Cal-QL: Calibrated Offline RL Pre-Training for Efficient Online Fine-Tuning

A compelling use case of offline reinforcement learning (RL) is to obtain a policy initialization from existing datasets followed by fast online fine-tuning with limited interaction. However, existing offline RL methods tend to behave poorly during fine-tuning. In this paper, we devise an approach for learning an effective initialization from offline data that also enables fast online fine-tuning capabilities. Our approach, calibrated Q-learning (Cal-QL), accomplishes this by learning a conservative value function initialization that underestimates the value of the learned policy from offline data, while also being calibrated, in the sense that the learned Q-values are at a reasonable scale. We refer to this property as calibration, and define it formally as providing a lower bound on the true value function of the learned policy and an upper bound on the value of some other (suboptimal) reference policy, which may simply be the behavior policy. We show that offline RL algorithms that learn such calibrated value functions lead to effective online fine-tuning, enabling us to take the benefits of offline initializations in online fine-tuning. In practice, Cal-QL can be implemented on top of the conservative Q learning (CQL) for offline RL within a one-line code change. Empirically, Cal-QL outperforms state-of-the-art methods on 9/11 fine-tuning benchmark tasks that we study in this paper. Code and video are available at https://nakamotoo.github.io/Cal-QL

Paper

Similar papers

Peer review

Reviewer J85q5/10 · confidence 2/52023-07-06

Summary

The paper introduces a method called Calibrated Q-learning (Cal-QL) for efficient online fine-tuning. This method learns an effective initialization from offline data, which enables fast online fine-tuning. It does so by underestimating the value of the learned policy from offline data while ensuring that the learned Q-values are at a reasonable scale.

Strengths

This work focuses on the offline pre-training phase for better online fine-tuning and identifies why CQL unlearn initially, which is new to me. The paper is well-written and easy to follow.

Weaknesses

My main concerns are listed below: 1. Does the proposed approach increase the possibility of overestimation, given that it selects values larger than the behaviour policy? This question particularly arises when dealing with datasets that lack diversity or expert data. 2. The baselines, such as CQL and AWAC, were tested on locomotion tasks, such as Hopper, Walker2d, and HalfCheetah. However, I did not see experiment results on these datasets. Does your approach perform poorly on these tasks? 3. Some newly presented offline-to-online baselines, such as off2on[1], are not included. [1] [Offline-to-Online Reinforcement Learning via Balanced Replay and Pessimistic Q-Ensemble

Questions

Why not conduct experiments on continuous control tasks similar to those performed by CQL and AWAC?

Rating

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

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

2 fair

Contribution

2 fair

Limitations

not applicable

Authorsrebuttal2023-08-16

Dear Reviewer J85q, Thank you for your feedback! We were wondering if you have gotten a chance to go over our responses -- we have now added the results on D4RL locomotion datasets, added an experiment to compare Cal-QL to SPOT (Wu et al 2022, a baseline similar to Off2On), and also provided clarifications on your other concerns. We would appreciate it if you could have a look. We are happy to discuss further. Thanks!

Authorsrebuttal2023-08-20

Any remaining concerns

Dear Reviewer J85q, As the author-reviewer discussion period draws to the close in less than 48 hours, we were hoping to check in with you if you have gotten a chance to go over our responses and if your concerns are addressed. We have performed several additional experiments (details in the rebuttal PDF) and would be grateful if you engage in a discussion with us. We are happy to address any remaining concerns. Thank you so much!

Reviewer zhZZ5/10 · confidence 4/52023-07-06

Summary

The paper proposed a new approach to Offine-RL Online Fine-Tuning called calibrated Q-learning (Cal-QL). This approach aims to learn an effective initialization from offline data that enables fast online fine-tuning capabilities. Cal-QL accomplishes this by learning a conservative value function initialization that underestimates the value of the learned policy from offline data while ensuring that the learned Q-values are at a reasonable scale. The contributions of this paper include: 1) The paper theoretically analyzes the cumulative regret obtained by online fine-tuning, when the value function pre-training with Cal-QL. 2) Empirically, Cal-QL outperforms previous methods on 9/11 fine-tuning benchmark tasks.

Strengths

1. Writing is clear and understandable. 2. The paper proposes a new definition (calibration) and gives the theoretical analysis of Cal-QL. 3. The method can be implemented on top of conservative Q-learning (CQL) for offline RL with few code changes. 4. The Cal-QL is better than the baseline methods in most tasks.

Weaknesses

1. The assumption of Cal-QL seems a bit too strong, that Cal-QL will be particularly effective in controlling the efficiency of online fine-tuning only when the reference policy μ is close to the narrow expert policy. 2. Cal-QL is also most related to methods that utilize a pessimistic RL algorithm for offline training but incorporate exploration in fine-tuning [1,2]. However, these two important baseline methods are missing from the experimental comparison results (Figure 6, Table 1 and 2). 3. The coefficient (\alpha) of the conservative term (in equation 3.1) is an important parameter that directly affects the experimental results. However, the hyperparameter (\alpha) ablation experiments are missing.

Questions

1. How many random seeds are used for the experimental results of Table 1 and Table 2? Same as Figure 6, six random seeds? Moreover, how many random seeds are used in Figure 7 and 8? 2. In Appendix Table 3, for the visual-manipulation task, the paper utilized separate values for offline (\alpha = 5) and online (\alpha = 0.5) phases for the final results. Is the proposed algorithm sensitive to hyperparameters (\alpha and mixing ratio)? To what extent do the final experimental results of different tasks depend on the hyperparameters?

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

1. It would be better if there are more comparative experiments with important baseline methods [1,2]. 2. To test the sensitivity of the method to hyperparameters, there should be more experimental comparisons and analyses of hyperparameters \alpha and mixing ratio. References: [1] S. Lee, Y. Seo, K. Lee, P. Abbeel, and J. Shin. Offline-to-online reinforcement learning via balanced replay and pessimistic q-ensemble. In Conference on Robot Learning, pages 1702–1712. PMLR, 2022. [2] J. Wu, H. Wu, Z. Qiu, J. Wang, and M. Long. Supported policy optimization for offline reinforcement learning. arXiv preprint arXiv:2202.06239, 2022.

Reviewer zucz6/10 · confidence 3/52023-07-07

Summary

The paper proposes a new offline to online algorithm, Cal-QL, that aims to mitigate the unlearning during the beginning of the online fine-tune stage. The paper first conducts empirical experiments to observe that the reason for the unlearning is due to overly pessimistic function values, and thus proposes the Cal-QL algorithm where lower bound the estimated value by the value of the reference policy. With the change to CQL, the proposed algorithm can outperform other Off2On algorithms on various tasks and shows consistent improvements, without the unlearning phenomena. The paper also proposes a theoretical algorithm and shows the regret bound under the bilinear model setting.

Strengths

1. The paper investigates empirically the reason for the initial unlearning phenomena at the beginning of the fine-tuning stage, shows good evidence for the issue of overly pessimistic value estimation, and thus well motivate the calibrated algorithm design. 2. The paper compares with a good range of baselines in offline2online methods, and shows consistent improvements over the previous baselines, and indeed resolves the unlearning issue at the beginning of the fine-tuning stage, and thus provides good practical contribution. 3. The overall algorithm is simple and a good amount of implementation details are provided in the paper. 4. Some theory is also provided to make the paper more complete.

Weaknesses

1. Although it is good to provide many baselines to compare, the current baselines presentation is slightly confusing. I understand that RLPD and HyQ are reasonable baselines to compare with, but the current paper is not clearly comparing with RLPD or HyQ? One would guess that Hybrid RL means HyQ and SAC+offline means RLPD, but it is not clear that in the experiments the paper is indeed comparing with the exact algorithms. 2. The theory algorithm seems to have some gap with the actual Cal-QL algorithm. 3. The completeness assumption in the calibrated case seems strong, because the paper is focusing on a certain strange subset of function class. For example, can one show the completeness assumption holds even in the tabular case?

Questions

See above.

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

3 good

Contribution

3 good

Limitations

N/A

Reviewer vsVq6/10 · confidence 3/52023-07-08

Summary

Initially, this paper conducts an empirical investigation to understand why the offline policy derived from prior methods is inadequate as an initialization for online fine-tuning. Building upon these empirical findings and the derived hypothesis, the authors design Cal-QL, a method differing from CQL by only a single line, aiming to calibrate the Q function. The superiority of Cal-QL is demonstrated through both theoretical and empirical studies.

Strengths

- The authors present meticulous empirical and theoretical explorations justifying the proposed calibration. - The method, while simple, proves to be quite effective. - The paper boasts commendable writing and organization.

Weaknesses

- In section 4.1, Figure 3 shows a clear increase in estimated Q values correlating with the increase in normalized scores. However, it does not clarify whether this increase stems from in-distribution data, out-distribution data, or both. Therefore, the connection between the results in Figure 3 and the conclusion in lines 162-164, which suggest that the rise in estimated Q values is due to the in-distribution data, may not be logically stringent. - Although the locomotion datasets in D4RL are extensively used as benchmarks, the authors did not test Cal-QL on these datasets. It would enhance the study if the authors could include comparisons with baselines on D4RL locomotion datasets.

Questions

Please see Weaknesses.

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

The authors have adequately addressed the limitations and the potential negative societal impact of their work.

Authorsrebuttal2023-08-16

Dear Reviewer vsVq, Thank you for your feedback! We were wondering if you have gotten a chance to go over our responses -- we have now added the results on D4RL locomotion datasets, and also provided clarifications on your other concerns. We would appreciate it if you could have a look. We are happy to discuss further. Thanks!

Reviewer vsVq2023-08-19

Rebuttal response

I am satisfied with the authors' comprehensive response, as it has resolved all of my issues. I recommend that they incorporate these new results and explanations into the updated version of the paper and include error bars in the newly introduced Figure B. With consideration of the thorough response, I'm pleased to increase my rating to 6.

Authorsrebuttal2023-08-20

Thank you for raising your score!

Thank you so much for replying to us and for raising your score! We are glad that your concerns are resolved -- we will incorporate these results and explanations into the final version of the paper, and also include error bars in Figure B. Thanks so much!

Authorsrebuttal2023-08-20

Any remaining concerns

Dear Reviewer zhZZ, As the author-reviewer discussion period draws to the close in less than 48 hours, we were hoping to check in with you if you have gotten a chance to go over our responses and if your concerns are addressed. We have performed several additional experiments (details in the rebuttal PDF) and would be grateful if you engage in a discussion with us. We are happy to address any remaining concerns. Thank you so much!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC