arXiv:2611.17603v1 Announce Type: new Abstract: KL-regularized reinforcement learning is the objective behind RL fine-tuning of language models, and its data are often collected by the reference policy itself. However, beyond bandits and deterministic transitions, little is known about how much offline data it needs. In this work, we study finite discounted Markov decision processes with unknown stochastic transitions, $K$ state-action pairs and effective horizon $H$. For data that cover the regularized optimum with concentrability $C\geq 4$, every estimator needs $\Omega(KCH^3/(\epsilon\max{\epsilon,\beta}))$ samples on some instances when $\epsilon$ is small, where $\beta$ is the temperature. Under a fixed design, pessimistic soft value iteration with a variance-aware penalty attains this rate with high probability, up to logarithmic factors, when $\beta\geq 1.4H$. When the data come from the reference policy itself, we show that it covers the optimum up to a power law whose exponent grows as $H/\beta$; in the same temperature range, this power law alone gives the estimator a sample complexity polynomial in $1/\epsilon$, with no coverage assumption. For $\beta\geq H$ the power can also increase the sample size needed by a factor polynomial in $1/\epsilon$. Finally, the plug-in estimator, which is minimax optimal with a generative model, can fail on reference data at small temperatures when many actions are rarely seen, while the pessimistic estimator does not.
Offline KL-Regularized Reinforcement Learning with Stochastic Transitions
A new arXiv paper (2611.17603v1) establishes sample-complexity bounds for offline KL-regularized reinforcement learning in finite discounted Markov decision processes with unknown stochastic transitions, K state-action pairs and effective horizon H. The authors prove a lower bound of Ω(KCH³/(ε max{ε,β})) samples for any estimator when data cover the regularized optimum with concentrability C ≥ 4 and ε is small, where β is the temperature, and show that pessimistic soft value iteration with a variance-aware penalty matches this rate up to logarithmic factors under a fixed design when β ≥ 1.4H. When data come from the reference policy itself, the paper shows that coverage of the optimum follows a power law whose exponent grows as H/β, giving polynomial-in-1/ε sample complexity without a coverage assumption in the same temperature range, while the plug-in estimator that is minimax optimal with a generative model can fail on reference data at small temperatures when many actions are rarely seen.
Run your AI side-project on zahid.host
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.