Bi-Level Offline Policy Optimization with Limited Exploration

We study offline reinforcement learning (RL) which seeks to learn a good policy based on a fixed, pre-collected dataset. A fundamental challenge behind this task is the distributional shift due to the dataset lacking sufficient exploration, especially under function approximation. To tackle this issue, we propose a bi-level structured policy optimization algorithm that models a hierarchical interaction between the policy (upper-level) and the value function (lower-level). The lower level focuses on constructing a confidence set of value estimates that maintain sufficiently small weighted average Bellman errors, while controlling uncertainty arising from distribution mismatch. Subsequently, at the upper level, the policy aims to maximize a conservative value estimate from the confidence set formed at the lower level. This novel formulation preserves the maximum flexibility of the implicitly induced exploratory data distribution, enabling the power of model extrapolation. In practice, it can be solved through a computationally efficient, penalized adversarial estimation procedure. Our theoretical regret guarantees do not rely on any data-coverage and completeness-type assumptions, only requiring realizability. These guarantees also demonstrate that the learned policy represents the "best effort" among all policies, as no other policies can outperform it. We evaluate our model using a blend of synthetic, benchmark, and real-world datasets for offline RL, showing that it performs competitively with state-of-the-art methods.

Paper

References (60)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer q6tp7/10 · confidence 3/52023-07-05

Summary

This manuscript proposes a bi-level structure for solving offline RL problem. The authors design the lower level constructing a set of value estimates, while the upper level maximizing the conservative value estimate. Specifically, the author provides a penalized adversarial estimation algorithm. The superiority of the proposed method is shown from both real world and synthetic data.

Strengths

1. The hierarchical structure for solving offline RL problem is very interesting. 2. The proposed method considers a confidence set of value estimates as well as the uncertainty, and it effectively prevents from being too pessimistic. 3. The manuscript provides solid theoretical results, and it shows the proposed method could work without requiring data-coverage, which sounds quite useful in offline RL scenario.

Weaknesses

1. It is suggested to add some important steps in Algorithm 1 from Line 270 ~ Line 281, otherwise the Algorithm 1 is too simple while there are many details in 270 ~ Line 281, which is somehow difficult to read. 2. The authors considers the synthetic data (CarPole) and real-world data (Type I Diabetes), as a common benchmark in offline RL research, the D4RL tasks are widely studied in offline RL experiments. It is suggested to consider some D4RL task in the experiment.

Questions

1. There are some confusion in understanding the definition of Consistency above line 154, should the $\mathcal{Q}_{\epsilon_n}$ be related with policy $\pi$? 2. In line 3 of Algorithm 1, it is required to solve a maxmin problem with iterative method, so how is the time consumption of the proposed method? Could the author provides some experiment study or conceptual analysis on this issue? 3. Any practical suggestions on selecting $\mathbb{G}(x)$ and the initial function of $\tau(s,a)$? The author also mentions the proposed method does not require data-coverage, could this issue be demonstrated from experiments? Say using a biased policy as behavior policy to generate data to learn offline RL policies.

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

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

see questions and limitations

Reviewer C4A27/10 · confidence 3/52023-07-05

Summary

Offline Reinforcement Learning suffers from distribution shift, arising as a consequence of limited exploration and function approximation. The paper aims to address these two challenges by proposing a bi-level hierarchical framework for policy optimization. The upper-level optimizes a policy improvement objective using a pessimistic value function. The lower level learns a conservative estimate of the value function by maintaining a confidence set for the upper level. Development of the framework is carried out by first constructing robust intervals for addressing approximation bias and then using these intervals in uncertainty quantification. Motivated by its theoretical results, the work further proposes an empirical penalized adversarial estimation algorithm which incorporates the bi-level objective using an adversarial estimation loss. While the lower level is optimized efficiently using maximum mean discrepancy principle, the upper level utilizes function approximation for learning the proximal mapping.

Strengths

* The paper is well written and organized. Specifically, the paper builds the framework in a structured way. * While the paper borrows from past results in tackling distribution shift and constrained optimization, the bi-level approach and its abstraction between policy and value objectives is novel.

Weaknesses

* **Empirical Evaluation:** My main concerns are centered at the empirical evaluation and experiments carried out to assess the bi-level framework. Experiments do not validate the theoretical claims of the work and only evaluate the performance of the framework. Ideally the empirical evaluation should validate whether regret bounds of Theorem 4.1, Theorem 5.1 or Theorem 5.2 would hold or not. Authors can compare empirically whether the bound of Theorem 5.2 is tight and that if $\lambda$ and $c^{*}$ are chosen as per the set criteria, regret is close to the bound of Theorem 5.2. Similarly, experiments do not demonstrate the benefits of bi-level approach in addressing distribution shift. It would be worthwhile to compare the performance with and without the upper/lower levels and assess if regret bounds still hold. * **Significance of Information-Theoretic Results:** Section 4 appears to be disconnected from the rest of the paper. While the Information Theoretic algorithm does not provide an empirically feasible method, it produces a baseline policy improvement scheme using bias and uncertainty errors. However, the significance of these results is unclear. To the best of my knowledge, the information theoretic results are not used to generate the penalized adversarial estimation algorithm. Neither they are used in the empirical evaluation or in producing a tighter regret bound. It would be helpful if the authors could explain the significance and utility of these results and how they should be interpreted. It would also be helpful if the paper could throw light on the benefit of near-optimal regret in Corollary 4.1. * **Utility of $\tilde{\sigma_{n}}$:** From my understanding, $\tilde{\sigma_{n}}$ trades off pessimism for policy improvement. This indicates that the utility of $\tilde{\sigma_{n}}$ is not well understood. Setting $\tilde{\sigma_{n}}$ to a lower value can hinder policy performance as the lower level will yield an overly conservative value function. On the other hand, setting $\tilde{\sigma_{n}}$ to a higher value leads to distribution shift. It would be helpful to have an optimal value of $\tilde{\sigma_{n}}$ characterized or a criterion for selecting $\tilde{\sigma_{n}}$ based on policy performance or pessimism. ## Minors * line 29: amplifying -> amplifies * line 32: to contain -> contains * line 79: discounted -> discount * line 125: sepcified -> specified * line 137: established the CI for $J(\pi)$ is presented in Theorem 3.1. * line 234: concentrability -> concentratibility * line 275: conditional -> conditioned * line 329: set propertly choose -> properly choose

Questions

* Can the empirical evaluation demonstrate the bounds of Theorem 4.1 and Theorem 5.2? Can experiments show the benefit of bi-level approach in tackling distribution shift (i.e- comparing with and without upper/lower levels)? * Can you please explain how are the Information Theoretic results useful? Are these results used to generate the penalized adversarial estimation algorithm? If not, then how should one interpret Theorem 4.1 and Corollary 4.1? * What is the optimal $\tilde{\sigma_{n}}$? Is there a criterion for setting $\tilde{\sigma_{n}}$?

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

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

2 fair

Presentation

3 good

Contribution

4 excellent

Limitations

Authors have discussed limitations and future work.

Reviewer UsZq6/10 · confidence 2/52023-07-07

Summary

This paper presents a bi-level structured policy optimization algorithm for offline reinforcement learning (RL). The algorithm aims to address the distributional shift and function approximation challenges in offline RL by modeling a hierarchical interaction between the policy and the value function. The lower level focuses on constructing a confidence set of value estimates to control uncertainty arising from distribution mismatch, while the upper level aims to maximize a conservative value estimate. The paper provides theoretical guarantees for the algorithm and evaluates its performance on both synthetic and real-world datasets.

Strengths

- The paper proposes a novel bi-level structured policy optimization algorithm that addresses the challenges of distributional shift and function approximation in offline RL. This algorithm offers a comprehensive solution for real-world scenarios. - The algorithm preserves the flexibility of the exploratory data distribution and enables model extrapolation, enhancing its power. - The paper provides theoretical regret guarantees for the proposed algorithm without relying on data-coverage and completeness-type assumptions, increasing its applicability.

Weaknesses

- There is only a weak connection between several parts of the paper. It is hard to find a central claim that supports the logical flow. - There remain some issues unsolved in the paper. See the questions for details.

Questions

- At the lower level, how can the confidence set, which constraints model learning, effectively prevent overly pessimistic evaluation? - How can the proposed algorithm build on weaker conditions than the global data coverage? Does it benefit from the occupancy distribution estimation?

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

Yes

Authorsrebuttal2023-08-16

Follow-up on Our Response to Your Feedback

We greatly appreciate your valuable and constructive comments, as well as the time and effort devoted! Following your insightful suggestions, we have conducted ablation studies and additional experiments to clarify the point of preventing overly pessimistic evaluation. Furthermore, we have provided justification for our data-coverage conditions, and the utility of the occupancy distribution. Additionally, we have also strengthened the central claim, ensuring coherence across the paper. All the details can be founded in our response and the attached one-page PDF file. We sincerely hope our additional clarifications and experiments can fully address your questions and can be helpful in the evaluation of our results. We are eagerly looking forward to your kind feedback.

Reviewer 7QtD7/10 · confidence 4/52023-07-09

Summary

This paper proposed a pessimistic offline RL method. Different from the previous method in the function approximation setting, the method proposed in this paper does not require coverage and completeness assumption. Moreover, in terms of algorithm design, this paper derives a lower confidence bound for RL objective function and explicitly optimizes it to obtain the policy, which is the first attempt in this direction. Both theoretical and empirical results are presented to support the effectiveness of this algorithm design.

Strengths

(1) This paper proposes a novel algorithm design to address the offline policy optimization problem by explicitly optimizing a lower bound of RL objective function. This design is first proposed to my best knowledge, which is valuable. (2) The algorithm design proposed in this paper has a very global optimality guarantee under weaker assumptions compared with previous works. (3) The author also derives a practical version of the algorithm which also has a theoretical guarantee. Moreover, to demonstrate the effectiveness of this practical version, several experiment results are presented.

Weaknesses

The experiment is not very strong: (1) Most of the standard simulation experiments are not considered thus the practical effectiveness of the algorithm is not convincing. The author might need to include more experiments are also included in OptiDICE paper (hopper, walker, halfcheetah, maze etc) (2) I was also wondering how the algorithm proposed in this paper compared most recent baseline such as TD3 + BC?

Questions

(1) If $\mathcal{U}^\tau_{\infty}=+\infty$, then how would the constant $\mathcal{U}^\tau_2$ and $\mathcal{U}^*_2$ look like, will them become unbounded as well? If that is the case, then the proposed algorithm still needs completeness assumption and coverage assumption to derive the bound in Theorem 5.1.? I don't think requiring coverage assumption for this paper should be a reason for rejection since this paper already made a lot of contributions, but the author needs to provide a clear discussion for this part otherwise it will be somehow misleading since it is not obvious how to achieve global optimal policy without any coverage requirements. (2) In linear MDP the algorithm complexity is $n^{-1/4}$? If so then why the algorithm has worse dependence on $n$? Can the author also provide some high-level explanation? (3) OptDICE has been proven to outperform CQL and other baselines (see the original paper of OptiDICE, why OptiDICE performs much worse than CQL? I will raise my score if the author can answer the above questions appropriately.

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

No limitation.

Authorsrebuttal2023-08-16

Eagerly Looking Forward to Your Feedback on Our Response

We are very grateful for your time and effort devoted to reviewing our work, as well as for giving us insightful, valuable, and encouraging comments! Following your constructive suggestions, we have conducted additional four sets of experiments on D4RL benchmarks and comparing with the most recent baseline TD3+BC. Furthermore, we have clarified the boundedness on $\mathcal{U}^\tau_2 (\mathcal{U}^\star_2)$, the coverage and completeness-type assumptions, and the sample complexity, among other points. All these details are provided in our response and the accompanying one-page PDF file. We sincerely hope our response can appropriately address your concerns and can be helpful in the evaluation of our work. We are eagerly looking forward to your kind feedback!

Reviewer 75WN4/10 · confidence 4/52023-07-31

Summary

This paper addresses the problem of offline reinforcement learning (RL). Offline RL involves computing a policy based on a static/pre-collected dataset, with no further interaction with the environment permitted. A significant hurdle in offline RL is the distributional shift resulting from insufficient exploration within the dataset. Moreover, previous work may necessitate the assumption of Bellman completeness for the utilized function class. To overcome these challenges, the authors propose a bi-level structured policy optimization algorithm. This algorithm models a hierarchical interaction between the policy and the value function. At the lower level, the focus is on constructing a set of confident value estimates. The upper level, in contrast, aims to maximize a conservative estimate derived from this confidence set. The algorithm can be solved efficiently through a penalized adversarial estimation procedure. The authors provide theoretical regret guarantees for their algorithm and assess its performance using both synthetic and real-world datasets. Performance is compared against several well-known offline RL algorithms, such as CQL, IQL, and COMBO.

Strengths

Strengths: 1. The concept of optimizing the confidence lower and upper bounds in equation (4) over τ within Ω, with the aim of tightening the Confidence Interval (CI), is intriguing and holds potential for generating a more precise confidence interval. 2. The paper, overall, is well-written and well-structured, demonstrating a clear and logical progression of ideas. 3. The authors have given a comprehensive view of their algorithm by providing both theoretical analysis and empirical versions. Additionally, the experimental results demonstrate the effectiveness of the proposed algorithm, as shown on various synthetic datasets as well as a real-world dataset.

Weaknesses

Weaknesses: 1. Complexity: The algorithm, in comparison to alternatives such as IQL and COMBO, is notably more complex. It employs a game-theoretical formulation and requires the resolution of two optimization problems at each iteration. This results in an increase in both the number of hyperparameters that require tuning and the overall computational cost, and potentially instability in the optimization (i.e., convergence). 2. Ambiguities and Clarifications in Section 3 (Lines 119 to 126): - Shouldn't $G (x_∗) = \sup_x \{x \cdot x_∗ − G(x)\}$ be revised to $G_* (x_∗) = \sup_x \{x \cdot x_∗ − G(x)\}$? - For $J(\pi) = q_\pi(s_0, a) + E_\mu[\lambda G_*((B^\pi q^\pi (s, a) − q^\pi (s, a))/(1 − \gamma))]$, where $\lambda \geq 0$, the initial term appears as if it should be $q_{\pi}(s_0, \pi)$. Additionally, could you explain how $\lambda$ is typically chosen? Does the choice of $\lambda$ significantly impact the algorithm's performance? - The derivation from $J(\pi) = q_\pi(s_0, a) + E_\mu[\lambda G_*((B^\pi q^\pi (s, a) − q^\pi (s, a))/(1 − \gamma))]$ to equation (1) is unclear. Why is $\lambda$ merely a coefficient of $G(x)$? Shouldn't it also be a coefficient of $x · (B^\pi q^\pi (s, a) − q^\pi(s, a))$? - The terms L(τ,q,π) and ξ(G,τ) require explicit definitions. 3. Justification for Assumption 2: The justifiability of Assumption 2 is uncertain. It appears to be more of a requirement for feasible derivation rather than a naturally occurring condition. Can it be guaranteed that $\tau(s,a)$ won't reach infinity in real-world problems? More justification on this point would be appreciated. 4. Need for Intuition: More insight into why your framework doesn't require the Bellman completeness assumption would be beneficial. 5. Additional Evaluations: It would be helpful if you could evaluate your proposed algorithm using more standard benchmarks, such as D4RL. This would significantly aid in positioning your work within the existing literature.

Questions

See Weakness.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

See Weakness.

Reviewer C4A22023-08-13

Response to Author's Comments

I thank the authors for providing a detailed response and addressing most of my concerns. After going through authors' responses and other reviewers' comments, my concern regarding the significance of information theoretic results still remains unaddressed. * **Significance of Information-Theoretic Results**- While the information-theoretic bound is used for decomposition analysis, I still struggle to see its direct connection with the penalized adversarial estimation algorithm. If the central contribution of this result is in characterizing the role of off-support and on-support errors, then the result can be considered as an intermediate step in the overall algorithm development. At best, we can conclude that the result leads to a theoretically tighter bound of Theorem 4.1 whose practical implementation still remains an open question.

Authorsrebuttal2023-08-14

Response to Reviewer C4A2

Thanks for your timely response and for agreeing with our explanations to most of your questions! To better answer your question and avoid any unnecessary confusion, we first briefly summarize the logical flow of the paper. 1. We propose a robust OPE procedure. 2. We leverage the above procedure to construct a bi-level policy optimization algorithm, which is our information-theoretic algorithm. 3. We establish strong regret guarantees for the information-theoretic algorithm, e.g., Theorem 4.1, and we name the results as information-theoretic results. 4. We find our information-theoretic algorithm is computationally inefficient due to solving constrained optimization, while the information-theoretic results demonstrate desirable theoretical properties of the algorithm. 5. We propose a practical algorithm, i.e., the penalized adversarial estimation, for implementation and empirical evaluation. 6. We study the theoretical properties of the practical algorithm, i.e., Theorem 5.1, and find it inherits desirable characteristics of the information-theoretic algorithm but with sample complexity degradation. 7. We perform a study of the practical algorithm in linear MDPs. > **The importance of the information-theoretic results**: By the logical flow, we first develop the information-theoretical algorithm and it is natural to study its theoretical properties. In addition, there are several major reasons that explain why the information-theoretic results matter. - As a theoretical-based paper (while we still provide a practical algorithm with comprehensive empirical evaluation), our information-theoretic results help to position our work in the existing literature and highlight our theoretical contributions in offline RL. We can fairly compare our results to the existing information-theoretic results, e.g., [1,2,3,4], in terms of the function approximation and coverage conditions, sample complexity, horizon dependency, and bound tightness. Following the convention, all these existing algorithms are purely information-theoretic without practical implementation. - The information-theoretic findings enhance the understanding of the practical algorithm. Specifically, a comparison between Theorem 4.1 and 5.1 reveals that Theorem 4.1 has a tighter bound than Theorem 5.1. This delivers an important message: to attain computational efficiency, the practical algorithm compromises on sample efficiency, i.e., $O(n^{-1/4})$ vs $O(n^{-1/2})$. - The information-theoretic results provide a well theoretical justification for the information-theoretic algorithm. While the algorithm currently grapples with computational efficiency challenges, resolving these issues could offer significant advantages. This paves the way for developing a more refined algorithm in offline RL. > **The connection between the information-theoretic results and the practical algorithm**: By the logical flow, although the information-theoretic results are not directly developed for the practical algorithm, they do have some indirect relationships. The error decomposition technique is shared between the information-theoretic results and the results for the practical algorithm, while the development of the latter is much more involved (optimization error, mirror-descent update, and a trajectory-dependent bound). Second, following a similar proofing strategy, the results in Theorem 4.2, Corollary 4.1, 4.2 can be established for the practical algorithm as well. To keep completeness, we will add these new results in the updated supplementary materials. > **Bound tightness**: In terms of the dependency on $n$ or horizon, Theorem 4.1 is an ultimate regret bound and has achieved the minimax results [5]. Furthermore, as indicated by Theorem 4.1, it observes a trade-off between bias $\epsilon_b$ and variance $\epsilon_\sigma$, affecting the bound tightness. In fact, Theorem 4.1 intrinsically chooses the best bias-variance decomposition through the best choice of $\mathcal{U}^\star_2$. It's important to note that this decomposition is solely a proof technique, not a knob in the algorithm. Combine with the adaptation $\rho$ via $\min\{\rho: ||\rho/\mu||_{L_2(\mu)} \leq \mathcal{U}^\star_2\}$, we can conclude that our regret bound is hard to be further improved unless with additional conditions e.g., coverage or explicit MDP structure. We hope these explanations adequately address your concern. We remain dedicated to refining our paper and welcome any further comments. Once again, thank you for your invaluable contribution to our work! **References:** [1] Jin, Y., et al. (2021), Is pessimism provably efficient for offline rl, ICML [2] Uehara, M., et al. (2022), Pessimistic model-based offline rl, ICLR [3] Zhan, W., et al. (2022), Offline rl with realizability and single-policy concentrability, COLT [4] Chen, J., et al. (2022), Offline rl under value and realizability, UAI [5] Xie, T., ta al, (2020), Q* approximation for batch rl, UAI

Reviewer C4A22023-08-18

Response to Comments

Thank you for the response. The discussion on tightness of bound and importance of results above clarifies my concern. I encourage the authors to add the above discussion as part of the Appendix. It would also be worthwhile to tailor Section 4 in order to highlight the tightness of bound and its role in motivating the empirical algorithm. Since the authors have addressed all of my concerns pertaining to the validation of theoretical results (regret bound, ablation study and information-theoretic results) and utility of the most important hyperparameter, I would like to raise my score and support acceptance of the work. I thank the authors for their efforts.

Authorsrebuttal2023-08-18

Thank you for your constructive and positive feedback

We are grateful for your recognition and support for the acceptance of our work! It's glad to know our response sufficiently addressed all your concerns. We really enjoy having such insightful discussions with the reviewer. Your invaluable feedback during the discussion and rebuttal phase has definitely enhanced the quality of our paper. Following your constructive suggestions, we will incorporate a discussion section in the updated Appendix and emphasize the significance of the regret bound and its influence on the practical algorithm in Section 4. Once again, thank you for the time and contributions you've devoted to our work!

Reviewer 75WN2023-08-14

Thank you for the response. I appreciate the additional results. However, it seems that the performance was significantly influenced by the hyperparameters $ \lambda $ and $ c^\star $, with the best results achieved around 50, and the lowest at 30. I wonder if the authors could provide more details about the experimental setup for the D4RL experiments, as the devil is often in the details. Also, did you report the results with the best hyperparameters tuned online? Did you tune the hyperparameters for each task as well? If that's the case, it might be unfair to claim the algorithm performs the best, as methods like COMBO, they did not tune the hyperparamters online.

Authorsrebuttal2023-08-15

Response to Reviewer 75WN

Thank you for your kind reply! We think many of your suggestions and comments are of great value to make our paper stronger. > **Sensitivity Analysis:** Thanks for your careful observation! In the sensitivity analysis, we intentionally challenge our algorithm and test whether it works in extreme cases (over-pessimism and inadequate-pessimism) **by setting** $\lambda, c^*$ **with extremely abnormal values**. We can find that the hyperparameter values that result in the $30$ return are set at $\lambda = c^*=10$ (over-pessimism). These values are **scaled 100 times larger than the** $\lambda=c^*=0.1$ **yielded by our selection rule**. A similar pattern can be observed when we set $\lambda=c^*=0.0001$ (inadequate pessimism), which is 100 times smaller than our selected choice. In contrast, when the $\lambda, c^*$ are set in a relatively normal range, e.g., $\lambda, c^* \in [1,0.001]$, the policy performance is **quite stable** with the highest around $50$ and the lowest around $47$ return. > **Experiments details and hyperparameter tuning:** For all D4RL tasks, we **fully offline** tune our hyperparameters following our selection rule (discussed later), instead of online tuning. We provide the experiment details as follows. **Our algorithm:** We use separate $2$-layer fully connected neural networks for function approximation, where each hidden layer width is $256$ with ReLU activation. The policy is Gaussian, with the mean and the standard deviation parameterized by the neural network. We use Adam as the optimizer with a minibatch size $256$. We set the learning rate $\eta^t$ for the $t$-th iteration be $\frac{\eta^0}{1+0.3\cdot t^{1/4}}$, where the $\eta^0$ is fixed to be $1e-3$. This learning rate follows the stepsize rule in Theorem A.1. We set the stepsize $\zeta = 3e-4$, which was selected offline with a heuristic: we consider the set {$1e-4, 3e-4, 1e-5$}, and found $\zeta = 3e-4$ works well for almost all halfcheetah and maze2d tasks. For each D4RL task, our hyperparameters $c^{*}$ and $\lambda$ are set **fully offline** by following the selection rule satisfying the conditions in Theorem 5.2, i.e., $\frac{2\cdot n^{1/4}}{3\cdot d\log(\bar{V}\sqrt{n})}$. We provide our selected hyperparameters in each task as follows. | Gym locomotion | Hypereparameters | |----------------------------|------------------| | walker2d-medium | 0.25 | | walker2d-medium-replay | 0.1 | | walker2d-medium-expert | 0.35 | | walker2d-random | 0.25 | | hopper-medium | 0.4 | | hopper-medium-replay | 0.25 | | hopper-medium-expert | 0.5 | | hopper-random | 0.4 | | halfcheetah-medium | 0.25 | | halfcheetah-medium-replay | 0.1 | | halfcheetah-medium-expert | 0.35 | | halfcheetah-random | 0.25 | | **Maze2d** | | | maze2d-umaze | 2 | | maze2d-medium | 2.25 | | maze2d-large | 2.5 | **We provide the experiment details for the competing methods as follows:** TD3+BC: We mostly follow default setups in the original paper but we set $\lambda=\frac{\alpha}{\frac{1}{n}\sum_{(s, a)}|q(s, a)|}$ as suggested in [1], to improve the policy performance. CQL: We follow the author-released github default setup but we modify the actor learning rate and use a fixed $\alpha$ instead of the Lagrange variant. This modification is to match the hyperparameters defined in the original paper, and [1] found the original hyperparameters performed better. IQL: We follow the author-released default setups. COMBO: We take the results for Gym locomotion from the original paper, and re-run the algorithm in Maze2d. In Maze2d tasks, we apply the author-released selection rule to the set {$1e-4, 3e-4$} for the $q$-function learning rate and the set {$1e-5, 3e-5, 1e-4$} for the policy learning rate: $1e-4$ and $3e-5$ are selected. We consider $\beta \in \{0.01, 0.1, 0.5, 1, 2.5\}$ for conservatism degree, and select $\beta = 1$ for implementation. We leave others as default setups. ATAC: We take the results for Gym locomotion from the original paper, and re-run the algorithm in Maze2d. We follow the author-released online selection rule for $\beta$ in the set {$4^{-4}, 4^{-3}, 4^{-2}, 4^{-1}, 4,4^2, 4^3$}. And we finally select $\beta = 4$, $4^{-1}$ and $4^{-1}$ for the maze2d-umaze, medium and large, respectively. We leave others as default setups. OptiDCIE, BCQ, BEAR: Results are from the original or D4RL paper. $~$ We hope these explanations adequately address your concern. We really enjoy having such insightful discussions with the reviewer, and thank you once again for your invaluable contributions to our work! **Reference** [1] Fujimoto, S., et al. 2021, A minimalist approach to offline rl. Neurips

Authorsrebuttal2023-08-18

Additional Response to Reviewer 75WN

We performed **additional sensitivity analyses on more D4RL environments** (we just exhausted our HPC resources:) ) and hope the new experiment results can help to resolve the reviewer's concern about hyperparameter-tunning. $~$ We again greatly appreciate the reviewer for your valuable comments and time devoted! All your suggestions have been reflected in our revised paper. We kindly ask the reviewer to consider raising their evaluation if their concerns were appropriately addressed, and we always welcome further feedback! $~$ The following tables provide the experiments of sensitivity analyses on the values of the hyperparameters vs policy performance on the additional D4RL benchmarks (**hopper, walker2d, maze2d**), in addition to the results (halfcheetah) we previously presented in the one-page PDF. Each number in the following tables is the normalized score of the policy at the last iteration of training, averaged over $3$ random seeds. $~$ **(Gym) Hopper-medium-replay**: Our selection rule chooses $\lambda=c^*=0.25$ with the policy performance $114.0 \pm 2.4$. | $c^{*}(\text{col}), \lambda(\text{row})$ | 2.5 | 1 | 0.1 | 0.01 | 0.0025 | 0.001 | |-----------------------------------------|--------------|--------------|--------------|--------------|--------------|--------------| | **2.5** | $108.1\pm2.7$| $109.7\pm2.4$| $111.6\pm3.1$| $111.1\pm2.7$| $109.5\pm2.1$| $108.3\pm4.4$| | **1** | $109.3\pm1.7$| $111.8\pm2.4$| $113.2\pm2.6$| $112.9\pm2.2$| $112.0\pm3.0$| $110.7\pm3.3$| | **0.1** | $112.6\pm2.1$| $113.3\pm2.0$| $114.4\pm2.9$| $114.6\pm2.1$| $113.2\pm2.9$| $112.5\pm3.4$| | **0.01** | $111.8\pm2.8$| $112.0\pm3.6$| $114.6\pm2.9$| $114.2\pm3.3$| $113.1\pm2.7$| $110.2\pm3.4$| | **0.0025** | $109.7\pm2.6$| $111.5\pm3.2$| $113.8\pm2.6$| $113.3\pm4.4$| $112.6\pm3.7$| $110.1\pm3.7$| | **0.001** | $108.2\pm3.0$| $109.5\pm3.8$| $111.9\pm3.5$| $111.2\pm2.6$| $109.8\pm3.1$| $108.4\pm4.6$| $~$ **(Gym) Walker2d-medium-replay**: Our selection rule chooses $\lambda=c^*=0.1$ with the policy performance $101.2\pm3.2$. | $c^{*}(\text{col}), \lambda(\text{row})$ | 2.5 | 1 | 0.1 | 0.01 | 0.0025 | 0.001 | |-----------------------------------------|--------------|--------------|--------------|--------------|--------------|--------------| | **2.5** | $95.8\pm2.5$ | $97.4\pm2.8$ | $98.4\pm2.5$ | $99.1\pm3.2$ | $97.8\pm3.0$ | $97.9\pm3.4$ | | **1** | $97.3\pm2.7$ | $98.0\pm3.1$ | $98.8\pm2.8$ | $99.4\pm3.4$ | $98.7\pm2.7$ | $98.1\pm3.2$ | |**0.1** | $97.4\pm2.8$ | $98.3\pm2.9$ | $101.2\pm3.2$| $101.3\pm3.4$| $98.9\pm3.6$ | $97.5\pm4.2$ | | **0.01** | $98.2\pm2.8$ | $99.5\pm2.9$ | $101.7\pm3.9$| $102.6\pm3.4$| $100.2\pm3.1$| $98.4\pm3.3$ | | **0.0025** | $98.0\pm3.6$ | $97.5\pm4.2$ | $100.1\pm3.6$| $100.8\pm3.5$| $99.2\pm4.2$ | $97.4\pm4.0$ | |**0.001** | $97.2\pm3.8$ | $98.5\pm3.3$ | $98.2\pm3.8$ | $99.3\pm3.6$ | $97.8\pm5.2$ | $98.2\pm4.1$ | $~$ **Maze2d-medium**: Our selection rule chooses $\lambda=c^*=2.25$ with the policy performance $138.1\pm 7.6$. | $c^{*}(\text{col}), \lambda(\text{row})$ | 15 | 10 | 5 | 2.5 | 1 | 0.5 | |----------------------------------------|---------------|--------------|---------------|---------------|---------------|---------------| |**15** | $134.5\pm4.6$ | $133.9\pm5.8$ | $134.8\pm4.5$ | $136.7\pm6.2$ | $134.8\pm6.0$ | $134.9\pm5.4$ | |**10** | $133.7\pm4.2$ | $136.7\pm5.1$ | $135.8\pm6.8$ | $138.4\pm7.4$ | $135.7\pm12.2$ | $137.5\pm8.2$ | |**5** | $133.9\pm5.8$ | $137.3\pm6.9$ | $136.6\pm5.5$ | $138.3\pm9.2$ | $134.9\pm7.0$ | $135.1\pm5.2$ | |**2.5** | $137.5\pm6.3$ | $135.8\pm5.9$ | $140.7\pm10.9$ | $138.9\pm9.2$ | $132.2\pm8.1$ | $133.7\pm6.5$ | |**1** | $134.0\pm4.2$ | $137.2\pm10.7$| $133.8\pm6.9$ | $137.3\pm9.5$ | $138.2\pm5.2$ | $137.6\pm8.0$ | |**0.5** | $135.2\pm11.8$ | $133.7\pm8.3$ | $136.1\pm7.8$ | $134.5\pm6.7$ | $137.2\pm9.2$ | $135.5\pm7.1$ | $~$ From the tables, we can see that, our algorithm demonstrates robustness over a wide value range of hyperparameters. Also, the policy performance under our hyperparameter choice is close to the best performance in the table, which indicates the effectiveness of our proposed hyperparameter selection rule.

Reviewer UsZq2023-08-19

Response

I acknowledge the authors' rebuttal and I remain my rating towards acceptance.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC