Offline KL-Regularized Reinforcement Learning with Stochastic Transitions
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, state-action pairs and effective horizon . For data that cover the regularized optimum with concentrability , every estimator needs samples on some instances when is small, where 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 . 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 ; in the same temperature range, this power law alone gives the estimator a sample complexity polynomial in , with no coverage assumption. For the power can also increase the sample size needed by a factor polynomial in . 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.
Figures & tables
| Data | Lower bound | Upper bound (PSVI) |
|---|---|---|
| fixed design (D1) | (Theorem 7 ) | for (Theorem 5 ); for every (Theorem 4 ); for every after a burn-in on the support of (Proposition 35 ) |
| reference trajectories (D2) | for (Theorem 14 ); for , (Theorem 13 ) | as for (D1) without the burn-in result, on pairs with expected count at least of order ; the other pairs can cost a factor (Theorems 4 (b) and 5 (b)); with no coverage assumption, for (Corollary 15 ) |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Statement | Instances | Measured |
|---|---|---|
| Lemmas 1 , 3 | random MDPs and slacks | ratio of to at most ; path-KL identity to |
| Lemma 2 | 60000 draws, variance term active in 43%, | smallest partial derivative (negative in 3003 draws with the floor halved); no decrease along ; no Lipschitz excess beyond rounding; additive penalty: derivative sums up to where unclipped |
| Proposition 25 | random MDPs, 27 settings of 100 datasets each, | deviation at most of the penalty; at most of |
| Proposition 25 | self-loop instance, , , 27 settings of 100 datasets each | variance term active at both pairs of in every dataset; deviation at most of the penalty; at most of |
| Proposition 28 | Hoeffding penalty, trajectory streams, , 240 datasets | deviation at most of the penalty |
| Propositions 25 , 28 | Bernstein penalty, self-loop instance under (D2), , 240 datasets | deviation at most of the penalty; at most of |
KL-Regularized RL
Kullback–Leibler-regularized reinforcement learning uses divergence penalties to balance reward-seeking policy updates with staying near a reference policy.
See all 63 papers