This article provides a comprehensive understanding of optimization in deep learning, with a primary focus on the challenges of gradient vanishing and gradient exploding, which normally lead to diminished model representational ability and training instability, respectively. We analyze these two challenges through several strategic measures, including the improvement of gradient flow and the imposition of constraints on a network's Lipschitz constant. To help understand the current optimization methodologies, we categorize them into two classes: explicit optimization and implicit optimization. Explicit optimization methods involve direct manipulation of optimizer parameters, including weight, gradient, learning rate, and weight decay. Implicit optimization methods, by contrast, focus on improving the overall landscape of a network by enhancing its modules, such as residual shortcuts, normalization methods, attention mechanisms, and activations. In this article, we provide an in-depth analysis of these two optimization classes and undertake a thorough examination of the Jacobian matrices and the Lipschitz constants of many widely used deep learning modules, highlighting existing issues as well as potential improvements. Moreover, we also conduct a series of analytical experiments to substantiate our theoretical discussions. This article does not aim to propose a new optimizer or network. Rather, our intention is to present a comprehensive understanding of optimization in deep learning. We hope that this article will assist readers in gaining a deeper insight in this field and encourages the development of more robust, efficient, and high-performing models.

Single-call stochastic extragradient methods, like stochastic past extragradient (SPEG) and stochastic optimistic gradient (SOG), have gained a lot of interest in recent years and are one of the most efficient algorithms for solving large-scale min-max optimization and variational inequalities problems (VIP) appearing in various machine learning tasks. However, despite their undoubted popularity, current convergence analyses of SPEG and SOG require a bounded variance assumption. In addition, several important questions regarding the convergence properties of these methods are still open, including mini-batching, efficient step-size selection, and convergence guarantees under different sampling strategies. In this work, we address these questions and provide convergence guarantees for two large classes of structured non-monotone VIPs: (i) quasi-strongly monotone problems (a generalization of strongly monotone problems) and (ii) weak Minty variational inequalities (a generalization of monotone and Minty VIPs). We introduce the expected residual condition, explain its benefits, and show how it can be used to obtain a strictly weaker bound than previously used growth conditions, expected co-coercivity, or bounded variance assumptions. Equipped with this condition, we provide theoretical guarantees for the convergence of single-call extragradient methods for different step-size selections, including constant, decreasing, and step-size-switching rules. Furthermore, our convergence analysis holds under the arbitrary sampling paradigm, which includes importance sampling and various mini-batching strategies as special cases.

Description of the problem and related works Fitting a system of nonparametric ordinary differential equations (ODEs) ẋ = f (t, x) to longitudinal data could lead to scientific breakthroughs in disciplines where ODEs or dynamical systems have been used for a long time, including physics, chemistry, and biology, see [1]. By nonparametric, we mean that there is no need to specify the functional form of the vector-field f using a pre-defined finite dimensional parameter. Instead, this force field belongs to a functional space and the number of parameters that characterize this vector field depends on the amount of data available. This provides a great advantage in situations where the form of the vector field is unknown but data is available for learning. The functional spaces considered are Reproducing Kernel Hilbert Spaces (RKHS) [2], allowing for efficient optimization among other desirable properties. A particular difficulty arises when the data is sparse and noisy. This is often the case for longitudinal healthcare data obtained during hospital visits.

We present an integrated prediction-optimization (PredOpt) framework to efficiently solve sequential decision-making problems by predicting the values of binary decision variables in an optimal solution. We address the key issues of sequential dependence, infeasibility, and generalization in machine learning (ML) to make predictions for optimal solutions to combinatorial problems. The sequential nature of the combinatorial optimization problems considered is captured with recurrent neural networks and a sliding-attention window. We integrate an attention-based encoder-decoder neural network architecture with an infeasibility-elimination and generalization framework to learn high-quality feasible solutions to time-dependent optimization problems. In this framework, the required level of predictions is optimized to eliminate the infeasibility of the ML predictions. These predictions are then fixed in mixed-integer programming (MIP) problems to solve them quickly with the aid of a commercial solver. We demonstrate our approach to tackling the two well-known dynamic NP-Hard optimization problems: multi-item capacitated lot-sizing (MCLSP) and multi-dimensional knapsack (MSMK). Our results show that models trained on shorter and smaller-dimensional instances can be successfully used to predict longer and larger-dimensional problems. The solution time can be reduced by three orders of magnitude with an average optimality gap below 0.1%. We compare PredOpt with various specially designed heuristics and show that our framework outperforms them. PredOpt can be advantageous for solving dynamic MIP problems that need to be solved instantly and repetitively.

Over the recent past data-driven algorithms for solving stochastic optimal control problems in face of model uncertainty have become an increasingly active area of research. However, for singular controls and underlying diffusion dynamics the analysis has so far been restricted to the scalar case. In this paper we fill this gap by studying a multivariate singular control problem for reversible diffusions with controls of reflection type. Our contributions are threefold. We first explicitly determine the long-run average costs as a domain-dependent functional, showing that the control problem can be equivalently characterized as a shape optimization problem. For given diffusion dynamics, assuming the optimal domain to be strongly star-shaped, we then propose a gradient descent algorithm based on polytope approximations to numerically determine a cost-minimizing domain. Finally, we investigate data-driven solutions when the diffusion dynamics are unknown to the controller. Using techniques from nonparametric statistics for stochastic processes, we construct an optimal domain estimator, whose static regret is bounded by the minimax optimal estimation rate of the unreflected process' invariant density. In the most challenging situation, when the dynamics must be learned simultaneously to controlling the process, we develop an episodic learning algorithm to overcome the emerging exploration-exploitation dilemma and show that given the static regret as a baseline, the loss in its sublinear regret per time unit is of natural order compared to the one-dimensional case.

Implementing intelligent control of robots is a difficult task, especially when dealing with complex black-box systems, because of the lack of visibility and understanding of how these robots work internally. This paper proposes an Intelligent Social Learning (ISL) algorithm to enable intelligent control of black-box robotic systems. Inspired by mutual learning among individuals in human social groups, ISL includes learning, imitation, and self-study styles. Individuals in the learning style use the Levy flight search strategy to learn from the best performer and form the closest relationships. In the imitation style, individuals mimic the best performer with a second-level rapport by employing a random perturbation strategy. In the self-study style, individuals learn independently using a normal distribution sampling method while maintaining a distant relationship with the best performer. Individuals in the population are regarded as autonomous intelligent agents in each style. Neural networks perform strategic actions in three styles to interact with the environment and the robot and iteratively optimize the network policy. Overall, ISL builds on the principles of intelligent optimization, incorporating ideas from reinforcement learning, and possesses strong search capabilities, fast computation speed, fewer hyperparameters, and insensitivity to sparse rewards. The proposed ISL algorithm is compared with four state-of-the-art methods on six continuous control benchmark cases in MuJoCo to verify its effectiveness and advantages. Furthermore, ISL is adopted in the simulation and experimental grasping tasks of the UR3 robot for validations, and satisfactory solutions are yielded.

Gaussian Mixture Models (GMMs) are one of the most potent parametric density models used extensively in many applications. Flexibly-tied factorization of the covariance matrices in GMMs is a powerful approach for coping with the challenges of common GMMs when faced with high-dimensional data and complex densities which often demand a large number of Gaussian components. However, the expectation-maximization algorithm for fitting flexibly-tied GMMs still encounters difficulties with streaming and very large dimensional data. To overcome these challenges, this paper suggests the use of first-order stochastic optimization algorithms. Specifically, we propose a new stochastic optimization algorithm on the manifold of orthogonal matrices. Through numerous empirical results on both synthetic and real datasets, we observe that stochastic optimization methods can outperform the expectation-maximization algorithm in terms of attaining better likelihood, needing fewer epochs for convergence, and consuming less time per each epoch.

Parameterized convex minorant (PCM) method is proposed for the approximation of the objective function in amortized optimization. In the proposed method, the objective function approximator is expressed by the sum of a PCM and a nonnegative gap function, where the objective function approximator is bounded from below by the PCM convex in the optimization variable. The proposed objective function approximator is a universal approximator for continuous functions, and the global minimizer of the PCM attains the global minimum of the objective function approximator. Therefore, the global minimizer of the objective function approximator can be obtained by a single convex optimization. As a realization of the proposed method, extended parameterized log-sum-exp network is proposed by utilizing a parameterized log-sum-exp network as the PCM. Numerical simulation is performed for parameterized non-convex objective function approximation and for learning-based nonlinear model predictive control to demonstrate the performance and characteristics of the proposed method. The simulation results support that the proposed method can be used to learn objective functions and to find a global minimizer reliably and quickly by using convex optimization algorithms.

In this paper, we present a novel Reduced Robustified NMPC (R$^2$NMPC) algorithm that has the same complexity as an equivalent nominal NMPC while enhancing it with robustified constraints based on the dynamics of ellipsoidal uncertainty sets. This promises both a closed-loop- and constraint satisfaction performance equivalent to common Robustified NMPC approaches, while drastically reducing the computational complexity. The main idea lies in approximating the ellipsoidal uncertainty sets propagation over the prediction horizon with the system dynamics' sensitivities inferred from the last optimal control problem (OCP) solution, and similarly for the gradients to robustify the constraints. Thus, we do not require the decision variables related to the uncertainty propagation within the OCP, rendering it computationally tractable. Next, we illustrate the real-time control capabilities of our algorithm in handling a complex, high-dimensional, and highly nonlinear system, namely the trajectory following of an autonomous passenger vehicle modeled with a dynamic nonlinear single-track model. Our experimental findings, alongside a comparative assessment against other Robust NMPC approaches, affirm the robustness of our method in effectively tracking an optimal racetrack trajectory while satisfying the nonlinear constraints. This performance is achieved while fully utilizing the vehicle's interface limits, even at high speeds of up to 37.5m/s, and successfully managing state estimation disturbances. Remarkably, our approach maintains a mean solving frequency of 144Hz.

In this paper, we propose an interactive genetic algorithm for solving multi-objective combinatorial optimization problems under preference imprecision. More precisely, we consider problems where the decision maker's preferences over solutions can be represented by a parameterized aggregation function (e.g., a weighted sum, an OWA operator, a Choquet integral), and we assume that the parameters are initially not known by the recommendation system. In order to quickly make a good recommendation, we combine elicitation and search in the following way: 1) we use regret-based elicitation techniques to reduce the parameter space in a efficient way, 2) genetic operators are applied on parameter instances (instead of solutions) to better explore the parameter space, and 3) we generate promising solutions (population) using existing solving methods designed for the problem with known preferences. Our algorithm, called RIGA, can be applied to any multi-objective combinatorial optimization problem provided that the aggregation function is linear in its parameters and that a (near-)optimal solution can be efficiently determined for the problem with known preferences. We also study its theoretical performances: RIGA can be implemented in such way that it runs in polynomial time while asking no more than a polynomial number of queries. The method is tested on the multi-objective knapsack and traveling salesman problems. For several performance indicators (computation times, gap to optimality and number of queries), RIGA obtains better results than state-of-the-art algorithms.