Anas Barakat

Ilyas Fatkhullin, Anas Barakat, Anastasia Kireeva et al. · eth-zurich

Recently, the impressive empirical success of policy gradient (PG) methods has catalyzed the development of their theoretical foundations. Despite the huge efforts directed at the design of efficient stochastic PG-type algorithms, the understanding of their convergence to a globally optimal policy is still limited. In this work, we develop improved global convergence guarantees for a general class of Fisher-non-degenerate parameterized policies which allows to address the case of continuous state action spaces. First, we propose a Normalized Policy Gradient method with Implicit Gradient Transport (N-PG-IGT) and derive a $\tilde{\mathcal{O}}(\varepsilon^{-2.5})$ sample complexity of this method for finding a global $\varepsilon$-optimal policy. Improving over the previously known $\tilde{\mathcal{O}}(\varepsilon^{-3})$ complexity, this algorithm does not require the use of importance sampling or second-order information and samples only one trajectory per iteration. Second, we further improve this complexity to $\tilde{ \mathcal{\mathcal{O}} }(\varepsilon^{-2})$ by considering a Hessian-Aided Recursive Policy Gradient ((N)-HARPG) algorithm enhanced with a correction based on a Hessian-vector product. Interestingly, both algorithms are $(i)$ simple and easy to implement: single-loop, do not require large batches of trajectories and sample at most two trajectories per iteration; $(ii)$ computationally and memory efficient: they do not require expensive subroutines at each iteration and can be implemented with memory linear in the dimension of parameters.

Anas Barakat, Ilyas Fatkhullin, Niao He · eth-zurich

We consider the reinforcement learning (RL) problem with general utilities which consists in maximizing a function of the state-action occupancy measure. Beyond the standard cumulative reward RL setting, this problem includes as particular cases constrained RL, pure exploration and learning from demonstrations among others. For this problem, we propose a simpler single-loop parameter-free normalized policy gradient algorithm. Implementing a recursive momentum variance reduction mechanism, our algorithm achieves $\tilde{\mathcal{O}}(ε^{-3})$ and $\tilde{\mathcal{O}}(ε^{-2})$ sample complexities for $ε$-first-order stationarity and $ε$-global optimality respectively, under adequate assumptions. We further address the setting of large finite state action spaces via linear function approximation of the occupancy measure and show a $\tilde{\mathcal{O}}(ε^{-4})$ sample complexity for a simple policy gradient method with a linear regression subroutine.

Jiduan Wu, Anas Barakat, Ilyas Fatkhullin et al. · eth-zurich

Zero-sum Linear Quadratic (LQ) games are fundamental in optimal control and can be used (i)~as a dynamic game formulation for risk-sensitive or robust control and (ii)~as a benchmark setting for multi-agent reinforcement learning with two competing agents in continuous state-control spaces. In contrast to the well-studied single-agent linear quadratic regulator problem, zero-sum LQ games entail solving a challenging nonconvex-nonconcave min-max problem with an objective function that lacks coercivity. Recently, Zhang et al. showed that an~$ε$-Nash equilibrium (NE) of finite horizon zero-sum LQ games can be learned via nested model-free Natural Policy Gradient (NPG) algorithms with poly$(1/ε)$ sample complexity. In this work, we propose a simpler nested Zeroth-Order (ZO) algorithm improving sample complexity by several orders of magnitude and guaranteeing convergence of the last iterate. Our main results are two-fold: (i) in the deterministic setting, we establish the first global last-iterate linear convergence result for the nested algorithm that seeks NE of zero-sum LQ games; (ii) in the model-free setting, we establish a~$\widetilde{\mathcal{O}}(ε^{-2})$ sample complexity using a single-point ZO estimator. For our last-iterate convergence results, our analysis leverages the Implicit Regularization (IR) property and a new gradient domination condition for the primal function. Our key improvements in the sample complexity rely on a more sample-efficient nested algorithm design and a finer control of the ZO natural gradient estimation error utilizing the structure endowed by the finite-horizon setting.

Anas Barakat, Souradip Chakraborty, Khushbu Pahwa et al.

Pass@k is a widely used performance metric for verifiable large language model tasks, including mathematical reasoning, code generation, and short-answer reasoning. It defines success if any of $k$ independently sampled solutions passes a verifier. This multi-sample inference metric has motivated inference-aware fine-tuning methods that directly optimize pass@$k$. However, prior work reports a recurring trade-off: pass@k improves while pass@1 degrades under such methods. This trade-off is practically important because pass@1 often remains a hard operational constraint due to latency and cost budgets, imperfect verifier coverage, and the need for a reliable single-shot fallback. We study the origin of this trade-off and provide a theoretical characterization of when pass@k policy optimization can reduce pass@1 through gradient conflict induced by prompt interference. We show that pass@$k$ policy gradients can conflict with pass@1 gradients because pass@$k$ optimization implicitly reweights prompts toward low-success prompts; when these prompts are what we term negatively interfering, their upweighting can rotate the pass@k update direction away from the pass@1 direction. We illustrate our theoretical findings with large language model experiments on verifiable mathematical reasoning tasks.

Anas Barakat, Pascal Bianchi, Julien Lehmann

Actor-critic methods integrating target networks have exhibited a stupendous empirical success in deep reinforcement learning. However, a theoretical understanding of the use of target networks in actor-critic methods is largely missing in the literature. In this paper, we reduce this gap between theory and practice by proposing the first theoretical analysis of an online target-based actor-critic algorithm with linear function approximation in the discounted reward setting. Our algorithm uses three different timescales: one for the actor and two for the critic. Instead of using the standard single timescale temporal difference (TD) learning algorithm as a critic, we use a two timescales target-based version of TD learning closely inspired from practical actor-critic algorithms implementing target networks. First, we establish asymptotic convergence results for both the critic and the actor under Markovian sampling. Then, we provide a finite-time analysis showing the impact of incorporating a target network into actor-critic methods.

A. Barakat, P. Bianchi, W. Hachem et al.

In this paper, a general stochastic optimization procedure is studied, unifying several variants of the stochastic gradient descent such as, among others, the stochastic heavy ball method, the Stochastic Nesterov Accelerated Gradient algorithm (S-NAG), and the widely used Adam algorithm. The algorithm is seen as a noisy Euler discretization of a non-autonomous ordinary differential equation, recently introduced by Belotto da Silva and Gazeau, which is analyzed in depth. Assuming that the objective function is non-convex and differentiable, the stability and the almost sure convergence of the iterates to the set of critical points are established. A noteworthy special case is the convergence proof of S-NAG in a non-convex setting. Under some assumptions, the convergence rate is provided under the form of a Central Limit Theorem. Finally, the non-convergence of the algorithm to undesired critical points, such as local maxima or saddle points, is established. Here, the main ingredient is a new avoidance of traps result for non-autonomous settings, which is of independent interest.

Anas Barakat, Pascal Bianchi

Although ADAM is a very popular algorithm for optimizing the weights of neural networks, it has been recently shown that it can diverge even in simple convex optimization examples. Several variants of ADAM have been proposed to circumvent this convergence issue. In this work, we study the ADAM algorithm for smooth nonconvex optimization under a boundedness assumption on the adaptive learning rate. The bound on the adaptive step size depends on the Lipschitz constant of the gradient of the objective function and provides safe theoretical adaptive step sizes. Under this boundedness assumption, we show a novel first order convergence rate result in both deterministic and stochastic contexts. Furthermore, we establish convergence rates of the function value sequence using the Kurdyka-Lojasiewicz property.

Anas Barakat, Pascal Bianchi

Adam is a popular variant of stochastic gradient descent for finding a local minimizer of a function. In the constant stepsize regime, assuming that the objective function is differentiable and non-convex, we establish the convergence in the long run of the iterates to a stationary point under a stability condition. The key ingredient is the introduction of a continuous-time version of Adam, under the form of a non-autonomous ordinary differential equation. This continuous-time system is a relevant approximation of the Adam iterates, in the sense that the interpolated Adam process converges weakly towards the solution to the ODE. The existence and the uniqueness of the solution are established. We further show the convergence of the solution towards the critical points of the objective function and quantify its convergence rate under a Lojasiewicz assumption. Then, we introduce a novel decreasing stepsize version of Adam. Under mild assumptions, it is shown that the iterates are almost surely bounded and converge almost surely to critical points of the objective function. Finally, we analyze the fluctuations of the algorithm by means of a conditional central limit theorem.