Stochastic networks and queueing systems often lead to Markov decision processes (MDPs) with large state and action spaces as well as nonconvex objective functions, which hinders the convergence of many reinforcement learning (RL) algorithms. Policy-gradient methods perform well on MDPs with large state and action spaces, but they sometimes experience slow convergence due to the high variance of the gradient estimator. In this paper, we show that some of these difficulties can be circumvented by exploiting the structure of the underlying MDP. We first introduce a new family of gradient estimators called score-aware gradient estimators (SAGEs). When the stationary distribution of the MDP belongs to an exponential family parametrized by the policy parameters, SAGEs allow us to estimate the policy gradient without relying on value-function estimation, contrary to classical policy-gradient methods like actor-critic. To demonstrate their applicability, we examine two common control problems arising in stochastic networks and queueing systems whose stationary distributions have a product-form, a special case of exponential families. As a second contribution, we show that, under appropriate assumptions, the policy under a SAGE-based policy-gradient method has a large probability of converging to an optimal policy, provided that it starts sufficiently close to it, even with a nonconvex objective function and multiple maximizers. Our key assumptions are that, locally around a maximizer, a nondegeneracy property of the Hessian of the objective function holds and a Lyapunov function exists. Finally, we conduct a numerical comparison between a SAGE-based policy-gradient method and an actor-critic algorithm. The results demonstrate that the SAGE-based method finds close-to-optimal policies more rapidly, highlighting its superior performance over the traditional actor-critic method.

Reward evaluation of episodes becomes a bottleneck in a broad range of reinforcement learning tasks. Our aim in this paper is to select a small but representative subset of a large batch of episodes, only on which we actually compute rewards for more efficient policy gradient iterations. We build a Gaussian process modeling of discounted returns or rewards to derive a positive definite kernel on the space of episodes, run an ``episodic" kernel quadrature method to compress the information of sample episodes, and pass the reduced episodes to the policy network for gradient updates. We present the theoretical background of this procedure as well as its numerical illustrations in MuJoCo tasks.

This paper studies distributed online learning under Byzantine attacks. The performance of an online learning algorithm is often characterized by (adversarial) regret, which evaluates the quality of one-step-ahead decision-making when an environment provides adversarial losses, and a sublinear bound is preferred. But we prove that, even with a class of state-of-the-art robust aggregation rules, in an adversarial environment and in the presence of Byzantine participants, distributed online gradient descent can only achieve a linear adversarial regret bound, which is tight. This is the inevitable consequence of Byzantine attacks, even though we can control the constant of the linear adversarial regret to a reasonable level. Interestingly, when the environment is not fully adversarial so that the losses of the honest participants are i.i.d. (independent and identically distributed), we show that sublinear stochastic regret, in contrast to the aforementioned adversarial regret, is possible. We develop a Byzantine-robust distributed online momentum algorithm to attain such a sublinear stochastic regret bound. Extensive numerical experiments corroborate our theoretical analysis.

Stochastic Gradient Descent (SGD) and its variants are almost universally used to train neural networks and to fit a variety of other parametric models. An important hyperparameter in this context is the batch size, which determines how many samples are processed before an update of the parameters occurs. Previous studies have demonstrated the benefits of using variable batch sizes. In this work, we will theoretically derive optimal batch size schedules for SGD and similar algorithms, up to an error that is quadratic in the learning rate. To achieve this, we approximate the discrete process of parameter updates using a family of stochastic differential equations indexed by the learning rate. To better handle the state-dependent diffusion coefficient, we further expand the solution of this family into a series with respect to the learning rate. Using this setup, we derive a continuous-time optimal batch size schedule for a large family of diffusion coefficients and then apply the results in the setting of linear regression.

Adaptive optimizers, such as Adam, have achieved remarkable success in deep learning. A key component of these optimizers is the so-called preconditioning matrix, providing enhanced gradient information and regulating the step size of each gradient direction. In this paper, we propose a novel approach to designing the preconditioning matrix by utilizing the gradient difference between two successive steps as the diagonal elements. These diagonal elements are closely related to the Hessian and can be perceived as an approximation of the inner product between the Hessian row vectors and difference of the adjacent parameter vectors. Additionally, we introduce an auto-switching function that enables the preconditioning matrix to switch dynamically between Stochastic Gradient Descent (SGD) and the adaptive optimizer. Based on these two techniques, we develop a new optimizer named AGD that enhances the generalization performance. We evaluate AGD on public datasets of Natural Language Processing (NLP), Computer Vision (CV), and Recommendation Systems (RecSys). Our experimental results demonstrate that AGD outperforms the state-of-the-art (SOTA) optimizers, achieving highly competitive or significantly better predictive performance. Furthermore, we analyze how AGD is able to switch automatically between SGD and the adaptive optimizer and its actual effects on various scenarios.

Selecting the best hyperparameters for a particular optimization instance, such as the learning rate and momentum, is an important but nonconvex problem. As a result, iterative optimization methods such as hypergradient descent lack global optimality guarantees in general. We propose an online nonstochastic control methodology for mathematical optimization. First, we formalize the setting of meta-optimization, an online learning formulation of learning the best optimization algorithm from a class of methods. The meta-optimization problem over gradient-based methods can be framed as a feedback control problem over the choice of hyperparameters, including the learning rate, momentum, and the preconditioner. Although the original optimal control problem is nonconvex, we show how recent methods from online nonstochastic control using convex relaxations can be used to overcome the challenge of nonconvexity, and obtain regret guarantees against the best offline solution. This guarantees that in meta-optimization, given a sequence of optimization problems, we can learn a method that attains convergence comparable to that of the best optimization method in hindsight from a class of methods.

Several recent works have studied the convergence \textit{in high probability} of stochastic gradient descent (SGD) and its clipped variant. Compared to vanilla SGD, clipped SGD is practically more stable and has the additional theoretical benefit of logarithmic dependence on the failure probability. However, the convergence of other practical nonlinear variants of SGD, e.g., sign SGD, quantized SGD and normalized SGD, that achieve improved communication efficiency or accelerated convergence is much less understood. In this work, we study the convergence bounds \textit{in high probability} of a broad class of nonlinear SGD methods. For strongly convex loss functions with Lipschitz continuous gradients, we prove a logarithmic dependence on the failure probability, even when the noise is heavy-tailed. Strictly more general than the results for clipped SGD, our results hold for any nonlinearity with bounded (component-wise or joint) outputs, such as clipping, normalization, and quantization. Further, existing results with heavy-tailed noise assume bounded $\eta$-th central moments, with $\eta \in (1,2]$. In contrast, our refined analysis works even for $\eta=1$, strictly relaxing the noise moment assumptions in the literature.

Neural networks have demonstrated success in various domains, yet their performance can be significantly degraded by even a small input perturbation. Consequently, the construction of such perturbations, known as adversarial attacks, has gained significant attention, many of which fall within "white-box" scenarios where we have full access to the neural network. Existing attack algorithms, such as the projected gradient descent (PGD), commonly take the sign function on the raw gradient before updating adversarial inputs, thereby neglecting gradient magnitude information. In this paper, we present a theoretical analysis of how such sign-based update algorithm influences step-wise attack performance, as well as its caveat. We also interpret why previous attempts of directly using raw gradients failed. Based on that, we further propose a new raw gradient descent (RGD) algorithm that eliminates the use of sign. Specifically, we convert the constrained optimization problem into an unconstrained one, by introducing a new hidden variable of non-clipped perturbation that can move beyond the constraint. The effectiveness of the proposed RGD algorithm has been demonstrated extensively in experiments, outperforming PGD and other competitors in various settings, without incurring any additional computational overhead. The codes is available in https://github.com/JunjieYang97/RGD.

In this paper, we extend mean-field Langevin dynamics to minimax optimization over probability distributions for the first time with symmetric and provably convergent updates. We propose mean-field Langevin averaged gradient (MFL-AG), a single-loop algorithm that implements gradient descent ascent in the distribution spaces with a novel weighted averaging, and establish average-iterate convergence to the mixed Nash equilibrium. We also study both time and particle discretization regimes and prove a new uniform-in-time propagation of chaos result which accounts for the dependency of the particle interactions on all previous distributions. Furthermore, we propose mean-field Langevin anchored best response (MFL-ABR), a symmetric double-loop algorithm based on best response dynamics with linear last-iterate convergence. Finally, we study applications to zero-sum Markov games and conduct simulations demonstrating long-term optimality.

Understanding the generalization abilities of modern machine learning algorithms has been a major research topic over the past decades. In recent years, the learning dynamics of Stochastic Gradient Descent (SGD) have been related to heavy-tailed dynamics. This has been successfully applied to generalization theory by exploiting the fractal properties of those dynamics. However, the derived bounds depend on mutual information (decoupling) terms that are beyond the reach of computability. In this work, we prove generalization bounds over the trajectory of a class of heavy-tailed dynamics, without those mutual information terms. Instead, we introduce a geometric decoupling term by comparing the learning dynamics (depending on the empirical risk) with an expected one (depending on the population risk). We further upper-bound this geometric term, by using techniques from the heavy-tailed and the fractal literature, making it fully computable. Moreover, as an attempt to tighten the bounds, we propose a PAC-Bayesian setting based on perturbed dynamics, in which the same geometric term plays a crucial role and can still be bounded using the techniques described above.