English

Actor-Critics Can Achieve Optimal Sample Efficiency

Machine Learning 2025-05-07 v1 Artificial Intelligence Machine Learning

Abstract

Actor-critic algorithms have become a cornerstone in reinforcement learning (RL), leveraging the strengths of both policy-based and value-based methods. Despite recent progress in understanding their statistical efficiency, no existing work has successfully learned an ϵ\epsilon-optimal policy with a sample complexity of O(1/ϵ2)O(1/\epsilon^2) trajectories with general function approximation when strategic exploration is necessary. We address this open problem by introducing a novel actor-critic algorithm that attains a sample-complexity of O(dH5logA/ϵ2+dH4logF/ϵ2)O(dH^5 \log|\mathcal{A}|/\epsilon^2 + d H^4 \log|\mathcal{F}|/ \epsilon^2) trajectories, and accompanying T\sqrt{T} regret when the Bellman eluder dimension dd does not increase with TT at more than a logT\log T rate. Here, F\mathcal{F} is the critic function class, A\mathcal{A} is the action space, and HH is the horizon in the finite horizon MDP setting. Our algorithm integrates optimism, off-policy critic estimation targeting the optimal Q-function, and rare-switching policy resets. We extend this to the setting of Hybrid RL, showing that initializing the critic with offline data yields sample efficiency gains compared to purely offline or online RL. Further, utilizing access to offline data, we provide a \textit{non-optimistic} provably efficient actor-critic algorithm that only additionally requires NoffcoffdH4/ϵ2N_{\text{off}} \geq c_{\text{off}}^*dH^4/\epsilon^2 in exchange for omitting optimism, where coffc_{\text{off}}^* is the single-policy concentrability coefficient and NoffN_{\text{off}} is the number of offline samples. This addresses another open problem in the literature. We further provide numerical experiments to support our theoretical findings.

Keywords

Cite

@article{arxiv.2505.03710,
  title  = {Actor-Critics Can Achieve Optimal Sample Efficiency},
  author = {Kevin Tan and Wei Fan and Yuting Wei},
  journal= {arXiv preprint arXiv:2505.03710},
  year   = {2025}
}

Comments

Accepted to ICML 2025

R2 v1 2026-06-28T23:23:18.172Z