Offline Minimax Soft-Q-learning Under Realizability and Partial Coverage

In offline reinforcement learning (RL) we have no opportunity to explore so we must make assumptions that the data is sufficient to guide picking a good policy, taking the form of assuming some coverage, realizability, Bellman completeness, and/or hard margin (gap). In this work we propose value-based algorithms for offline RL with PAC guarantees under just partial coverage, specifically, coverage of just a single comparator policy, and realizability of soft (entropy-regularized) Q-function of the single policy and a related function defined as a saddle point of certain minimax optimization problem. This offers refined and generally more lax conditions for offline RL. We further show an analogous result for vanilla Q-functions under a soft margin condition. To attain these guarantees, we leverage novel minimax learning algorithms to accurately estimate soft or vanilla Q-functions with $L^2$-convergence guarantees. Our algorithms' loss functions arise from casting the estimation problems as nonlinear convex optimization problems and Lagrangifying.

Paper

Similar papers

Peer review

Reviewer CJ1Z6/10 · confidence 4/52023-06-27

Summary

This paper presents algorithms based on soft Q-learning for offline RL with finite-sample guarantees. The guarantees are given for general function approximation hold under weaker requirements such as a (partially) weaker variant of single-policy concentrability and avoids the Bellman-completeness assumption.

Strengths

- The paper proposes algorithms based on soft Q-learning that only requires single-policy concentrability and realizability, without Bellman completeness. While a couple of recent papers such as Zhan et al. 2022 have also accomplished this, the guarantees presented in this work hold under a weaker single-policy coverage requirement and the MQP algorithm avoids the realizability of regularized optimal solutions, unlike the PRO-RL algorithm of Zhan et al. 2022. - The proposed algorithms appear to be implementable compared to prior works such as Chen and Jiang.

Weaknesses

- The dual variables $l(s,a)$ here seems to be closely connected to importance weights $w(s,a)$. The Bellman consistency term in (4) seems to be similar to the importance weighted regularization considered in Zhu et al. 2023, in which there is a max_w. The main difference is using $\Omega$ here. In that paper, the variable $w(s,a)$ are importance weights, which ensure that the average Bellman error is small under relevant distributions (including the data and optimal distributions). Would you please provide a comparison between the two objectives and comment on the connections between the variable $l(s,a)$ in your paper vs. $w(s,a)$ in their paper? Does the function $\Omega$ contribute to avoiding the Bellman completeness assumption? What is the key factor in your proposed objective or analysis that removes the need for Bellman completeness compared to Zhu et al? - Convergence rates are slower than 1/sqrt{N} (or involve the gap), which are achieved by some of the recent works such as Zhu et al. 2023. **References** Zhu, Hanlin, et al. "Importance Weighted Actor-Critic for Optimal Conservative Offline Reinforcement Learning." arXiv preprint arXiv:2301.12714 (2023)

Questions

Please refer to the questions in the above segment.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Yes

Reviewer 3Bh27/10 · confidence 4/52023-07-05

Summary

This paper proposes a q-learning based algorithm for solving offline minimax problem in reinforcement learning. Under certain partial coverage and realizability assumption for the regularized problem, they prove $O(1/\epsilon^4)$ sample complexity.

Strengths

This is the first paper for minimax problem in offline RL that only requires realizability assumption and partial data coverage (without completeness and all-policy concentrability).

Weaknesses

Weakness and comments: 1. The partial data coverage and realizability assumptions are made for the regularized problem instead of the original problem. It is unclear what the relationship between these assumptions with the standard partial coverage and realizability assumptions for the original problem as in ``Revisiting the linear-programming framework for offline rl with general function approximation''. 2. The technique seems to be similar to paper ``Refined value-based offline rl under realizability and partial coverage ''. The authors should explain more about the novelty. 3. References missing: ``Revisiting the linear-programming framework for offline rl with general function approximation'' also proposes an algorithm which requires $1/\epsilon^2$ sample complexity without completeness assumption for offline RL.

Questions

None

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

None

Reviewer 6n4K7/10 · confidence 2/52023-07-06

Summary

This paper proposed value-based algorithms for offline RL without Bellman completeness and full coverage of data support. By proposing the optimization objective of selecting the soft-Q from the set, where element in set satisfies the empirical soft-Bellman operator and the selected Q should minimize the square value under the expectation of behavior data distribution, and transforming it into a minimax problem, MSQP and MSQ could provide the property of convergence and sample complexity guarantee only with partial coverage and realizability.

Strengths

1. This paper propose novel offline algorithms without the assumption of full coverage and Bellman completeness. 2. The proposed algorithms are in a value-based manner, and avoid the assumption of initial distribution of ground truth MDP. 3. The proposed algorithms could have the property of convergence and sample complexity guarantee under more related assumptions. 4. This paper is well written and easy to follow.

Weaknesses

1. There lack some intuitions for the proposed learning objective in Equation 2. and Equation 3., since the square value of Q is not an explicit constraint for offline learning problem.

Questions

1. Could the author explain more about the intuitions of the proposed learning objective in Equation 2, though the convergence and sample complexity guarantee is provided?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

It could be better if empirical results beyond tabular are provided.

Reviewer 4xjy6/10 · confidence 4/52023-07-20

Summary

This submission studies offline RL with function approximation. Under the relatively mild assumption of partial coverage and realizability of the function approximations, the paper proposes the algorithm M(S)QP that approximates the (soft-)Q-function by solving a minimax optimization problem, and establishes guarantees of $L^2$-convergence of the Q-function estimator and the sub-optimality gap of the learned policy.

Strengths

A novelty of this work is that it studies offline RL based on soft-Q-function, while previous works mostly focus on the standard Q-function or the LP formulation. There are also several technical improvements compared to prior works, including $L^2$-convergence of the estimators and slightly milder technical assumptions.

Weaknesses

(1) The proposed schemes are based on minimax optimization and hence are mainly of theoretical interest, and the dual function $l$ is hard to interpret. (2) The bounds on PAC sample complexity are either worse than the results in related literatures, or require an extra margin condition.

Questions

(1) In Remark 1, the authors remark that the minimax optimization is computationally feasible when $\mathcal{L}$ is chosen to be RKHS or linear function classes. I would like to know if there is a concrete class of offline RL problems where such a choice of $\mathcal{L}$ satisfies realizability, or otherwise this remark is vacuous. (2) The bounds presented in this paper seem to be far from optimal (e.g. $O(n^{-1/4})$ convergence or $O(\epsilon^{-8})$ rate), and they are also worse than the results in previous works (e.g. the case $\alpha=0$ requires an extra assumption). I wonder if this is an artifact of the analysis or a price of milder assumptions. (3) Line 303: "In the following, we demonstrate that MQP, which is a special version of MSQP when $\alpha \rightarrow 0$, can achieve a faster rate of $O\left(1 / \epsilon^2\right)$." In my opinion, this remark is slightly misleading, because the faster $O\left(1 / \epsilon^2\right)$ rate is achieved under a stronger assumption of "soft margin". For example, in tabular offline RL with gap $\Delta>0$ and single policy concentrability $C^*$, the minimax optimal sample complexity is of order $\frac{SC^*}{\Delta\epsilon}$ (Wang et al., ignoring the horizon factor), which is better than the gap-independent sample complexity $\frac{SC^*}{\epsilon^2}$ as $\epsilon\to 0$. Xinqi Wang, Qiwen Cui, and Simon S. Du. "On Gap-dependent Bounds for Offline Reinforcement Learning." (4) Writing: (a) The notation $\left\|d\_{\pi\_\alpha^{\star}, P\_b} / P\_b\right\|\_{\infty}$ is ambiguous, because the measure $d\_{\pi\_\alpha^{\star}, P\_b}$ is defined on both the space $\mathcal{S}$ and the space $\mathcal{S}\times\mathcal{A}$, and so does the measure $P\_b$. (b) Line 187 & 188 (typos): $d\_{\pi\_{\alpha^*, P\_b}}$ -> $d\_{\pi\_\alpha^*, P\_b} $

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer oZk75/10 · confidence 4/52023-07-28

Summary

This paper studies offline reinforcement learning and proposes a soft Q-learning algorithm based on reformulating the solution to the Bellman equation as a saddle point of a certain minimax optimization problem. The authors then provide PAC guarantees for estimating the entropy-regularized Q-function, which in turn gives the sample complexity bounds. A key feature of this work is that it requires weaker assumptions (only partial coverage assumption is required) compared with existing literature.

Strengths

Theoretical analysis of Q-learning with function approximation is one of the key topics in reinforcement learning (RL). This paper provides a sample complexity bound of soft Q-learning under only a partial coverage assumption, which might be of interest to the broad community of RL.

Weaknesses

I have two major concerns about this work. (1) I am confused about the reformulation of solving the Bellman equation as a minimax optimization problem. It is well-known that both the vanilla Bellman operator and the entropy regularized Bellman operator are contraction mappings, which imply the uniqueness of both $q^*$ and $q_\alpha^*$. In that case, it is unclear what Eq. (2) means since there is really no optimization problem here. Am I missing anything? (2) I read the proof of Theorem 5 and have some questions. In the second inequality in line 637, the event $q^*(s,a')-q^*(s,\pi^*(s))\leq 0$ happens surely because the optimal policy is greedy with respect to the optimal Q-function. In that case, why do we need to add this event to the indicator? In addition, since $q^*(s,a')-q^*(s,\pi^*(s))\leq 0$ happens surely, the first term on the right-hand side of the first inequality in line 638 is exactly equal to $1$. As a result, the bound seems to be meaningless and Assumption 7 is redundant. What am I missing here?

Questions

See the weakness section.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

No negative societal impact.

Reviewer oZk72023-08-13

Acknowledgment of Rebuttal

Thank the authors for the detailed response. Q1: I am confused about the reformulation of solving the Bellman equation as a minimax optimization problem... I see the point now. In the next version, please make it clear the reason for introducing the optimization problem. Q2: I read the proof of Theorem 5 and have some questions... I checked the proof again and have the following comments: (1) In the rebuttal, the authors state that "The argument then proceeds in line 638 by splitting the range of $\hat{q}(s,a')-\hat{q}(s,\pi^*(s))$ in this indicator from $(-\infty, 0)$ into the mutually exclusive $(-\infty, -t]$ and $[-t,0]$..." I believe the events are split using $q^*(s,a')-q^*(s,\pi^*(s))$ instead of $\hat{q}(s,a')-\hat{q}(s,\pi^*(s))$. (2) How was the second inequality obtained from the first inequality in Line 641? (3) In general, I feel the appendix lacks quality, which is unfortunate for theoretical work. Besides the problem pointed out in my review, many steps in the proof are missing, which makes it hard for the reader to follow. Also, some equations are too long, which is a minor issue.

Authorsrebuttal2023-08-14

Thank you very much for your prompt response. **Q2** 1) Indeed, you are correct. We apologize for the typo. 2) Here is the detailed derivation for Line 641. We will expand it to make sure it's evidently clear. * In the second inequality of Line 641, we optimize the term in the second line in Line 641 with respect to $t$. * Set $\nu=2(E_{s \sim d, a\sim \pi_b^{\diamond}(s)} [|\hat q(s,a)-q^\star(s,a)|^2 ]+E_{s \sim d, a\sim \pi^\star(s)}[|\hat q(s,a)-q^\star(s,a)|^2 ])$ so that the second equation line on line 641 reads $|\mathcal A|(c(t/t_0)^\beta+2\nu/t^2)$. This holds for any $t$. * Set $t=t_0^{\beta/(2+\beta)}\nu^{1/(2+\beta)}$ so that $(t/t_0)^\beta=\nu/t^2$. * Obtain the bound $2c|\mathcal A|t_0^{-2\beta/(2+\beta}\nu^{\beta/(2+\beta)}$. * Note that $2^{\beta/(2+\beta)}\leq 2$ and $(x+y)^{\beta/(2+\beta)}\leq x^{\beta/(2+\beta)}+y^{\beta/(2+\beta)}$ for any $x,y>0$ since $\beta/(2+\beta)\in(0,1)$. * Using this we obtain the bound $4c|\mathcal A|t_0^{-2\beta/(2+\beta)}(E_{s \sim d, a\sim \pi_b^{\diamond}(s)} [|\hat q(s,a)-q^*(s,a)|^2 ]^{\beta/(2+\beta)}+E_{s \sim d, a\sim \pi^*(s)}[|\hat q(s,a)-q^*(s,a)|^2 ]^{\beta/(2+\beta)})$. * In the proof we ignore the constant factor of 4 since we do not care about universal constants -- we wrap that into the ``Poly" term in the expression in Theorem 5. But when revising we will make the 4 factor clear in the proof so it's easy to read and follow. 3) Thank you for the suggestion. We will go through the appendix and expand the arguments into more steps and add short in-words explanations to make it easier to follow the proofs.

Reviewer oZk72023-08-14

Thank the authors for the reply. Q2: The explanation for the inequality in Line 641 is clear, although many steps are skipped in the original proof. I do not have further questions and I have raised my score accordingly given that the authors have promised to revise the appendix.

Reviewer 4xjy2023-08-18

Thanks for the detailed response, which clarifies all my concerns.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC