Replicability in Reinforcement Learning

We initiate the mathematical study of replicability as an algorithmic property in the context of reinforcement learning (RL). We focus on the fundamental setting of discounted tabular MDPs with access to a generative model. Inspired by Impagliazzo et al. [2022], we say that an RL algorithm is replicable if, with high probability, it outputs the exact same policy after two executions on i.i.d. samples drawn from the generator when its internal randomness is the same. We first provide an efficient $ρ$-replicable algorithm for $(\varepsilon, δ)$-optimal policy estimation with sample and time complexity $\widetilde O\left(\frac{N^3\cdot\log(1/δ)}{(1-γ)^5\cdot\varepsilon^2\cdotρ^2}\right)$, where $N$ is the number of state-action pairs. Next, for the subclass of deterministic algorithms, we provide a lower bound of order $Ω\left(\frac{N^3}{(1-γ)^3\cdot\varepsilon^2\cdotρ^2}\right)$. Then, we study a relaxed version of replicability proposed by Kalavasis et al. [2023] called TV indistinguishability. We design a computationally efficient TV indistinguishable algorithm for policy estimation whose sample complexity is $\widetilde O\left(\frac{N^2\cdot\log(1/δ)}{(1-γ)^5\cdot\varepsilon^2\cdotρ^2}\right)$. At the cost of $\exp(N)$ running time, we transform these TV indistinguishable algorithms to $ρ$-replicable ones without increasing their sample complexity. Finally, we introduce the notion of approximate-replicability where we only require that two outputted policies are close under an appropriate statistical divergence (e.g., Renyi) and show an improved sample complexity of $\widetilde O\left(\frac{N\cdot\log(1/δ)}{(1-γ)^5\cdot\varepsilon^2\cdotρ^2}\right)$.

Paper

Similar papers

Peer review

Reviewer 1f6r6/10 · confidence 4/52023-07-01

Summary

This work studies the question of reproducibility in reinforcement learning (RL). They define reproducibility as an algorithm returning the same policy on two different random draws from the environment, with probability at least $\rho$. In the generative model setting, they show that there exists an algorithm which is $\rho$-reproducible, returns an $\epsilon$-optimal policy with probability at least $1-\delta$, and collects at most $O(\frac{N^3 \log 1/\delta}{(1-\gamma)^5 \epsilon^2 \rho^2})$ samples, for $N$ the number of state-actions. They also show a lower bound that this is tight (up to a factor of $1/(1-\gamma)^2$). As the $N^3$ complexity could be prohibitively large in many cases, they show that by relaxing the definition of reproducibility somewhat, they are able to obtain a complexity that scales with $N$ instead.

Strengths

1. To my knowledge, the setting is novel: I do not think the question of reproducibility has been previously considered in the RL literature. Furthermore, the question of algorithmic reproducibility has in general seen attention recently in the broader machine learning literature, so there is general interest in the setting. 2. The results paint a fairly complete picture of the problem, with nearly matching upper and lower bounds. In addition, they show that the $N^3$ dependence can be reduced by considering a slightly relaxed formulation of the problem.

Weaknesses

1. The paper does not clearly motivated why we should care about reproducibility in RL. While reproducibility in science in general is an issue, in RL typically the goal is simply to find a policy that (approximately) maximizes the reward, and we don’t necessarily care whether two runs of the same algorithm produce identical policies or not, as long as they have similar performance. The paper does not give clear justification for why this problem is important, but I think this is necessary given the novelty of it. 2. Definition 2.6 should make clear that the algorithm is able to request which state-actions the $n(\epsilon,\delta)$ are from (if it is indeed able to). As it currently reads, it is ambiguous if the algorithm can request which state-actions they are from, or if it is simply given $n(\epsilon,\delta)$ samples from the generator—the latter option is clearly not possible since if all samples are from the same state-action it is not possible to learn a good policy. 3. Definition 2.9 is also somewhat unclear for a similar reason. My reading is that $\bar{S},\bar{S}’ \sim G$ means that samples from a particular set of state-actions are generated from $G$. Would this exclude the case when all the samples are coming from the state-action? Also, does it exclude an adaptive algorithm which chooses which state-actions to sample from based on the data it has already seen? 4. Using the notation $\bar{r}$ for the randomness of the algorithm is somewhat confusing as $r$ is already used for the reward. 5. I understand that space is limited, but if possible it would be helpful to give some more explanation for the algorithm, in particular the replicable rounding procedure, since this is non-standard in the RL literature. Some intuition on why the $N^3$ dependence arises would also be helpful.

Questions

The primary question I have is on motivating the problem setting, as I stated in the weaknesses section.

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Yes.

Reviewer GbLw7/10 · confidence 4/52023-07-07

Summary

The paper studies replicable reinforcement learning algorithms in the tabular MDP setting with an oracle generative model. The paper gives the first lower-bound for the sample complexity of a $\rho$-replicable $(\varepsilon, \delta)$-optimal algorithm. To obtain this lower-bound, the paper first builds an information-theoretical lower-bound for the sample complexity of estimating $N$ independent possibly biased coins, and then reduces the multiple coin estimation problem to some leveled tabular MDP learning with an oracle generative model. By proposing an algorithm (which is basically running an existing RL algorithm for tabular MDPs with a generative model with some carefully designed precision parameters, and then properly rounding the value function) with $\tilde O(N^3)$ dependency of $N$ in the sample complexity upper-bound, the paper shows that the proposed sample complexity lower-bound is indeed sharp in $N$. The paper also studies a weaker replicability formulation called approximate-replicability and gives an algorithm with $\tilde O(N)$ sample complexity.

Strengths

The paper is clearly written and easy to follow for readers with background on recent theoretical RL works. The given sample complexity lower-bound in $N$ is tight. The proposed algorithms in both formulations are built upon existing oracle algorithms and simple post-processing and hence easy to understand.

Weaknesses

In my opinion, the structure of the paper can be slightly improved. The reductions between the optimal Q-function estimation and the optimal policy, the multiple coin estimation problem and RL on tabular MDPs are at a high level easy to understand. The author may consider compress these contents and move some details of the important technical steps (e.g., the replicable rounding step and how to develop the sample complexity lower-bound for the replicable multiple coin estimation problem) to the main text, in order to bring more insight to the readers after reading the main text.

Questions

Can you provide some insights and comments on the replicability of RL in the pure exploration setting or episodic regret minimization setting (hence both without discounting factors)?

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

The paper considers RL on tabular MDPs with a generative model, it would be more exciting to see results on weaker settings.

Reviewer o1wo6/10 · confidence 3/52023-07-11

Summary

The paper makes a significant contribution by introducing a theoretical study of replicability in reinforcement learning. It focuses on discounted tabular Markov Decision Processes (MDPs) with generative models and explores two definitions of replicability: exact and approximate versions. For the exact version of replicability, the authors provide a lower bound and propose an algorithm that achieves near-optimal performance, with the exception of dependence on the discount factor $\gamma$. In the case of the approximate version of replicability, the authors show an improved sample complexity.

Strengths

- The paper provides a comprehensive study on the theory of replicability, offering insights into both the exact and approximate versions of replicability. Notably, for the exact version, the authors present both upper and lower bounds, which align closely. - The concept of replicability in reinforcement learning is highly intriguing, and this paper serves as a pioneering work in advancing the study of theory of replicability in RL. By introducing and delving into this concept, the authors open up new avenues for exploring replicability and its implications in RL research. - Technically, the paper exhibits strong foundations and analysis, showcasing a solid understanding of the subject matter. The presentation of the research is clear and well-motivated, ensuring that readers can grasp the significance and implications of the findings effectively.

Weaknesses

- While the topic of studying replicability in RL (Reinforcement Learning) is intriguing, I have reservations about its significance. For instance, if there exists a single optimal policy, it logically follows that all RL algorithms should eventually converge to the same policy. Hence, replicability is inherently implied. - I understand that this is primarily a theoretical paper; however, it would be advantageous to include some experimental analysis. Specifically, exploring the behavior of classical RL algorithms in the tabular case with regards to replicability would provide valuable insights.

Questions

How do classic RL algorithms in the tabular case behave concerning the two notions of replicability discussed in the paper?

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

2 fair

Limitations

The authors have addressed the limitations.

Reviewer i28M2/10 · confidence 4/52023-07-20

Summary

Reproducibility is a big problem in RL. This paper builds on Impagliazzo (2022) on replicability in learning, and develops a replicability framework for RL. It focuses on a discounted tabular MDP setting with a generative model. The replicability problem is given shared internal randomness, how many samples are needed for a learning algorithm to output the exact same policy with probability (1-$\rho$) across two executions. It shows that $\rho$-replicability requires $O(N^2/\rho^2)$ more samples than would otherwise be needed for an \epsilon-optimal Q function, where N is the number of state-action pairs. This matches the lower bound which they also provide. There is also extensions to approximately replicable policy estimation.

Strengths

1. The paper presents a nice framework for replicability in RL which is related to the reproducibility problem. 2. The paper is well-written, and it is easy to understand. 3. The results are interesting though not surprising.

Weaknesses

My main issue with the paper is the underlying premise that the reproducibility crisis in ML and RL is due to the replication difficulty owing to randomness. It is not! As the authors themselves remark, a lot of the reproducibility issues arise due to the need for hyper-parameters, code not being aligned with the algorithms presented in the papers, training details, environments not being set up in the same way and results reproducible only on trajectories with specific random seeds, etc. As we all well know, if we want to estimate the mean of a distribution from n samples, we will not get the same estimate from two different experiments. That is why confidence intervals should always be reported: That is the standard way, in some sense of ensuring replicability in statistics. And to get say 95\% confidence intervals, it is fairly easy to calculate the number of samples (or sample trajectories) needed. That calculation I would expect would scale as O(N^2) as well. So the question is what more have we learnt from the results in this paper? I think the results in this paper are essentially confidence interval type calculations just dressed up nicely as $\rho$-replicability.

Questions

Q1. Can the authors through some experimental work justify why their notion of replicability (particularly approximate replicability) is more suitable than reporting say 95% confidence intervals? If not experimental work, could you argue this through a simple example? Q2. Could you elucidate how the replicability framework you introduce can practically help resolve the reproducibility issues in AI/RL?

Rating

2: Strong Reject: For instance, a paper with major technical flaws, and/or poor evaluation, limited impact, poor reproducibility and mostly unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Please explain how the limitations of your framework can be overcome with further work to make it useful for practically addressing the reproducibility issues in AI/RL?

Reviewer 1f6r2023-08-16

Reply to rebuttal

I would like to thank the authors for their response. I will keep my score as is.

Reviewer GbLw2023-08-19

Thanks for the rebuttal

The rebuttal looks good to me. I will keep my score.

Authorsrebuttal2023-08-20

Reviewer i28M

Dear Reviewer i28M, We would be grateful if you could let us know if our explanation made sense to you. We hope the discussion with other reviewers has also helped address any lingering concerns. Thanks a lot!

Reviewer i28M2023-08-21

I am not convinced replicability in the policy space is a worthy goal, or even possible. We can have two different policies with the same value function. And if we care only about policy replicability, we should do imitation learning, perhaps even just behavior cloning. And the problem of replicability from different datasets due to the distribution shift problem, not randomness. Having considered the results in the manuscript, and the authors' response, I find my original score to be too generous. So I am revising it.

Authorsrebuttal2023-08-21

We respectfully disagree with the reviewer's comments. > I am not convinced replicability in the policy space is a worthy goal, or even possible. The interest in the recent line of work on the formal study of replicability in ML initiated by Impagliazzo et al., which studies a similar problem as we do, shows that replicability in the space of classifiers/models and not just utility, is a topic of interest in the ML community. The comment that this goal might not even be possible is factually incorrect, since our results show that it is, indeed, possible. In fact, in the case of approximate replicability, we show that this property can be achieved with poly-logarithmic sample complexity overhead. This discussion also contradicts the comment that our results are not surprising, which was raised by the same reviewer. > We can have two different policies with the same value function. As we explained, in many applications having the same value function is *not* the end goal. > And if we care only about policy replicability, we should do imitation learning, perhaps even just behavior cloning. We are a bit confused about this comment. Our work provides a black-box transformation to achieve replicability in the policy space, no matter what the underlying algorithm we begin with is. We don't see why restricting ourselves to imitation learning and behavior cloning would provide any benefit. In the case of exact replicability, we provide lower bounds which illustrate that the approach the reviewer is suggesting will not lead to improved sample complexity compared to our current results. We underline again that in the case of approximate replicability, our transformation comes, essentially, at no cost in the sample complexity of the algorithm. We would be happy to provide further clarifications if the reviewer can explain what they mean by their comment. > And the problem of replicability from different datasets due to the distribution shift problem, not randomness. This is factually incorrect. Even if the datasets are coming from the same distribution, as we have argued in our work and in our rebuttal, the policies that the algorithm outputs can be significantly different. Of course, the problem of replicability becomes even more difficult in the setting of distribution shifts, but this is beyond the scope of our work. We hope and believe that our work will provide the foundation to study the problem of replicability in more complex environments, like the ones that have distribution shifts.

Reviewer i28M2023-08-21

I respectfully strongly disagree. I think the claims above are a bunch of baloney. First, the replicability problem as defined in this paper has nothing to do with the reproducibility in AI/ML crisis. Second, if the optimal action in any state is not unique, there is no way to "replicate" the policy. If that is not the goal (contrary to above claims now), and , then we are really talking about "replicability" in the value function space, in which case I don't really see what the big deal is! Suppose I want to estimate mean of a distribution in two separate iid experiments. If I want the returned results to be within epsilon, it is pretty straightforward to calculate how many samples one would need. At some level, this is what is being done in this paper. And it is neither surprising, not very useful. I am revising my score further.

Authorsrebuttal2023-08-21

The reviewer is being utterly disrespectful. We refuse to entertain conversations that use unprofessional and derogatory terms for our work, like "baloney". The reviewer should stick to the scientific part of the work and avoid such behaviors. > First, the replicability problem as defined in this paper has nothing to do with the reproducibility in AI/ML crisis The reviewer fails to see the point in this line of work that has led to publications in very well-established conferences like STOC/COLT/ICLR/ICML. > Second, if the optimal action in any state is not unique, there is no way to "replicate" the policy. Again, this is factually incorrect and it shows minimal to zero understanding of our paper. We show that it is indeed possible to do that. > If that is not the goal (contrary to above claims now), and , then we are really talking about "replicability" in the value function space, in which case I don't really see what the big deal is! We are not talking about replicability in the value space, we are talking about replicability in the policy space. We formally prove that this is possible. If the reviewer cannot understand the main results of our paper, it is impossible to have any discussion of scientific merit. > Suppose I want to estimate mean of a distribution in two separate iid experiments. If I want the returned results to be within epsilon, it is pretty straightforward to calculate how many samples one would need. Again, we underline that this is *not* what we are doing. This shows some type of replicability in the value space and not in the policy space. Our work shows that replicability in the policy space is indeed possible, and in the case of the approximate replicability definition, this can be achieved with minimal overhead in the sample complexity. > At some level, this is what is being done in this paper. And it is neither surprising, not very useful. I am revising my score further. The behavior of the reviewer shows that they are acting completely unprofessionally. Adjusting the score based on personal sentiment rather than scientific merit is *not* what the reviewing process is about.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC