Corruption-Robust Offline Reinforcement Learning with General Function Approximation

We investigate the problem of corruption robustness in offline reinforcement learning (RL) with general function approximation, where an adversary can corrupt each sample in the offline dataset, and the corruption level $ζ\geq0$ quantifies the cumulative corruption amount over $n$ episodes and $H$ steps. Our goal is to find a policy that is robust to such corruption and minimizes the suboptimality gap with respect to the optimal policy for the uncorrupted Markov decision processes (MDPs). Drawing inspiration from the uncertainty-weighting technique from the robust online RL setting \citep{he2022nearly,ye2022corruptionrobust}, we design a new uncertainty weight iteration procedure to efficiently compute on batched samples and propose a corruption-robust algorithm for offline RL. Notably, under the assumption of single policy coverage and the knowledge of $ζ$, our proposed algorithm achieves a suboptimality bound that is worsened by an additive factor of $\mathcal{O}(ζ(C(\widehat{\mathcal{F}},μ)n)^{-1})$ due to the corruption. Here $\widehat{\mathcal{F}}$ is the confidence set, and the dataset $\mathcal{Z}_n^H$, and $C(\widehat{\mathcal{F}},μ)$ is a coefficient that depends on $\widehat{\mathcal{F}}$ and the underlying data distribution $μ$. When specialized to linear MDPs, the corruption-dependent error term reduces to $\mathcal{O}(ζd n^{-1})$ with $d$ being the dimension of the feature map, which matches the existing lower bound for corrupted linear MDPs. This suggests that our analysis is tight in terms of the corruption-dependent term.

Paper

Similar papers

Peer review

Reviewer wcW26/10 · confidence 3/52023-06-22

Summary

The paper studies offline RL with general function approximation in the presence of adversarial corruption. Specifically, an adversary can corrupt the rewards and transitions in the offline dataset with a fixed cumulative budget. The corruption formulation encompasses prior corruption formulations. The proposed algorithm is based on an uncertainty-weight iteration subroutine, inspired by similar techniques in prior robust online RL papers. The authors conduct a finite-sample analysis, showing a tighter result on the corruption term. The theory is accompanied by a practical implementation tested over a couple of D4RL tasks.

Strengths

- The authors propose a corruption formulation, which admits prior formulations such as fixed fraction contamination of Zhang et al. as special cases. - The theoretical result improves over the most relevant prior work Zhang et al. from $\sqrt{\epsilon}$ dependency to $\epsilon$ dependency of the corruption term, which matches the lower bound when specialized to the linear case. - Inspired by the uncertainty weighting approach in robust online RL, the authors propose a similar weight iteration method in the offline RL case. However, in the offline case, the - The authors implement their algorithm over a subset of continuous control tasks in D4RL.

Weaknesses

- The main limitation of this work is that the proposed algorithm requires the corruption level to be known. I did not find any attempts on possible heuristics for estimating the corruption level, which limits the practicality of the proposed approach. - The experiments section is generally weak. The practical implementation is limted to only two continous control tasks (Walker and Half Cheetah) from D4RL that are artificially corrupted. It would be good to see and compare performance with other tasks in D4RL as well. Furthermore, the approach is compared with uncertainty-based offline RL algorithms that are not designed to be corruption robust, making the comparison unfair.

Questions

Suggestions: - Perhaps comparison with an implementation of Zhang et al instead of uncertainty-based offline RL is more appropriate.

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some 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

Yes, clearly stated.

Reviewer GirP6/10 · confidence 3/52023-06-26

Summary

The research investigates the issue of corruption robustness in offline reinforcement learning with general function approximation, where an adversary can corrupt samples in the offline dataset. The objective is to find a policy that remains robust against such corruption and minimizes the suboptimality gap compared to the optimal policy in uncorrupted Markov decision processes. To address this, a new uncertainty weight iteration procedure inspired by the robust online RL setting is designed, enabling efficient computation on batched samples. The proposed corruption-robust algorithm achieves a suboptimality bound, considering single policy coverage and knowledge of the corruption level. The analysis reveals that the corruption-dependent error term matches the existing lower bound for corrupted linear MDPs, indicating the tightness of the analysis in terms of corruption dependence.

Strengths

I think this work really pushes the data-corruption solving real-world RL community research efforts further by answering: > Can we design a generic algorithm for data-corrupted offline RL in the context of function approximation? The main contribution: algorithmic introduction to uncertainty reweighting idea. Both theoretical and practical performances have been shown. This is good for publication at NeurIPS.

Weaknesses

I have only one minor weakness for this work as follows: On Assumption 2.1. The recent offline RL theoretical guarantees proposing pessimism is to avoid completeness assumption and having weaker concentrabilities, like that of [23]. I don't see what is pessimism helping in this work if not to avoid the completeness. Maybe I am missing something. I am okay with this assumption if the work is trying to address the corruption-robust problem which requires (?) completeness?

Questions

please see weaknesses for more.

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some 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

Limitations is marked Yes, but I couldn't find the discussion in the main paper.

Reviewer p9Rd7/10 · confidence 4/52023-07-05

Summary

This paper proposed a new robust offline RL algorithm called the Corruption Robust PEVI (CR-PEVI). The CR-PEVI achieves smaller suboptimality error compared to existing methods. The key idea behind CR-PEVI is to utilize the uncertainty-weighting technique when solving the least-squares problem in PEVI. The weights in the proposed method do not have closed-form solution, thus the authors proposed an iterative approach and relies on the monotone convergence theorem to prove convergence to the target weights. Based on that and several other assumptions, e.g., dataset needs to be well-explored, the paper proved an upper bound on the suboptimality of the policy learned by the CR-PEVI algorithm. When specializing to the linear function approximation with dimension d, the proposed CR-PEVI achieves O(eta*d/n) convergence rate. Experiments demonstrate the effectiveness of the CR-PEVI method.

Strengths

(1). This paper studied robust offline RL, which is a hot and timely topic in recent years. Prior works mostly consider test-time attacks that are relatively easier to defend. However, less is understood about robustness against training-time attacks. This paper pushes the frontier in this domain and gained deeper understanding on how to design poisoning-resilient offline RL algorithms in the function approximation scenario, which is important and fascinating work. (2). The theoretical analysis is intriguing. The authors combined the uncertainty-weighting technique and the PEVI algorithm to propose a new algorithm that is robust to offline poisoning attacks. The analysis is based on several critical theoretical results established in the paper, which are complete and solid. Overall, the paper has strong theory contributions. (3). The empirical results are also convincing. On all the tasks, the CR-PEVI achieves much better performance compared to baselines.

Weaknesses

(1). It seems like the definition of cumulative corruption is not commonly used by prior works. It is defined as the gap between the Bellman operators induced by the clean dataset and corrupted dataset, which is less intuitive and harder to quantize. In contrast, it makes more sense to define corruption level as the real modification on the reward/state transition over the trajectories. (2). In the experiment, the setup does not strictly conform with attacks under limited budget. The authors implemented different attacks without reporting the corruption level for every different attack and victim algorithm. Furthermore, there is not parameter study on the corruption level eta as well. This makes me wonder if the comparison to prior works, in particular, Zhang et al, is fair enough. (3) Missing a few related references on training-time attacks in RL: [1]. Policy Poisoning in Batch Reinforcement Learning and Control [2]. Adaptive Reward-Poisoning Attacks against Reinforcement Learning

Questions

(1). Why the paper chooses to define the cumulative corruption based on the Bellman operator instead of the modification on the dataset? Is the goal to favor theoretical analysis? How does the theory changes if the corruption level is defined differently, e.g., total change on the rewards? (2). It's important to report the corruption level induced by different attack and victim RL algorithms in the experiment section. Furthermore, I was wondering how the cumulative corruption level affects the policy learned by CR-PEVI in practical scenarios.

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

Yes

Reviewer H58a6/10 · confidence 2/52023-07-07

Summary

This work considers corrupted offline RL with general function approximation, and propose an algorithm based on PEVI and the uncertainty weighting technique. Under certain date coverage and well-explored dataset assumptions, the proposed algorithm achieves neal-optimal bound, where the corruption-independent term matches with the uncorrupted bound while the corruption-dependent term nearly matches the lower bound for linear models. Futher, experiments showcase the efficetiveness of the algorithm.

Strengths

1 This work is well-organized and well-written. 2 The studied corrupted offline RL with general function approximation is not being widely explored. 3 Adapting uncertainty-related weights technique to offline setting may be of interest.

Weaknesses

1 The uncertainty-weighting technique and the PEVI are not new. 2 This work extends the corrupted online setting to corrupted offline setting

Questions

1 Is the result comparable to the existing corrupted online RL with general function approximation? 2 Could the authors comment more on Eqn. (6) and give intuition? It is unclear what tuning parameter $\alpha$ and $\lambda$ are designed for in this uncertainty-related weights. 3 For the lower bound shown in Theorem 2, it would be great if the authors provide any insight. In line 12, ''$\widehat{F}$'' -> (or <-) ''$\hat{F}$''.

Rating

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

3 good

Contribution

2 fair

Limitations

NA

Authorsrebuttal2023-08-15

Thank you for raising the score. We're glad to know that our response has addressed your concerns. We will incorporate these clarifications into the final version.

Reviewer GirP2023-08-20

The rebuttal addressed my concerns. I have adjusted my rating considering the rebuttal and other reviewers’ concerns.

Authorsrebuttal2023-08-21

We're grateful for your positive feedback. We are pleased to learn that our response has effectively addressed the concerns raised. We've noticed that the current ratings are still unchanged. We kindly remind the reviewer to adjust the rating if the reviewer forgot.

Reviewer p9Rd2023-08-20

Thank you for your response. My questions are carefully addressed in the rebuttal. I have increased my score accordingly. There are a few related works about poisoning attacks in RL, and the authors should consider discussing them in the paper. Ma et al. Policy poisoning in batch reinforcement learning and control (NeurIPS 19). Zhang et al. Adaptive Reward-Poisoning Attacks against Reinforcement Learning (ICML 20).

Authorsrebuttal2023-08-21

Thank you for increasing the score! We will be sure to discuss the additional related works pointed out by the reviewer in the final version.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC