Minimax Optimal and Computationally Efficient Algorithms for Distributionally Robust Offline Reinforcement Learning

Distributionally robust offline reinforcement learning (RL), which seeks robust policy training against environment perturbation by modeling dynamics uncertainty, calls for function approximations when facing large state-action spaces. However, the consideration of dynamics uncertainty introduces essential nonlinearity and computational burden, posing unique challenges for analyzing and practically employing function approximation. Focusing on a basic setting where the nominal model and perturbed models are linearly parameterized, we propose minimax optimal and computationally efficient algorithms realizing function approximation and initiate the study on instance-dependent suboptimality analysis in the context of robust offline RL. Our results uncover that function approximation in robust offline RL is essentially distinct from and probably harder than that in standard offline RL. Our algorithms and theoretical results crucially depend on a novel function approximation mechanism incorporating variance information, a new procedure of suboptimality and estimation uncertainty decomposition, a quantification of the robust value function shrinkage, and a meticulously designed family of hard instances, which might be of independent interest.

Paper

References (67)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer d55R6/10 · confidence 3/52024-06-17

Summary

This paper studies robust offline RL with function approximation under the specific setting of $d$-rectangular linear DRMDPs, where the nominal environment is a linear MDP with simplex feature space. The authors propose two learning algorithms and establish instance-dependent upper bounds for the suboptimality. The derived information-theoretic lower bound shows that the proposed VA-DRPVI algorithm is near-optimal up to a factor of $\widetilde{O}(\sqrt{d})$.

Strengths

(+) The paper derives an information-theoretic lower bound, which depends on the novel uncertainty function $\Phi((\Sigma_h^*)^{-1},s)$ and could be of independent interest to the community. (+) The suboptimality for the proposed VA-DRPVI algorithm nearly matches the lower bound.

Weaknesses

(-) The derived suboptimality bounds (in Theorems 3.4 and 4.2) and the lower bound (in Theorem 5.1) require the number $K$ of the trajectories in the offline dataset to scale with $\text{poly}(d,H)$, which could be quite restrictive in the offline setting.

Questions

1. Is it possible to establish suboptimality bounds and the lower bound for universal $K$? I guess there would be some "burn-in" terms. 2. Ma et al. (2023) consider a similar setting, except that the model uncertainty is measured by the KL divergence instead of the TV divergence in this paper. Can you point out the differences between this work and theirs? Also, are there any specific reasons for considering the TV divergence, and what are the unique challenges in the analysis? 3. The paper considers linear MDPs with simplex feature space. Can you explain the challenges in extending the results to general linear MDPs? 4. A minor suggestion on wording: In part 3 of the contribution section (the last paragraph on p.g. 2), the paper claims that "...which implies that VA-DRPVI is minimax optimal in the sense of information theory". However, as parameter $\beta_2$ scales with $\widetilde{O}(\sqrt{d})$, it can be better to switch to the word *near-optimal* instead of *minimax optimal*.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

Limitations have been addressed in the paper.

Reviewer nZV55/10 · confidence 4/52024-07-11

Summary

This paper proposes minimax optimal and computationally efficient algorithms with linear function approximation in the context of distributionally robust offline RL. The authors incorporate multiple new techniques in theoretical analysis, e.g., variance information, suboptimality and estimation uncertainty decomposition, robust value function shrinkage, and family of hard instances.

Strengths

1. The paper is technically sound with most claims supported sufficiently. 2. The theoretical analysis seems novel.

Weaknesses

Quality: 1. In Line 175, the expression requires calculating the minimum value of the entire space. This operation is computationally intractable when considering continuous state space, which violates the "computationally efficient" claim made by authors. 2. In Line 202, I don't find Assumption 4.1 and Remark 4.2 in the paper. Additionally, the "fail-state" assumption doesn't necessarily hold in multiple MuJoCo environments, i.e., some environments have negative minimum values. 3. There are no experimental results provided in the paper. Although it is a theoretical paper, it is better to include some toy experiments to verify the proposed algorithm. Clarity: 1. Since the authors only focus on linear function approximation, it is better to clarify it in the open questions (Lines 51-52). 2. (Lines 97-98) Assoud's method -> Assouad's method? Additionally, it is better to provide references about Assouad's method. 3. (Lines 201-203) It is better to provide references about the Nelder-Mead method. Significance: There are several works providing theoretical guarantees for general function approximation in online robust RL (e.g., [58]) and offline RL (e.g., [Chen and Jiang, 2019]). Some discussions are missing about why general function approximation is not considered here for offline robust RL settings. [Chen and Jiang, 2019] Information-Theoretic Considerations in Batch Reinforcement Learning, ICML 2019.

Questions

Please see the details in "weakness".

Rating

5

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

There is no potential negative social impact of this work.

Reviewer Qkxw6/10 · confidence 3/52024-07-13

Summary

This paper considers the distributionally robust Markov decision process (or sometimes robust MDP in the literature). In particular, it considers the linear RMDP with $d$-rectangularity TV uncertainty set, which decouples the uncertainty set from the state-action pair. Two algorithms are proposed to solve this, with the second algorithm being an improvement to the first. Extensive analysis is provided.

Strengths

1. The paper is well-written and easy-to-follow 2. Results are strong and thorough 3. Problem is well-motivated by introducing a hard instance (lower bound results)

Weaknesses

1. Comparison with other existing works is a bit unclear

Questions

1. Can authors provide more comparison with existing linear MDP with $d$-rectangularity? So besides the difference in the type of uncertainty set used, can you compare your work with [1] in details, especially the results and methodologies? 2. Continuing the first question, in particular, I noticed that your coverage assumption is stronger than [1]. It seems that you need the behavioral policy to cover states and actions that the optimal policy seldom visits as well. [1] states that with pessimism, they are able to achieve single-policy concentrability (w.r.t. only optimal policy). Could you share some insights? [1] Xiaoteng Ma, Zhipeng Liang, Li Xia, Jiheng Zhang, Jose Blanchet, Mingwen Liu, Qianchuan Zhao, and Zhengyuan Zhou. Distributionally robust offline reinforcement learning with linear function approximation. arXiv preprint arXiv:2209.06620, 2022. Minor things 1. Question marks on line 618

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

No other limitation I'd like to bring up

Reviewer mtm16/10 · confidence 5/52024-07-13

Summary

This paper presents a theoretical study of distributionally robust MDPs. Their findings show that function approximation is both different and harder in robust RL as compared to offline RL. They show matching information theoretic lower bounds for their novel algorithm.

Strengths

1. The paper has positioned itself and it's improvement over past work really well (Table 1). 2. The paper's result about the information theoretic lower bound has a lot of theoretical value. 3. The appendix is well written and the proofs are sound.

Weaknesses

1. More motivation on practical scenarios where DRMDPs are applicable will be useful. 2. Small simulation examples would further strengthen the paper to ensure that the constants in the sample complexities are reasonable.

Questions

There is also literature on theoretical multi-task RL, which while not focussed on the distributional robustness aspect, still tackle the problem of planning in an environment with uncertain dynamics. For example see : https://arxiv.org/pdf/2402.12570 and follow up works. I'd like the authors to cite and write the differences to Theorem 1 of the referenced paper, which use function approximation for the representation as well, and come up with uncertainty gaps for different state action pairs of the transition model, and still give a suboptimality gap on the policy planning in the downstream target task (see Theorem 2).

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

There are no significant technical limitations of this paper.

Reviewer nZV52024-08-12

Thank you for your clarification! I have increased my score, but I still have two more questions. 1. As I mentioned, the "fail-state" assumption doesn't necessarily hold in multiple MuJoCo environments, while the Nelder–Mead technique is a heuristic search method that cannot guarantee convergence to global optimal. In these cases, I think the suboptimality gap cannot hold anymore as shown in Theorem 4.2. 2. I admit that there are some challenges in the setting of d-rectangular linear DRMDPs. However, given existing online robust RL [58] and offline robust RL [26] have shown the theoretical guarantees under general function approximation, it is still not clear why authors don't consider general function approximation directly. In the authors' rebuttal, this part is somewhat dodged by highlighting the difficulty of d-rectangular linear DRMDPs.

Authorsrebuttal2024-08-12

Thank you for your reply. We answer your further questions as follows. --- ### Q1. About the "fail-state" assumption and Nelder–Mead technique. In theory, we would like to clarify that Theorem 4.2 does not rely on the “fail-state” assumption. The minimization over $s$ can be treated as an oracle, a common approach in theoretical RL analysis [1, 2, 3, 4]. Therefore, **our theoretical results are valid for any setting that satisfies the $d$-rectangular linear DRMDP structure**. Specifically, [1, 2, 3, 4] focus on standard linear MDPs and propose computationally efficient methods with the help of an oracle. For example, in the Bellman operator, the maximization over the action space $\mathcal{A}$, especially when continuous, is treated as an oracle. Similarly, in our work, assuming access to an oracle for minimization over the (potentially continuous) state space allows us to concentrate on the fundamental limits and challenges of $d$-rectangular linear DRMDPs with TV-divergence uncertainty sets. In practice, the minimization/maximization can be approximated. We acknowledge that in RL, a gap often exists between theoretical guarantees and actual performance in practice. However, our focus in this work is on providing a rigorous theoretical understanding of the offline $d$-rectangular DRMDP setting, which aligns with the approach of other theoretical research and meets NeurIPS standards. Furthermore, we plan to explore ways to address the issue of minimization over the state space in future work, which arises from the strong duality for TV distances (see Proposition G.1). As an alternative, we could consider replacing the TV-divergence uncertainty set with a Kullback-Leibler uncertainty set [6] or a $\chi^2$ uncertainty set [5], both leading to strong duality without the minimization term. We anticipate these approaches will lead to more practical algorithms for $d$-rectangular linear DRMDPs, though they are beyond the scope of this current paper. --- ### Q2. Why not consider general function approximation directly. We would like to further elaborate on this issue from the perspectives of uncertainty sets and algorithm design. **Panaganti et al. (2022) ([26] in our paper)**: [26] study offline $(s,a)$-rectangular DRMDPs with general function approximation. However, the $(s,a)$-rectangular uncertainty set can lead to overly conservative policies, particularly when the transition probabilities exhibit inherent structure [7,8]. To address this issue, our work focuses on structured uncertainty sets, specifically the $d$-rectangular linear DRMDP. This structured uncertainty set allows for a linear representation of robust value functions, resulting in less conservative policies compared to the general $(s,a)$-rectangular uncertainty set. This is why we do not follow [26] in studying general function approximation. In fact, our study of function approximation under structured uncertainty sets represents a significant contribution of our work, distinguishing it from [26]. **Zhou et al. (2023) ([58] in our paper)**: [58] explores the Double-Sampling uncertainty set and the Integral Probability Metric uncertainty set. They focus on scenarios with a simulator, proposing a robust Natural Actor-Critic algorithm with finite-time convergence guarantees, supported by extensive numerical experiments to demonstrate robustness. However, their method cannot be easily extended to address the $d$-rectangular uncertainty set of linear DRMDPs due to fundamental differences in uncertainty set formulations and algorithm design (value iteration-based versus policy gradient-based). We acknowledge that the method in [58] and our work are orthogonal. This distinction highlights potential avenues for complementary research in the future. We will add the above discussion in our revision. If you have further questions, we are more than happy to discuss with you. --- **References** [1] Jin et al. Provably efficient reinforcement learning with linear function approximation. In COLT 2020. [2] He et al. "Nearly minimax optimal reinforcement learning for linear markov decision processes." In ICML 2023. [3] Xiong et al. Nearly Minimax Optimal Offline Reinforcement Learning with Linear Function Approximation: Single-Agent MDP and Markov Game. In ICLR 2023. [4] Yin et al. "Near-optimal Offline Reinforcement Learning with Linear Representation: Leveraging Variance Information with Pessimism." In ICLR 2022. [5] Shi et al. The curious price of distributional robustness in reinforcement learning with a generative model. arXiv preprint, 2023. [6] Zhou et al. Finite-sample regret bound for distributionally robust offline tabular reinforcement learning. In AISTATS 2021. [7] Goyal et al. Robust markov decision processes: Beyond rectangularity. Mathematics of Operations Research, 48(1):203–226, 2023. [8] Ma et al. Distributionally robust offline reinforcement learning with linear function approximation. arXiv preprint, 2022.

Reviewer d55R2024-08-12

Thank you

Thank you for your detailed responses. I tend to maintain my score. Responding: After reading references [1] and [2], I can see that the lower bound on $K$ results from the concentration properties of variance estimators. Hence, it seems quite confusing that a suboptimality upper bound for uniform $K$ cannot be derived for Algorithm 1 as it does not involve such estimators.

Authorsrebuttal2024-08-12

Thank you for your reply and for acknowledging our responses. We answer your further questions as follows. --- ### Suboptimality upper bound for uniform $K$. We would like to acknowledge that your intuition is correct. For Algorithm 1, if we adopt the Suboptimality Decomposition Lemma C.1 proposed in our current paper, as well as the uniform concentration argument for analyzing the estimation error developed by Jin et al. (2021), we can get an instance-dependent suboptimality upper bound for uniform $K$ for Algorithm 1 without Assumption 3.3. However, this would lead to an additional $O(\sqrt{d})$ in the order of the upper bound of Algorithm 1. More importantly, the focus of our work is to explore the fundamental limit and intrinsic characteristics of the offline d-rectangular linear DRMDP. We proposed Algorithm 1 to use it as the backbone of the more advanced Algorithm 2, which could achieve the nearly minimax optimal suboptimality bound. Therefore, we analyze Algorithm 1 using the method presented in our paper so that it serves as a warm-start for us to work out the theoretical analysis pipeline needed to achieve our goal (near-optimal instance-dependent upper and lower bounds). Further, we would like to clarify that it is unsure if we can derive an upper bound for uniform $K$ for **Algorithm 2** due to the reason we state in the sequel. Note that Jin et al. (2021)’s analysis framework does not incorporate the reference-decomposition technique, and thus cannot be leveraged to achieve the near-optimal suboptimality in the $d$-rectangular linear DRMDP setting. Consequently, adapting Jin et al. (2021)’s analysis framework mentioned in the above paragraph for analyzing Algorithm 1 would be a distraction/detour from our main goal. Nevertheless, we will add a remark to discuss more on this in the next version of our manuscript. We would be more than happy to discuss more if you have further questions. **References** [1] Ying Jin, Zhuoran Yang, and Zhaoran Wang. Is pessimism provably efficient for offline RL? In International Conference on Machine Learning, pages 5084–5096. PMLR, 2021.

Reviewer d55R2024-08-13

Thank you

I appreciate the detailed explanation from the authors. I have no further questions.

Reviewer mtm12024-08-12

I thank the authors for their detailed response. I request to maintain my positive rating.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC