both functions are nonsmooth. We conjecture the tightness of these rates for four parameter regimes, based on strong numerical evidence obtained via performance estimation, as well as the leading constant in the asymptotic sublinear rate for two more regimes.

In this work, we focus on the challenge of transferring an autonomous driving controller from simulation to the real world (i.e. Sim2Real). We propose a data-efficient method for online and on-the-fly adaptation of parametrizable control architectures such that the target closed-loop performance is optimized while accounting for uncertainties as model mismatches, changes in the environment, and task variations. The novelty of the approach resides in leveraging black-box optimization enabled by Executable Digital Twins (xDTs) for data-driven parameter calibration through derivative-free methods to directly adapt the controller in real-time. The xDTs are augmented with Domain Randomization for robustness and allow for safe parameter exploration. The proposed method requires a minimal amount of interaction with the real-world as it pushes the exploration towards the xDTs. We validate our approach through real-world experiments, demonstrating its effectiveness in transferring and fine-tuning a NMPC with 9 parameters, in under 10 minutes. This eliminates the need for hours-long manual tuning and lengthy machine learning training and data collection phases. Our results show that the online adapted NMPC directly compensates for the Sim2Real gap and avoids overtuning in simulation. Importantly, a 75% improvement in tracking performance is achieved and the Sim2Real gap over the target performance is reduced from a factor of 876 to 1.033., Comment: This work has been accepted by IEEE for possible publication

We show that adaptive proximal gradient methods for convex problems are not restricted to traditional Lipschitzian assumptions. Our analysis reveals that a class of linesearch-free methods is still convergent under mere local H\"older gradient continuity, covering in particular continuously differentiable semi-algebraic functions. To mitigate the lack of local Lipschitz continuity, popular approaches revolve around $\varepsilon$-oracles and/or linesearch procedures. In contrast, we exploit plain H\"older inequalities not entailing any approximation, all while retaining the linesearch-free nature of adaptive schemes. Furthermore, we prove full sequence convergence without prior knowledge of local H\"older constants nor of the order of H\"older continuity. Numerical experiments make comparisons with baseline methods on diverse tasks from machine learning covering both the locally and the globally H\"older setting.

Efficient markets are characterised by profit-driven participants continuously refining their positions towards the latest insights. Margins for profit generation are generally small, shaping a difficult landscape for automated trading strategies. This paper introduces a novel intraday power trading strategy tailored for single-price balancing markets. The strategy relies on a strategically devised mixture model to forecast future system imbalance prices and is formulated as a stochastic optimization problem with decision-dependent distributions to address two primary challenges: (i) the impact of trading positions on the system imbalance price and (ii) the uncertainty inherent in the model. The first challenge is tackled by adjusting the model to account for price changes after taking a position. For the second challenge, a coherent risk measure is added to the cost function to take additional uncertainties into account. This paper introduces a methodology to select the tuning parameter of this risk measure adaptively by continuously quantifying the performance of the strategy on a window of recently observed data. The strategy is validated with a simulation on the Belgian electricity market using real-time market data. The adaptive tuning approach leads to higher absolute profits, while also reducing the number of trades., Comment: Submitted to Applied Energy [Elsevier]

The autonomous driving industry is continuously dealing with safety-critical scenarios, and nonlinear model predictive control (NMPC) is a powerful control strategy for handling such situations. However, standard safety constraints are not scalable and require a long NMPC horizon. Moreover, the adoption of NMPC in the automotive industry is limited by the heavy computation of numerical optimization routines. To address those issues, this paper presents a real-time capable NMPC for automated driving in urban environments, using control barrier functions (CBFs). Furthermore, the designed NMPC is based on a novel collocation transcription approach, named RESAFE/COL, that allows to reduce the number of optimization variables while still guaranteeing the continuous time (nonlinear) inequality constraints satisfaction, through regional convex hull approximation. RESAFE/COL is proven to be 5 times faster than multiple shooting and more tractable for embedded hardware without a decrease in the performance, nor accuracy and safety of the numerical solution. We validate our NMPC-CBF with RESAFE/COL on digital twins of the vehicle and the urban environment and show the safe controller's ability to improve crash avoidance by 91\%. Supplementary visual material can be found at https://youtu.be/_EnbfYwljp4., Comment: 2024 the authors. This work has been accepted to IFAC for publication under a Creative Commons Licence CC-BY-NC-ND

This paper studies data-driven iterative learning control (ILC) for linear time-invariant (LTI) systems with unknown dynamics, output disturbances and input box-constraints. Our main contributions are: 1) using a non-parametric data-driven representation of the system dynamics, for dealing with the unknown system dynamics in the context of ILC, 2) design of a fast ILC method for dealing with output disturbances, model uncertainty and input constraints. A complete design method is given in this paper, which consists of the data-driven representation, controller formulation, acceleration strategy and convergence analysis. A batch of numerical experiments and a case study on a high-precision robotic motion system are given in the end to show the effectiveness of the proposed method.

In this paper we study an unconventional inexact Augmented Lagrangian Method (ALM) for convex optimization problems, as first proposed by Bertsekas, herein the penalty term is a potentially non-Euclidean norm raised to a power between one and two. We analyze the algorithm through the lens of a nonlinear Proximal Point Method (PPM), as originally introduced by Luque, applied to the dual problem. While Luque analyzes the order of local convergence of the scheme with Euclidean norms our focus is on the non-Euclidean case which prevents us from using standard tools for the analysis such as the nonexpansiveness of the proximal mapping. To allow for errors in the primal update, we derive two implementable stopping criteria under which we analyze both the global and the local convergence rates of the algorithm. More specifically, we show that the method enjoys a fast sublinear global rate in general and a local superlinear rate under suitable growth assumptions. We also highlight that the power ALM can be interpreted as classical ALM with an implicitly defined penalty-parameter schedule, reducing its parameter dependence. Our experiments on a number of relevant problems suggest that for certain powers the method performs similarly to a classical ALM with fine-tuned adaptive penalty rule, despite involving fewer parameters.

In this paper we study a nonlinear dual space preconditioning approach for the relaxed Proximal Point Algorithm (PPA) with application to monotone inclusions, called anisotropic PPA. The algorithm is an instance of Luque's nonlinear PPA wherein the nonlinear preconditioner is chosen as the gradient of a Legendre convex function. Since the preconditioned operator is nonmonotone in general, convergence cannot be shown using standard arguments, unless the preconditioner exhibits isotropy (preserves directions) as in existing literature. To address the broader applicability we show convergence along subsequences invoking a Bregman version of Fej\`er-monotonicity in the dual space. Via a nonlinear generalization of Moreau's decomposition for operators, we provide a dual interpretation of the algorithm in terms of a forward iteration applied to a $D$-firmly nonexpansive mapping which involves the Bregman resolvent. For a suitable preconditioner, convergence rates of arbitrary order are derived under a mild H\"older growth condition. Finally, we discuss an anisotropic generalization of the proximal augmented Lagrangian method obtained via the propsed scheme. This has strong connections to Rockafellar's generalized and sharp Lagrangian functions.

The Chambolle-Pock algorithm (CPA), also known as the primal-dual hybrid gradient method, has gained popularity over the last decade due to its success in solving large-scale convex structured problems. This work extends its convergence analysis for problems with varying degrees of (non)monotonicity, quantified through a so-called oblique weak Minty condition on the associated primal-dual operator. Our results reveal novel stepsize and relaxation parameter ranges which do not only depend on the norm of the linear mapping, but also on its other singular values. In particular, in nonmonotone settings, in addition to the classical stepsize conditions, extra bounds on the stepsizes and relaxation parameters are required. On the other hand, in the strongly monotone setting, the relaxation parameter is allowed to exceed the classical upper bound of two. Moreover, we build upon the recently introduced class of semimonotone operators, providing sufficient convergence conditions for CPA when the individual operators are semimonotone. Since this class of operators encompasses traditional operator classes including (hypo)- and co(hypo)-monotone operators, this analysis recovers and extends existing results for CPA. Tightness of the proposed stepsize ranges is demonstrated through several examples., Comment: 29 pages

Building upon recent works on linesearch-free adaptive proximal gradient methods, this paper proposes adaPG$^{q,r}$, a framework that unifies and extends existing results by providing larger stepsize policies and improved lower bounds. Different choices of the parameters $q$ and $r$ are discussed and the efficacy of the resulting methods is demonstrated through numerical simulations. In an attempt to better understand the underlying theory, its convergence is established in a more general setting that allows for time-varying parameters. Finally, an adaptive alternating minimization algorithm is presented by exploring the dual setting. This algorithm not only incorporates additional adaptivity, but also expands its applicability beyond standard strongly convex settings.

Distributed learning and adaptation have received significant interest and found wide-ranging applications in machine learning and signal processing. While various approaches, such as shared-memory optimization, multi-task learning, and consensus-based learning (e.g., federated learning and learning over graphs), focus on optimizing either local costs or a global cost, there remains a need for further exploration of their interconnections. This paper specifically focuses on a scenario where agents collaborate towards a common task (i.e., optimizing a global cost equal to aggregated local costs) while effectively having distinct individual tasks (i.e., optimizing individual local parameters in a local cost). Each agent's actions can potentially impact other agents' performance through interactions. Notably, each agent has access to only its local zeroth-order oracle (i.e., cost function value) and shares scalar values, rather than gradient vectors, with other agents, leading to communication bandwidth efficiency and agent privacy. Agents employ zeroth-order optimization to update their parameters, and the asynchronous message-passing between them is subject to bounded but possibly random communication delays. This paper presents theoretical convergence analyses and establishes a convergence rate for nonconvex problems. Furthermore, it addresses the relevant use-case of deep learning-based resource allocation in communication networks and conducts numerical experiments in which agents, acting as transmitters, collaboratively train their individual policies to maximize a global reward, e.g., a sum of data rates.

Deep learning (DL) based resource allocation (RA) has recently gained significant attention due to its performance efficiency. However, most related studies assume an ideal case where the number of users and their utility demands, e.g., data rate constraints, are fixed, and the designed DL-based RA scheme exploits a policy trained only for these fixed parameters. Consequently, computationally complex policy retraining is required whenever these parameters change. In this paper, we introduce a DL-based resource allocator (ALCOR) that allows users to adjust their utility demands freely, such as based on their application layer requirements. ALCOR employs deep neural networks (DNNs) as the policy in a time-sharing problem. The underlying optimization algorithm iteratively optimizes the on-off status of users to satisfy their utility demands in expectation. The policy performs unconstrained RA (URA)--RA without considering user utility demands--among active users to maximize the sum utility (SU) at each time instant. Depending on the chosen URA scheme, ALCOR can perform RA in either a centralized or distributed scenario. Derived convergence analyses provide guarantees for ALCOR's convergence, and numerical experiments corroborate its effectiveness.

In this paper we derive a moment relaxation for large-scale nonsmooth optimization problems with graphical structure and spherical constraints. In contrast to classical moment relaxations for global polynomial optimization that suffer from the curse of dimensionality we exploit the partially separable structure of the optimization problem to reduce the dimensionality of the search space. Leveraging optimal transport and Kantorovich--Rubinstein duality we decouple the problem and derive a tractable dual subspace approximation of the infinite-dimensional problem using spherical harmonics. This allows us to tackle possibly nonpolynomial optimization problems with spherical constraints and geodesic coupling terms. We show that the duality gap vanishes in the limit by proving that a Lipschitz continuous dual multiplier on a unit sphere can be approximated as closely as desired in terms of a Lipschitz continuous polynomial. The formulation is applied to sphere-valued imaging problems with total variation regularization and graph-based simultaneous localization and mapping (SLAM). In imaging tasks our approach achieves small duality gaps for a moderate degree. In graph-based SLAM our approach often finds solutions which after refinement with a local method are near the ground truth solution.

This work presents PANTR, an efficient solver for nonconvex constrained optimization problems, that is well-suited as an inner solver for an augmented Lagrangian method. The proposed scheme combines forward-backward iterations with solutions to trust-region subproblems: the former ensures global convergence, whereas the latter enables fast update directions. We discuss how the algorithm is able to exploit exact Hessian information of the smooth objective term through a linear Newton approximation, while benefiting from the structure of box-constraints or l1-regularization. An open-source C++ implementation of PANTR is made available as part of the NLP solver library ALPAQA. Finally, the effectiveness of the proposed method is demonstrated in nonlinear model predictive control applications., Comment: Accepted for publication in IEEE Control Systems Letters

Asymmetric data naturally exist in real life, such as directed graphs. Different from the common kernel methods requiring Mercer kernels, this paper tackles the asymmetric kernel-based learning problem. We describe a nonlinear extension of the matrix Singular Value Decomposition through asymmetric kernels, namely KSVD. First, we construct two nonlinear feature mappings w.r.t. rows and columns of the given data matrix. The proposed optimization problem maximizes the variance of each mapping projected onto the subspace spanned by the other, subject to a mutual orthogonality constraint. Through Lagrangian duality, we show that it can be solved by the left and right singular vectors in the feature space induced by the asymmetric kernel. Moreover, we start from the integral equations with a pair of adjoint eigenfunctions corresponding to the singular vectors on an asymmetrical kernel, and extend the Nystr\"om method to asymmetric cases through the finite sample approximation, which can be applied to speedup the training in KSVD. Experiments show that asymmetric KSVD learns features outperforming Mercer-kernel based methods that resort to symmetrization, and also verify the effectiveness of the asymmetric Nystr\"om method.

In the context of deep learning with kernel machines, the deep Restricted Kernel Machine (DRKM) framework allows multiple levels of kernel PCA (KPCA) and Least-Squares Support Vector Machines (LSSVM) to be combined into a deep architecture using visible and hidden units. We propose a new method for DRKM classification coupling the objectives of KPCA and classification levels, with the hidden feature matrix lying on the Stiefel manifold. The classification level can be formulated as an LSSVM or as an MLP feature map, combining depth in terms of levels and layers. The classification level is expressed in its primal formulation, as the deep KPCA levels, in their dual formulation, can embed the most informative components of the data in a much lower dimensional space. The dual setting is independent of the dimension of the inputs and the primal setting is parametric, which makes the proposed method computationally efficient for both high-dimensional inputs and large datasets. In the experiments, we show that our developed algorithm can effectively learn from small datasets, while using less memory than the convolutional neural network (CNN) with high-dimensional data. and that models with multiple KPCA levels can outperform models with a single level. On the tested larger-scale datasets, DRKM is more energy efficient than CNN while maintaining comparable performance.

The goal of this paper is to revisit Kernel Principal Component Analysis (KPCA) through dualization of a difference of convex functions. This allows to naturally extend KPCA to multiple objective functions and leads to efficient gradient-based algorithms avoiding the expensive SVD of the Gram matrix. Particularly, we consider objective functions that can be written as Moreau envelopes, demonstrating how to promote robustness and sparsity within the same framework. The proposed method is evaluated on synthetic and real-world benchmarks, showing significant speedup in KPCA training time as well as highlighting the benefits in terms of robustness and sparsity., Comment: 15 pages, ICML 2023

The proximal point algorithm (PPA) is the most widely recognized method for solving inclusion problems and serves as the foundation for many numerical algorithms. Despite this popularity, its convergence results have been largely limited to the monotone setting. In this work, we study the convergence of (relaxed) preconditioned PPA for a class of nonmonotone problems that satisfy an oblique weak Minty condition. Additionally, we study the (relaxed) Douglas-Rachford splitting (DRS) method in the nonmonotone setting by establishing a connection between DRS and the preconditioned PPA with a positive semidefinite preconditioner. To better characterize the class of problems covered by our analysis, we introduce the class of semimonotone operators, offering a natural extension to (hypo)monotone and co(hypo)monotone operators, and describe some of their properties. Sufficient conditions for global convergence of DRS involving the sum of two semimonotone operators are provided. Notably, it is shown that DRS converges even when the sum of the involved operators (or of their inverses) is nonmonotone. Various example problems are provided, demonstrating the tightness of our convergence results and highlighting the wide range of applications our theory is able to cover., Comment: 44 pages

This paper studies proximal gradient iterations for solving simple bilevel optimization problems where both the upper and the lower level cost functions are split as the sum of differentiable and (possibly nonsmooth) proximable functions. We develop a novel convergence recipe for iteration varying stepsizes that relies on Barzilai-Borwein type local estimates for the differentiable terms. Leveraging the convergence recipe, under global Lipschitz gradient continuity, we establish convergence for a nonadaptive stepsize sequence, without requiring any strong convexity or linesearch. In the locally Lipschitz differentiable setting, we develop an adaptive linesearch method that introduces a systematic adaptive scheme enabling large and nonmonotonic stepsize sequences while being insensitive to the choice of hyperparameters and initialization. Numerical simulations are provided showcasing favorable convergence speed of our methods.

We present a novel framework for distributionally robust optimization (DRO), called cost-aware DRO (CADRO). The key idea of CADRO is to exploit the cost structure in the design of the ambiguity set to reduce conservatism. Particularly, the set specifically constrains the worst-case distribution along the direction in which the expected cost of an approximate solution increases most rapidly. We prove that CADRO provides both a high-confidence upper bound and a consistent estimator of the out-of-sample expected cost, and show empirically that it produces solutions that are substantially less conservative than existing DRO methods, while providing the same guarantees., Comment: Revision; More general formulation and additional experiments

We consider the worst-case expectation of a permutation invariant ambiguity set of discrete distributions as a proxy-cost for data-driven expected risk minimization. For this framework, we coin the term ordered risk minimization to highlight how results from order statistics inspired the proxy-cost. Specifically, we show how such costs serve as point-wise high-confidence upper bounds of the expected risk. The confidence level can be determined tightly for any sample size. Conversely we also illustrate how to calibrate the size of the ambiguity set such that the high-confidence upper bound has some user specified confidence. This calibration procedure notably supports $\phi$-divergence based ambiguity sets. Numerical experiments then illustrate how the resulting scheme both generalizes better and is less sensitive to tuning parameters compared to the empirical risk minimization approach.

There is a growing interest in data-driven control of nonlinear systems over the last years. In contrast to related works, this paper takes a step back and aims to solve the output matching problem, a problem closely related to the reference tracking control problem, for a broader class of nonlinear systems called output-generalized bilinear, thereby offering a new direction to explore for data-driven control of nonlinear systems. It is shown that discrete time linear parameter-varying systems are included in this model class, with affine systems easily shown to also be included. This paper proposes a method to solve the output matching problem and offers a way to parameterize the solution set with a minimal number of parameters. The proposed model class and method are illustrated using simulations of two real-life systems., Comment: Submitted to ECC 2023

Principal Component Analysis (PCA) and its nonlinear extension Kernel PCA (KPCA) are widely used across science and industry for data analysis and dimensionality reduction. Modern deep learning tools have achieved great empirical success, but a framework for deep principal component analysis is still lacking. Here we develop a deep kernel PCA methodology (DKPCA) to extract multiple levels of the most informative components of the data. Our scheme can effectively identify new hierarchical variables, called deep principal components, capturing the main characteristics of high-dimensional data through a simple and interpretable numerical optimization. We couple the principal components of multiple KPCA levels, theoretically showing that DKPCA creates both forward and backward dependency across levels, which has not been explored in kernel methods and yet is crucial to extract more informative features. Various experimental evaluations on multiple data types show that DKPCA finds more efficient and disentangled representations with higher explained variance in fewer principal components, compared to the shallow KPCA. We demonstrate that our method allows for effective hierarchical data exploration, with the ability to separate the key generative factors of the input data both for large datasets and when few training samples are available. Overall, DKPCA can facilitate the extraction of useful patterns from high-dimensional data by learning more informative features organized in different levels, giving diversified aspects to explore the variation factors in the data, while maintaining a simple mathematical formulation.

This paper introduces a new extragradient-type algorithm for a class of nonconvex-nonconcave minimax problems. It is well-known that finding a local solution for general minimax problems is computationally intractable. This observation has recently motivated the study of structures sufficient for convergence of first order methods in the more general setting of variational inequalities when the so-called weak Minty variational inequality (MVI) holds. This problem class captures non-trivial structures as we demonstrate with examples, for which a large family of existing algorithms provably converge to limit cycles. Our results require a less restrictive parameter range in the weak MVI compared to what is previously known, thus extending the applicability of our scheme. The proposed algorithm is applicable to constrained and regularized problems, and involves an adaptive stepsize allowing for potentially larger stepsizes. Our scheme also converges globally even in settings where the underlying operator exhibits limit cycles., Comment: Code accessible at: https://github.com/LIONS-EPFL/weak-minty-code/

This paper introduces a family of stochastic extragradient-type algorithms for a class of nonconvex-nonconcave problems characterized by the weak Minty variational inequality (MVI). Unlike existing results on extragradient methods in the monotone setting, employing diminishing stepsizes is no longer possible in the weak MVI setting. This has led to approaches such as increasing batch sizes per iteration which can however be prohibitively expensive. In contrast, our proposed methods involves two stepsizes and only requires one additional oracle evaluation per iteration. We show that it is possible to keep one fixed stepsize while it is only the second stepsize that is taken to be diminishing, making it interesting even in the monotone setting. Almost sure convergence is established and we provide a unified analysis for this family of schemes which contains a nonlinear generalization of the celebrated primal dual hybrid gradient algorithm., Comment: Code accessible at: https://github.com/LIONS-EPFL/stochastic-weak-minty-code

We present a deep Graph Convolutional Kernel Machine (GCKM) for semi-supervised node classification in graphs. The method is built of two main types of blocks: (i) We introduce unsupervised kernel machine layers propagating the node features in a one-hop neighborhood, using implicit node feature mappings. (ii) We specify a semi-supervised classification kernel machine through the lens of the Fenchel-Young inequality. We derive an effective initialization scheme and efficient end-to-end training algorithm in the dual variables for the full architecture. The main idea underlying GCKM is that, because of the unsupervised core, the final model can achieve higher performance in semi-supervised node classification when few labels are available for training. Experimental results demonstrate the effectiveness of the proposed framework., Comment: Accepted for publication in AAAI 2024

Backtracking linesearch is the de facto approach for minimizing continuously differentiable functions with locally Lipschitz gradient. In recent years, it has been shown that in the convex setting it is possible to avoid linesearch altogether, and to allow the stepsize to adapt based on a local smoothness estimate without any backtracks or evaluations of the function value. In this work we propose an adaptive proximal gradient method, adaPG, that uses novel estimates of the local smoothness modulus which leads to less conservative stepsize updates and that can additionally cope with nonsmooth terms. This idea is extended to the primal-dual setting where an adaptive three-term primal-dual algorithm, adaPD, is proposed which can be viewed as an extension of the PDHG method. Moreover, in this setting the "essentially" fully adaptive variant adaPD$^+$ is proposed that avoids evaluating the linear operator norm by invoking a backtracking procedure, that, remarkably, does not require extra gradient evaluations. Numerical simulations demonstrate the effectiveness of the proposed algorithms compared to the state of the art.

PANOC is an algorithm for nonconvex optimization that has recently gained popularity in real-time control applications due to its fast, global convergence. The present work proposes a variant of PANOC that makes use of Gauss-Newton directions to accelerate the method. Furthermore, we show that when applied to optimal control problems, the computation of this Gauss-Newton step can be cast as a linear quadratic regulator (LQR) problem, allowing for an efficient solution through the Riccati recursion. Finally, we demonstrate that the proposed algorithm is more than twice as fast as the traditional L-BFGS variant of PANOC when applied to an optimal control benchmark problem, and that the performance scales favorably with increasing horizon length., Comment: Submitted to the 2023 IFAC World Congress, Yokohama

This paper proposes an accelerated version of Feasible Sequential Linear Programming (FSLP): the AA($d$)-FSLP algorithm. FSLP preserves feasibility in all intermediate iterates by means of an iterative update strategy which is based on repeated evaluation of zero-order information. This technique was successfully applied to techniques such as Model Predictive Control and Moving Horizon Estimation, but it can exhibit slow convergence. Moreover, keeping all iterates feasible in FSLP entails a large number of additional constraint evaluations. In this paper, Anderson Acceleration (AA($d$)) is applied to the zero-order update strategy improving the convergence rate and therefore decreasing the number of constraint evaluations in the inner iterative procedure of the FSLP algorithm. AA($d$) achieves an improved contraction rate in the inner iterations, with proven local linear convergence. In addition, it is observed that due to the improved zero-order update strategy, AA($d$)-FSLP takes larger steps to find an optimal solution, yielding faster overall convergence. The performance of AA($d$)-FSLP is examined for a time-optimal point-to-point motion problem of a parallel SCARA robot. The reduction of the number of constraint evaluations and overall iterations compared to FSLP is successfully demonstrated., Comment: Submitted to the IFAC World Congress 2023

This paper presents a novel model-free and fully data-driven policy iteration scheme for quadratic regulation of linear dynamics with state- and input-multiplicative noise. The implementation is similar to the least-squares temporal difference scheme for Markov decision processes, estimating Q-functions by solving a least-squares problem with instrumental variables. The scheme is compared with a model-based system identification scheme and natural policy gradient through numerical experiments.

Risk-averse optimal control problems have gained a lot of attention in the last decade, mostly due to their attractive mathematical properties and practical importance. They can be seen as an interpolation between stochastic and robust optimal control approaches, allowing the designer to trade-off performance for robustness and vice-versa. Due to their stochastic nature, risk-averse problems are of a very large scale, involving millions of decision variables, which poses a challenge in terms of efficient computation. In this work, we propose a splitting for general risk-averse problems and show how to efficiently compute iterates on a GPU-enabled hardware. Moreover, we propose Spock - a new algorithm that utilizes the proposed splitting and takes advantage of the SuperMann scheme combined with fast directions from Anderson's acceleration method for enhanced convergence speed. We implement Spock in Julia as an open-source solver, which is amenable to warm-starting and massive parallelization.

Lane changing and lane merging remains a challenging task for autonomous driving, due to the strong interaction between the controlled vehicle and the uncertain behavior of the surrounding traffic participants. The interaction induces a dependence of the vehicles' states on the (stochastic) dynamics of the surrounding vehicles, increasing the difficulty of predicting future trajectories. Furthermore, the small relative distances cause traditional robust approaches to become overly conservative, necessitating control methods that are explicitly aware of inter-vehicle interaction. Towards these goals, we propose an interaction-aware stochastic model predictive control (MPC) strategy integrated with an online learning framework, which models a given driver's cooperation level as an unknown parameter in a state-dependent probability distribution. The online learning framework adaptively estimates the surrounding vehicle's cooperation level with the vehicle's past trajectory and combines this with a kinematic vehicle model to predict the probability of a multimodal future state trajectory. The learning is conducted with logistic regression which enables fast online computation. The multi-future prediction is used in the MPC algorithm to compute the optimal control input while satisfying safety constraints. We demonstrate our algorithm in an interactive lane changing scenario with drivers in different randomly selected cooperation levels.

Orthogonal collocation methods are direct approaches for solving optimal control problems (OCP). A high solution accuracy is achieved with few optimization variables, making it more favorable for embedded and real-time NMPC applications. However, collocation approaches lack a guarantee about the safety of the resulting trajectory as inequality constraints are only set on a finite number of collocation points. In this paper we propose a method to efficiently create a convex safety envelope containing the trajectory such that the solution fully satisfies the OCP constraints. We make use of Bernstein approximations of a polynomial's extrema and span the solution over an orthogonal basis using Legendre polynomials. The tightness of the safety envelope estimation, high accuracy in solving the underlying differential equations, fast rate of convergence and little conservatism are properties of the presented approach making it a suitable method for safe real-time NMPC deployment. We show that our method has comparable computational performance to pseudospectral approaches and can accurately approximate the original OCP up to 9 times more quickly than standard multiple-shooting method in autonomous driving applications, without adding complexity to the formulation., Comment: This paper appears in the proceedings of the 2023 European Control Conference (ECC)