Exclusively Penalized Q-learning for Offline Reinforcement Learning

Constraint-based offline reinforcement learning (RL) involves policy constraints or imposing penalties on the value function to mitigate overestimation errors caused by distributional shift. This paper focuses on a limitation in existing offline RL methods with penalized value function, indicating the potential for underestimation bias due to unnecessary bias introduced in the value function. To address this concern, we propose Exclusively Penalized Q-learning (EPQ), which reduces estimation bias in the value function by selectively penalizing states that are prone to inducing estimation errors. Numerical results show that our method significantly reduces underestimation bias and improves performance in various offline control tasks compared to other offline RL methods

Paper

References (57)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Hh4A5/10 · confidence 4/52024-06-21

Summary

This paper investigates an important problem in offline reinforcement learning (RL), say mitigating unnecessary conservatism in value function. The authors achieve this by selectively penalizing states that are prone to inducing estimation errors, i.e., $f$. The core idea is to train an exclusive penalty $P_\tau=f_\tau^{\pi,\hat{\beta}}(s) \left(\dfrac{\pi}{\hat{\beta}}-1\right)$, where $f_\tau^{\pi,\hat{\beta}}(s)$ is the penalty adaptation factor. $f_\tau^{\pi,\hat{\beta}}(s)$ assigns smaller weights on in-distribution transitions and a larger weight on out-of-distribution (OOD) transitions. Furthermore, the authors propose the *Prioritized Dataset* (PD) trick to reduce unnecessary bias. Based on PD, the authors derive the final optimization objective of their EPQ approach. The authors conduct experiments on numerous D4RL datasets and show that EPQ can outperform previous methods on many datasets.

Strengths

I appreciate that the authors set their focus on value-based offline RL methods and propose an interesting weighting method for mitigating the over-conservatism phenomenon in CQL, akin to mildly conservative Q-learning (MCQ) algorithm. Many recent offline RL algorithms study policy regularization approaches and somewhat neglect the advances of learning offline policies purely from a value function optimization perspective. The overall method is, as far as the reviewer can tell, novel. Though the proposed EPQ method can be seen as one variant of CQL, the proposed method is interesting and can address the unnecessary conservatism on in-distribution data. The presentation of this paper is very nice and I personally quite like it. The authors include many toy examples (e.g., Figure 3, Figure 4) and illustrations of their method (Figure 2). This is of great help in aiding the readers to quickly capture the key points that the authors would like to convey. The authors also compare their method against numerous strong offline RL value-based and policy regularization algorithms and demonstrate that EPQ exhibits quite strong performance on many datasets.

Weaknesses

This paper has the following potential drawbacks - EPQ introduces many hyperparameters, and one needs to manually find the optimal ones on a new dataset. This can impede the practical application of EPQ in real-world problems - The ablation study part is insufficient. The authors only conduct experiments on one single environment in the main text and the appendix (see Figure 6 and Figure 7). This ought to be evaluated on wider datasets to thoroughly examine the hyperparameter sensitivity and how different components/hyperparameters affect the performance of EPQ. The influence of the PD trick also should be investigated on wider datasets. Based on Figure 6, it turns out that EPQ with PD and EPQ without PD exhibit similar performance. - Equation 3 depicts that the target value is *corrected* by the introduced exclusive penalty term $P_\tau$. The final objective (Equation 4) is also derived based on this. Equation 3 reminds me of the RND-series in offline RL, e.g., SAC-RND [1], and SAC-DRND [2]. Typically, they also subtract a penalty term from the target value to pursue conservatism and mitigate overestimation. I think $P_\tau$ can also be viewed as a similar role. Any comments here? Are there any advantages of the introduced penalty term $P_\tau$ over the penalty term given by RND? How these methods can be connected? What if we directly optimize Equation 3 and tune $\alpha$? [1] Anti-Exploration by Random Network Distillation. ICML 2023 [2] Exploration and Anti-Exploration with Distributional Random Network Distillation. ICML 2024.

Questions

I have the following questions - Is there a reason that you do not report the standard deviation of EPQ on halfcheetah-expert dataset in Table 1? - it seems CQL($\alpha=0.0$) in Figure 5(a) middle early stops, it there a reason for this? - how do you expect EPQ to be applied in real-world tasks? Any suggestions or instructions on how to tune the introduced hyperparameters? Some minor points: - I noticed that one cannot be directed to the corresponding images/tables by clicking, e.g. Fig 2. This problem hinders the smooth reading and understanding of the content and should be fixed in the revision - The referenced figures/tables should be distinguished, e.g., Line 90 *Fig. 1(a) and (b)* should be *Fig. 1(a) and 1(b)*. Please check the manuscript and fix all such types of issues. - Equation 5, $Q\_( s,a)$ ==> $Q(s,a)$ - sometimes the authors just write *adroit* instead of *Adroit* (e.g., Line 232) - the captions in Figure 5 and Figure 6 are too simple

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

The authors include an honest discussion part on the limitations of their work in the appendix. I personally agree with that.

Reviewer eWrQ7/10 · confidence 4/52024-06-28

Summary

This paper showed that the popular conservative Q-learning introduces bias into the Q function by its restrictive penalty. It proposed to enforce an adaptive penalty term based on the dataset to avoid bias when the dataset support disagrees with the learned policy. Experimental results supported the authors' claims that the proposed method can effectively reduce estimation bias and outperform many baselines.

Strengths

Conservatism is common in offline RL methods. CQL is one of the most popular algorithms that enforces the conservatism by penalizing its Q value, and hence it runs the risk of introducing bias. This paper nicely illustrated a possible source of bias originating from the difference between dataset actions and policy actions, and proposed to correct the bias by the penalty adaptation factor. The authors justified the proposed penalty with both pedagogical examples and more challenging problems.

Weaknesses

While the paper is well-written in general, I believe the presentation can be further improved by adding making the exlanation more accesible, especially Figure 1 and 2. The two figures served a pivotal role in illustrating the downside of CQL and the core idea of the proposed EPQ. Figure 1 would benefit from separating the information into two parts: (1) overlaps of the dataset/policy actions and (2) estimation bias of CQL under different $\alpha$.

Questions

1. In Figure 2 the authors suggested the relationship > let $\tau_2 < \tau_1$ when $N_1 < N_2$\ However, it seems there is chance that if $\tau_2$ is only slightly smaller than $\tau_1$, then $\tau_2$ would only be penalizing $\pi_3$ and not covering $\pi_1, \pi_2$. More generally, this relates to choosing an appropriate threshold. In the experiments the authors swept $(c\cdot \rho, c\in[0, 10])$, but what was the observation or rule of thumb here to make sure the threshold is meaningful? 2. Another question concerns the implementation. It seems the empirical behavior policy $\hat{\beta}(a|s)$ plays a crucial role in the adaptive penalty. In continuous domains the offline datasets often comprise only a single copy for any action, making the estimation of $\hat{\beta}(a|s) = \frac{N(s,a)}{N(s)}$ rather hard and inaccurate. Therefore, existing work like [1, 2] often use sophisticated models to estimate $\hat{\beta}(a|s)$ with the hope that it would generalize better. By contrast, it is surprising to see that Algorithm 1 used only simple BC and that was sufficient to fuel the superior performance of EPQ. What could be the reason? References:\ [1] Mildly Conservative Q-Learning for Offline Reinforcement Learning [2] Supported Policy Optimization for Offline Reinforcement Learning

Rating

7

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

Potential negative societal impact does not apply and the authors have discussed some of the limitations in the appendix.

Reviewer RxmU8/10 · confidence 4/52024-07-03

Summary

The paper studies the problem of value estimation bias mitigation in offline reinforcement learning. Specifically, the paper takes the well-known Conservative Q-Learning algorithm as a starting point and improves its penalization scheme with an alternative one that applies provably less value underestimation bias. The paper reports results on the well-known D4RL benchmark and compares against alternative approaches.

Strengths

* The paper points to a key problem in model-free offline reinforcement learning, i.e. model underfit caused by overconservatism due to excessive penalty application. * The paper follows a solid methodology to the problem that starts from illustrative data-driven toy results, continues with theoretical analysis, and ends up with a well-understood and well-justified improvement to an established algorithm * The shown results are particularly comprehensive and strong. * The presentation of the paper at stellar level. It is extremely easy to follow the story line, even though it is technically quite dense. * The paper addresses the related literature well.

Weaknesses

The paper builds its whole problem statement and solution on a big assumption: The environment dynamics will not be learned. In other terms, the problems it highlights and addresses are specific to the model-free offline RL approaches, although learning an environment model is not such a major limitation in the offline setting, where the assumption is that data and compute sources are generous, training time is also not an issue, unlike the online MBRL setup. The paper seems to miss this key positioning element in problem formulation. Its presentation will improve if this choice is made more explicit and the impact of the proposed solution is presented accordingly. Minor: I found Figure 2 extremely difficult to grasp. The authors may consider simplifying it a little and extending the caption with an explanation of it.

Questions

* Is there a particular reason why the EDAC algorithm of An et al. [40] is not in the comparison list? It is also model free, reports results on the same benchmarks and it performs better than MISA on Mujoco Tasks, i.e. 85.2*18 = 1533.6. * Section 4.2 does not specify whether the estimation bias is compared between the predicted Q-value and the "discounted" observed return. Can the authors confirm that discounting has been applied in the results shown in Fig 6? By bare eye the bias looked to me too high.

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

4

Limitations

The paper does not address its own limitations and does not specify the potential negative societal impact of the presented work.

Reviewer 36me7/10 · confidence 4/52024-07-17

Summary

This paper introduces a novel approach to handling distribution shift in off-policy RL by the means of Q-function regularization. This is accomplished by modulating a penalty term that is overly conservative in CQL. The authors argue that CQL overcompensates for the distribution shift in cases where state-action pairs exist with non-negligible density in the dataset, but are still less likely under the reference policy than the learned policy. Based on this insight, the authors propose to modulate the CQL value function penalty by a decreasing term in the density of state-action pairs under the reference policy, which activates as soon as this density crosses a threshold. Moreover, the authors propose a prioritized sampling method to further reduce the underestimation bias. The resulting algorithm, called EPQ, achieves superior performance relative to its competitors across many familiar offline RL benchmarks.

Strengths

The motivation for this work was fairly strong, and the authors identified an interesting shortcoming with the overestimation correction in CQL. Figure 1 followed by Figure 4 do a nice job of depicting the influence of this shortcoming, and how the proposed method corrects it. Moreover, the EPQ algorithm is deployed on a large suite of benchmarks, and outperforms all competitors with remarkable consistency. To complement these results, the authors conducted experiments to verify that EPQ does in fact reduce value estimation bias, shown by comparing EPQ value estimates with Monte-Carlo estimates, as well as predictions from CQL. The experiments are conducted over four random seeds. While this is relatively few seeds, I think this is acceptable given the range of tasks that were tested.

Weaknesses

The presentation of the paper (e.g., writing, figures) can be improved. In particular, many of the figures were difficult to read and/or interpret. Both facets of Figure 2 took substantial effort to understand for me (in fact, I think I would have more easily understood the paper without having seen these figures; see Questions below). The statement of Theorem 3.1 is not precise enough, particularly for the latter claim. Moreover, I suspect there are some technical assumptions missing; see Questions below. Furthermore, I do not entirely understand the motivation for the prioritized dataset. Particularly, it is not clear to me that Theorem 3.1 (the theoretical justification for EPQ) actually applies with the prioritized dataset, since the prioritzation depends on the estimated Q-function being updated. Beyond that, the ablation of this feature is not very convincing. While there is a wealth of empirical results, confidence intervals from the baselines is largely lacking—this is especially relevant in Table 1. See for example the `door-cloned` row: EPQ is identified as the best, but its confidence region definitely overlaps that of CQL, and probably those of many of the baselines as well. The same goes for `relocate-cloned`. as well as (I'd suspect) many of the AntMaze tasks. That said, the results do suggest that EPQ frequently outperforms its competitors, and rarely does substantially worse. Finally, it would have been nice to see stronger heuristics for choosing the threshold parameter $\tau$. Figure 2a suggests that $\tau$ should be a function of the amount of data in the dataset, but this is not actually discussed anywhere. Rather, the authors claim to have found a choice for $\tau$ that is inversely proportional to the volume of the action space, but results for this choice are only given on one environment, precluding any conclusion that this choice/trend is good in general. The proof of Theorem 3.1 also gives a condition for determining when $\alpha$ is large enough, which is roughly inversely proportional to the lowest density under the reference policy over all states and actions – therefore, choosing $\alpha$ to be inversely proportional to the volume of the action space is only theoretically justified when the reference policy is uniform. ## Minor Issues The notation/definition of the Bellman operator $\mathcal{B}^\pi$ s not exactly correct. In its definition on line 61, $\mathcal{B}^\pi$ averages over all next states $s'$. Then, in the expression on line 63, you are taking an expectation over all state transitions $(s, a, s')$ in the dataset with $\mathcal{B}^\pi$ evaluated in this expectation. Since $s'$ isn't used anywhere in this expression explicitly, my assumption is that you're using this state as the state to bootstrap from in the application of $\mathcal{B}^\pi$; but then $\mathcal{B}^\pi$ on line 63 is not the same as its definition on line 61. On line 83, you refer to the "actual average return $G_0$", but $G_0$ was defined as the random return (not averaged) on line 56. Formatting of equation (2) is not nice – it almost looks like it's depicting two separate formulas. It may read easier if you instead colored the two factors and described their influence in the text below. In Figure 1 and Figure 4, it would be very helpful to see where $0$ lies on the y-axis on the estimation bias side. Figure 2a is very busy and difficult to interpret. Firstly, I think it would be better to highligh the magnitude of the penalty itself, as opposed ot the penalty reduction (which implicitly depends on some initial penalty, I'm guessing from CQL). Moreover, the relationship between the amount of data and $\tau$ should be discussed before this figure, even if superficially (e.g., with a sentence that says that $\tau$ decreases as you collect more data). Then, the figure would have a much more clear interpretation: increase the amount of data, and the penalty will be relaxed more aggressively. There appears to be a formatting error on line 140, "Proof) Proof…". There is a formatting error in equation (5), $Q_(s, a)$ -\> $Q(s, a)$. On line 243, $200k$ should be $200\mathrm{k}$. In Table 1, the "total" rows are not good indications of performance, firstly because the returns for the different environments are not normalized. That said, the results for EPQ still look good if you neglect the "total" rows. In Figure 5, it would help to use a different line style to emphasize which curve corresponds to EPQ. I found it difficult to distinguish EPQ from CQL ($\alpha=0.0$) – fortunately these curves generally occupied disjoint regions in the graphs.

Questions

In Figure 1, what is the relationship between $\tau_i$ and $N_i$ ($i=1,2$)? Such a relationship has not been discussed up to this point. In the proof of Theorem 3.1, I believe some assumptions are missing. If $\pi(\cdot\mid s)$ is ever supported on an action $a'$ such that $\hat{\beta}(a'\mid s) = 0$, then $\Delta^\pi_{EPQ}\to\infty$. Therefore, there would be no $\alpha$ large enough to underestimate the value function for your argument on line 465 as long as there exists a single $(s, a)$ in the dataset for which $\xi^\delta(s, a) > 0$. Since $\hat{\beta}$ was defined to be the empirical conditional distribution over actions from the dataset, this result actually suggests that $\alpha$ must be infinite whenever your dataset does not fully cover the action space (which is always the case in the experiments, where the action space is continuous). In table 1, why aren't confidence intervals given for the baseline methods? Why are there no confidence regions shown in Figure 6? Particularly, it would have been helpful to see these in Figure 6a. As it stands, the effect of the prioritized dataset depicted in this figure is a little underwhelming. With regard to the analysis of the penalty threshold, why should we scale $\tau$ linearly with the density of the uniform distribution over $\mathcal{A}$ (that is, inversely proportional to the volume of the action space)? Figure 6b does not indicate whether the choice of $\tau = 0.2\rho$ is actually a good choice across environments, so indeed it could have just been that $\tau=0.2\rho$ happens to work well in `hopper-medium` by chance. You have have investigated (or at least presented the results) that show how this form of scaling with the action space performs.

Rating

7

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

Limitations are mostly discussed, except for potentially missing assumptions in Theorem 3.1.

Reviewer 36me2024-08-08

Thanks to the authors for the detailed response. **Choice of $\tau$**: Thanks a lot for the discussion and the additional experimental data. This is helpful, and this (in my opinion) is a more useful argument / advice for choosing this parameter than what was originally stated in the submission. **Confidence intervals**: Thanks again for the additional experimental data, I am satisfied with these results now. **Penalizing constant $\alpha$**: I think I see what you mean, though it would be helpful maybe if you could write out this deduction explicitly. For example, explicitly show why $\alpha \geq \mathsf{corrected bound}$ leads to underestimation.

Authorsrebuttal2024-08-09

Thank you for the prompt response from the reviewer. To provide a more intuitive understanding of penalizing the constant $\alpha$, as requested by the reviewer, we will explicitly explain below how the value function of EPQ underestimates the true value. **Explicit derivation of Theorem 3.1 to show underestimation** We start from line 462 in Appendix A.1, which states that when $V_{k+1}$ converges to $V_\infty$, then the converged value function $V_\infty$ of EPQ satisfies $V_\infty(s) = V^\pi(s) + (I-\gamma P^\pi)^{-1}\cdot$ \{ $- \alpha\Delta_{EPQ}^{\pi}(s) + \mathbb{E}_{a\sim\pi}[\xi^\delta(s,a)]$\}, where $\xi^\delta(s,a)$ and $\Delta_{EPQ}^{\pi}(s)$ are positive $\forall~ s,a$, assuming $\pi\neq\hat{\beta}$. (if $\pi=\hat{\beta}$, then there will be no overestimation error.) If we choose the penalizing constant $\alpha$ that satisfies $\alpha \geq \max_{s,a\in D}[\xi^\delta(s,a)]\cdot\max_{s\in D} (\Delta_{EPQ}^\pi(s))^{-1}$, then $- \alpha\cdot\Delta\_{EPQ}^{\pi}(s) + \mathbb{E}\_{a\sim\pi}[\xi^\delta(s,a)] $ $\leq- \max\_{s,a\in D}[\xi^\delta(s,a)]\cdot \underbrace{\max\_{s\in D} (\Delta\_{EPQ}^\pi(s))^{-1} \cdot \Delta\_{EPQ}^{\pi}(s)}_{\geq 1} + \mathbb{E}\_{a\sim\pi}[\xi^\delta(s,a)]$ $\leq- \max\_{s,a\in D}[\xi^\delta(s,a)] + \mathbb{E}\_{a\sim\pi}[\xi^\delta(s,a)] \leq 0,~~~~\forall s,$ Since $I-\gamma P^\pi$ is non-singular $M$-matrix and the inverse of non-singular $M$-matrix is non-negative (Please see "M-matrix" in Wikipedia), i.e., all elements of $(I - \gamma P^\pi)^{-1}$ are non-negative, then $V_\infty(s) = V^\pi(s) + (I-\gamma P^\pi)^{-1}\cdot$ \{ $- \alpha\Delta\_{EPQ}^{\pi}(s) + \mathbb{E}\_{a\sim\pi}[\xi^\delta(s,a)]$ \}$\leq V^\pi(s),~\forall s$. Thus, we can conclude that $V_\infty$ of EPQ underestimates the true value $V^\pi$. We hope this explanation helps clarify the concept of Theorem 3.1.

Reviewer 36me2024-08-09

Thanks a lot for the clarification, I will raise my score.

Authorsrebuttal2024-08-12

Thank you for your thoughtful response. We will make sure to incorporate the points you mentioned into the paper. If you have any additional questions or need further information, please let us know.

Reviewer eWrQ2024-08-09

thanks for the response

Thanks for the detailed response. I think it makes sense for the new version to include the description posted here for choosing a suitable $\tau$. The authors have addressed my questions satisfactorily. I have raised my score to 7.

Authorsrebuttal2024-08-09

Thank you for the prompt response and valuable feedback. We're glad that our answers have been helpful for clarification. Please let us know if you have any additional questions or need further information.

Authorsrebuttal2024-08-09

Thank you for the prompt response and valuable feedback. We appreciate that our response has helped clarify the issue. Please let us know if you have any additional questions or need further information.

Reviewer Hh4A2024-08-11

Thanks for the clarifications. It would be good if the authors could incorporate my suggestions (e.g., ablation study, early stop issue) when preparing the camera-ready manuscript. The main drawback of this paper is that EPQ introduces many hyperparameters, and one needs to manually find the optimal ones on a new dataset, as commented. However, since existing methods like ReBARC also require significant hyperparameter search to achieve good performance, I think it is okay and believe this paper can be accepted. I would vote for acceptance.

Authorsrebuttal2024-08-12

Thank you for the prompt response. We are glad to hear that our answers have helped in understanding the paper. If you have any additional questions or need further information, please let us know.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC