Provably Efficient Offline Goal-Conditioned Reinforcement Learning with General Function Approximation and Single-Policy Concentrability

Goal-conditioned reinforcement learning (GCRL) refers to learning general-purpose skills that aim to reach diverse goals. In particular, offline GCRL only requires purely pre-collected datasets to perform training tasks without additional interactions with the environment. Although offline GCRL has become increasingly prevalent and many previous works have demonstrated its empirical success, the theoretical understanding of efficient offline GCRL algorithms is not well established, especially when the state space is huge and the offline dataset only covers the policy we aim to learn. In this paper, we provide a rigorous theoretical analysis of an existing empirically successful offline GCRL algorithm. We prove that under slight modification, this algorithm enjoys an $\widetilde{O}(\text{poly}(1/ε))$ sample complexity (where $ε$ is the desired suboptimality of the learned policy) with general function approximation thanks to the property of (semi-)strong convexity of the objective functions. We only require nearly minimal assumptions on the dataset (single-policy concentrability) and the function class (realizability). Moreover, this algorithm consists of two uninterleaved optimization steps, which we refer to as $V$-learning and policy learning, and is computationally stable since it does not involve minimax optimization. We also empirically validate our theory by showing that the modified algorithm outperforms the previous algorithm in various real-world environments. To the best of our knowledge, this is the first algorithm that is both provably efficient with general function approximation and single-policy concentrability, and empirically successful without requiring solving minimax optimization problems.

Paper

Similar papers

Peer review

Reviewer vRJf6/10 · confidence 4/52023-07-04

Summary

The paper proposes the VP-learning algorithm to solve offline goal-conditioned RL in the context of general function approximations. The algorithm is based on a previous empirically successful algorithm proposed by [1], and the author proves the finite sample complexity for VP-learning under mild assumptions. The proposed algorithm avoids minimax learning by using a duality formulation first identified by [1], which makes the proposed VP-learning algorithm friendly to implementation while also enjoying theoretical guarantees. **References:** [1] Ma J Y, Yan J, Jayaraman D, et al. Offline goal-conditioned reinforcement learning via $ f $-advantage regression[J]. Advances in Neural Information Processing Systems, 2022, 35: 310-323.

Strengths

- The VP-learning algorithm proposed in the paper enjoys finite sample guarantees under a mild single policy concentrability assumption. Compared with previous works on offline RL (single-task) algorithms with general function approximations, VP-learning does not involve minimax optimization, potentially making the algorithm more easy and stable to implement.

Weaknesses

- This theoretical paper aims to study the problem of offline goal-conditioned RL. But it seems that the whole theory does not essentially rely on the setting of goal-conditioned RL. The existing theoretical works this paper compares with are also for standard offline RL, which makes the motivation of the paper confusing. It's true that the standard offline RL can be interpreted as a special case of goal conditioned RL, but I hope to see more intuitions and messages on the problem of goal-conditioned RL itself. - The key path through which the VP-learning algorithm can avoid solving minimax optimization problems is based on [1], which makes the theoretical contributions and insights of the paper weakened.

Questions

- Regarding Assumption 4 (single policy realizability), what is the $\alpha$ in the equation $\pi_{\alpha}^{\star}\in\Pi$? Since in Theorem 2 and 3 this assumption is imposed before the choice of $\alpha$, the $\alpha$ in the assumption seems undetermined. I think this needs clarification since the paper assumes that the policy class $\Pi$ is finite, which necessitates a specific choice of $\pi_{\alpha}^{\star}$. - Regarding Assumption 5 (lower bound of policy), this seems a quite strong assumption which is not needed by previous single task offline RL algorithms with general function approximations, e.g., [2, 3, 4]. The author argues that the parameter $\tau$ can be very small, but a small $\tau$ will also increase the upper bound of the suboptimality of VP-learning due to Theorem 2&3. Is this assumption actually necessary for offline goal-conditioned RL? or is it only required by the specific analysis for VP-learning algorithm? **References:** [2] Zhan W, Huang B, Huang A, et al. Offline reinforcement learning with realizability and single-policy concentrability[C]//Conference on Learning Theory. PMLR, 2022: 2730-2775. [3] Uehara M, Sun W. Pessimistic model-based offline reinforcement learning under partial coverage[J]. arXiv preprint arXiv:2107.06226, 2021. [4] Xie T, Cheng C A, Jiang N, et al. Bellman-consistent pessimism for offline reinforcement learning[J]. Advances in neural information processing systems, 2021, 34: 6683-6694.

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

The above questions are potential limitations.

Reviewer quHd6/10 · confidence 4/52023-07-05

Summary

This paper provides a rigorous theoretical analysis to a modified version of existing offline Goal-conditioned RL algorithm, and proves that it has an $\tilde O (poly(1/\epsilon))$ sample complexity, where $\epsilon$ is the desired suboptimality of the learned policy. The algorithm requires minimal assumptions on the dataset and the function class and does not involve minimax optimization.

Strengths

The paper makes a novel contribution by providing a theoretical analysis of an existing offline GCRL algorithm. The paper is clearly written and well-organized. The theoretical analysis is well-explained.

Weaknesses

The derivation of this paper is flawed due to a mistake in Proposition 3.2. This mistake leads to the conclusion that strong duality holds, and we can recover the optimal policy by solving the dual problem. However, it can be seen that, at the limit of $\alpha\to0$, the dual problem (Equation (6)) solves a value function that evaluates the behavior policy. Therefore, solving the dual problem cannot recover the optimal value from suboptimal data, and strong duality does not hold. As a result, all subsequent derivation is invalid, and the claims are unsupported. The mistake in Proposition 3.2 stems from a citation of Proposition 4.2 of [1] which incorrectly establishes strong duality. Overall, I think the paper would have been a valuable contribution to the field of offline RL. However, the flawed derivation is a serious issue that needs to be addressed. [1] Ma, Jason Yecheng, et al. "Offline goal-conditioned reinforcement learning via $f$-advantage regression." Advances in Neural Information Processing Systems 35 (2022): 310-323. ===Post-rebuttal Update=== Previously raised concerns have been effectively addressed. Overall, I am optimistic that, with some revisions, this paper could meet the standards for acceptance. I am pleased to raise my score from 3 to 6 and look forward to seeing your further refinement of this paper.

Questions

Can you explain more on the raised issue?

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

The authors of the paper have discussed the limitations of their work.

Reviewer 39fV7/10 · confidence 1/52023-07-06

Summary

This paper aims at improving the theoretical understanding of offline goal-conditioned RL (GCRL). In particular, this paper modifies an existing offline GCRL algorithm and shows an O^˜ (poly(1/ϵ)) sample complexity under minimal assumptions of single-policy concentrability and realizability. Their algorithm, called VP-learning, consists of two uninterleaved optimization steps and has good empirical performance while retaining computational stability. Moreover, it can also be applied in non-goal-conditioned settings. There seems to be a theory-practice gap that this paper addresses. Namely, most provably efficient algorithms seems to require minimax optimization, while in practice that is not effective. Ideally, an algorithm should be practical with good sample complexity guarantees, which is a gap this paper aims to fill. In particular this paper provides guarantees for a modified version of an algorithm GoFAR, with modifications in the deterministic and stochastic MDP settings. They call their modified algorithm VP-learning. While other algorithms have been shown to be efficient under single-policy concentrability and realizability assumptions, they require solving minimax optimization problems.

Strengths

While other algorithms have been shown to be efficient under single-policy concentrability and realizability assumptions, they require solving minimax optimization problems. This algorithm does not. The algorithm they develop is built off of an algorithm with good empirical performance.

Weaknesses

While part of the pitch of the paper is that they desire an algorithm that has provable efficiency whilst also having good empirical performance, they do not have empirical results to demonstrate that their modified algorithm performs well.

Questions

How do you expect the modified algorithm, VP-learning to compare to GoFAR in terms of empirical performance? How do you expect the modifications will impact things adversely?

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

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

To my knowledge, the authors do adequately address the limitations of the paper.

Reviewer fB346/10 · confidence 2/52023-07-06

Summary

This paper establishes a rigorous theoretical analysis for offline goal-conditioned reinforcement learning algorithms (GCRL). To achieve that, the authors made a slight modification on top of an existing offline GCRL algorithm (GoFAR), achieve a polynomial sample complexity by regression instead minimax optimization.

Strengths

1. This paper is well-organized and provides a throughout survey of related work. 2. It seems to achieve an effective and reasonable improvement from the lens of theory.

Weaknesses

There is no experiment to support the correctness of the theoretical analysis.

Questions

N/A

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

3 good

Limitations

N/A

Reviewer quHd2023-08-13

I am pleased to raise my score from 3 to 6

I appreciate the effort you've put into addressing the concerns raised in my review. Upon considering your response, I am pleased to acknowledge that the primary concern I had previously expressed has been effectively addressed. However, the proposed fix raises an additional problem. The V-learning now involves a minmax optimization which eliminates a bright spot of the VP-learning algorithm. Furthermore, I am aligned with the viewpoints of Reviewer vRJf concerning the adoption of the GCRL setting. It appears that the inclusion of the GCRL setting introduces complexity to the notation without yielding any benefits. Contrary to the assertion in your rebuttal, I think this setting does not enhance the generality of the approach. The goal can be easily formulated as the additional state dimensions, making GCRL special cases of RL. Therefore, I recommend move to the RL setting which could also amplify the potential impact and reception of this paper. Overall, I am optimistic that, with some revisions, this paper could meet the standards for acceptance. I am pleased to raise my score from 3 to 6 and look forward to seeing your further refinement of this paper.

Authorsrebuttal2023-08-13

Thanks for your valuable feedback!

We thank the reviewer again for the helpful comments that make our results more solid, and we are pleased that we have addressed the concern! For GCRL vs single-task RL setting, it's a very good suggestion to view GCRL as a special case of RL and simplify the notation to avoid introducing complexity to readers. We will make corresponding modifications in the revision. For the additional problem that the reviewer mentioned regarding our proposed fix, actually, the $V$-learning is still a minimization problem. Note that our $V$-learning still has the form of $\min\_{V} L\_\alpha(V)$, and the objective $L\_\alpha(V) = \alpha( (1-\gamma)\mathbb{E}\_{(s,g)\sim(\rho,p(g))}[V(s;g)] + \mathbb{E}\_{(s,a,g)\sim \mu}[ I$ { $g'\_\*(A\_V(s,a;g)) \geq 0$ } $ \tilde{g}\_*(A\_V(s,a;g)) ])$ in our proposed fix does not involve a maximization problem. Therefore, after the fix, our VP-learning algorithm still enjoys the property that it does not involve a minimax optimization.

Reviewer vRJf2023-08-13

Response to Authors

I have read the rebuttal and the comments from other reviewers. Thanks for all your efforts dealing with my concerns. I agree now the theoretical contribution of the paper is enough, especially given that the technical concerns from Reviewer quHd can be addressed. Regarding my other questions, I hope that the authors could make them clearer in the revision. Given all that, I am pleased to raise my score to 6.

Authorsrebuttal2023-08-13

Thanks for your time reviewing our paper, reading our response, and providing helpful feedback! We will make the points you mentioned more clear in the revision.

Reviewer fB342023-08-21

Thanks for your response

I thank the authors for their response and the uploaded experiment results, which improve the credibility of this work. I do not doubt the correctness of their contribution to the theory, so I gave such a suggestion as I think it would be helpful for the readers to have a straightforward understanding of their work. I'll improve my score from 5 to 6. However, I still have some suggestions about their uploaded experiment results: I cannot get any information about the environment settings, e.g., what are FetchReach, FetchPick, FetchPush, FetchSlide, and HandReach? I hope the authors can add this part of the environment introduction and also the learning curve in their future revision.

Authorsrebuttal2023-08-21

Thanks for your suggestion

Thanks for your time reviewing our paper and reading our response, and thanks for raising your score! For the experiment settings, FetchReach, FetchPick, FetchPush, FetchSlide, and HandReach are all environments in the d4rl benchmark. We did not provide details of the environment in the global response since it is the same as [1], and thus we omit the details to keep the response more concise. We will add the environmental details in the revision. Thanks for your suggestion! **Reference:** [1] Ma J Y, Yan J, Jayaraman D, et al. Offline goal-conditioned reinforcement learning via f-advantage regression[J]. Advances in Neural Information Processing Systems, 2022, 35: 310-323.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC