Restless Bandits with Average Reward: Breaking the Uniform Global Attractor Assumption

We study the infinite-horizon restless bandit problem with the average reward criterion, in both discrete-time and continuous-time settings. A fundamental goal is to efficiently compute policies that achieve a diminishing optimality gap as the number of arms, $N$, grows large. Existing results on asymptotic optimality all rely on the uniform global attractor property (UGAP), a complex and challenging-to-verify assumption. In this paper, we propose a general, simulation-based framework, Follow-the-Virtual-Advice, that converts any single-armed policy into a policy for the original $N$-armed problem. This is done by simulating the single-armed policy on each arm and carefully steering the real state towards the simulated state. Our framework can be instantiated to produce a policy with an $O(1/\sqrt{N})$ optimality gap. In the discrete-time setting, our result holds under a simpler synchronization assumption, which covers some problem instances that violate UGAP. More notably, in the continuous-time setting, we do not require \emph{any} additional assumptions beyond the standard unichain condition. In both settings, our work is the first asymptotic optimality result that does not require UGAP.

Paper

Similar papers

Peer review

Reviewer TTwr4/10 · confidence 3/52023-07-03

Summary

This paper studies the infinite-horizon restless bandits problem and proposed a simulation-based framework, i.e., Follow-the-Virtual-Advice, to leverage single armed policy to solve a multi-armed problem, which gets rid of the difficult-to-verify condition, i.e., the uniform global attractor property. Removing this pre-condition is a significant improvement for RMABs.

Strengths

This paper studies the infinite-horizon restless bandits problem and proposed a simulation-based framework, i.e., Follow-the-Virtual-Advice, to leverage single armed policy to solve a multi-armed problem, which gets rid of the difficult-to-verify condition, i.e., the uniform global attractor property. Removing this pre-condition is a significant improvement for RMABs.

Weaknesses

1. The arguments on FTVA in Section 3.3 look to be a bit hands-waving since there is no clear evidence to support their claims. Based on the reviewer’s understandings, those arguments are possibly wrong. For example, in lines 181-184, the authors claim that even if the initial virtual state and real state of an arm are different, they will become identical in finite time by chance under mild assumptions in Section 4.1. However, Assumption 1 relies on the “observation” in lines 233-238, which is not true. All the transitions are stochastic, not deterministic. How could we guarantee that $S(t+1) =\hat{S}(t+1)$ if $S(t)$ is different with $\hat{S}(t)$? If Assumption 1 fails, the argument in lines 183-184 does not hold. The arguments in lines 185-189 are also not true. Even when the virtual process can always satisfy the budget constraint, how do we guarantee that the real process following the virtual actions does not violate the budget constraints. 2. The example in lines 194-215 is also not fully supported by evidence. The single-agent policy defined in eq. (8) is stationary stochastic policy, and agent selects actions according to certain probabilities. What do you mean by the preferred action at each state? Though this particular example may only have one action at each state, there is no preferred action in general. How to guarantee the descriptions in lines 207-209 to be true? This example shows that the real and virtual process has different state distributions, which contradicts Section 3.3. See comment 1. 3. A very important related work is missing. [Ghosh22] also gets rid of the global attractor assumption and considers a much more challenging setting with heterogeneous arms and multi-actions. An outstanding limitation of this work is that all arms must be the same. Ghosh, A., Nagaraj, D., Jain, M., & Tambe, M. (2022). Indexability is Not Enough for Whittle: Improved, Near-Optimal Algorithms for Restless Bandits. arXiv preprint arXiv:2211.00112. 4. The references cited in this paper are not precise. For example, as far as the reviewer knows, [Ver16] does not consider the single-armed problem. Hence, the argument in lines 140-141 are not precise. Another example is in lines 123-124. The reviewer is not aware of the meaning of these sentences. How are they related to the RMAB problem in eqs. (1)-(2)?

Questions

See comments in Weaknesses.

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

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

2 fair

Contribution

2 fair

Limitations

no negative societal impact.

Reviewer S8QK7/10 · confidence 4/52023-07-05

Summary

This paper presents an algorithm providing an asymptotically optimal policy to solve restless bandits with N arms when N goes to infinity without using classical assumptions on the expected deterministic dynamics (uniform global attractor).

Strengths

1. The idea to couple the actual trajectory under the relaxed policy with a virtual one, sampled independently for each bandit is a very nice idea. 2. Removing the UGAP assumption is also a strong point, although this is not the only paper going in this direction (see for example Zhang X, Frazier PI (2022) Near-optimality for infinite-horizon restless bandits with many arms. arXiv preprint arXiv:2203.15853). 3. The paper makes a good literature review and a good use of previous work in this domain.

Weaknesses

1. The discussion on the synchronization assumption, and especially the sufficient conditions could be made more precise. The link with UGAP is also missing. Also there is a gap between the synchronization of each arm with the synchronization of the whole bandit (see for example the PhD thesis of Kimang Khun). 2. The fact the algorithm is based on independent sampling of each arm makes the \sqrt{N} bound tight, wuth little hope for improvement, while other approaches based on the fluid dynamics could still work with UGAP (using Cezaro averaging for example) with a higher chance to have a better convergence rate. 3. The comparison with previous algorithms is a little unfair, especially because of initial conditions.

Questions

1. Can you change the initial conditions of the numerical experiments to make a fairer comparisons with existing solutions. 2. Provide answers to the weak points listed above.

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

4 excellent

Limitations

I did not see any limitations.

Reviewer s6TA6/10 · confidence 3/52023-07-05

Summary

In the late 80s and early 90s, Whittle formalized the restless multi-armed bandit framework and Weber & Weiss proved that Whittle’s index policy, which solves a relaxed version of the optimization problem, is asymptotically optimal under conditions of indexability. Since then, RMAB problems have attracted much attention for their wide application in diverse areas. This paper contributes a new technique that is not an index policy. Instead, the simulation framework requires a new “synchronization assumption” for the discretized time case. The contributions of this paper are purely theoretical, including a new algorithmic approach (Follow-the-Virtual-Advice) and optimality gap bounds.

Strengths

This paper introduces an original technique for solving two-action RMAB problems in both discretized and continuous time spaces, and provides bounds on conversion loss and gaps in optimality for both. It is clear that the only assumption that must be enforced in the discrete case is the new “synchronization assumption”. There is a nice discussion about this assumption and when it holds in the appendix for novice readers. The submission appears to be technically sound, and claims are well supported by theoretical analysis.

Weaknesses

After decades of essentially little-to-no change in the state-of-the-art techniques that are deployed in practice, it is wonderful to see new approaches. The authors should be made aware of [1], which aims to address the same problem and also does so in a new, novel way. I am surprised by the use of “uniform global attractor property” rather than “indexability” throughout the paper. With regards to clarity, I encourage the authors to include more explanations to inform the reader. The contributions of this paper are purely theoretical. Open-source code as well as empirical results comparing FTVA against other state-of-the-art implementations would be helpful for the research community. [1] Abheek Ghosh, Dheeraj Nagaraj, Manish Jain, and Milind Tambe. 2023. Indexability is Not Enough for Whittle: Improved, Near-Optimal Algorithms for Restless Bandits. In Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems (AAMAS '23). International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC, 1294–1302.

Questions

Are the authors aware of [1]? That paper also claims to contribute a novel new approach for the discrete-time problem. Not only do they consider average-reward, but also any discounted total expected reward in both the finite and infinite horizon case. How does your approach differ? How do your optimality guarantees differ? [1] Abheek Ghosh, Dheeraj Nagaraj, Manish Jain, and Milind Tambe. 2023. Indexability is Not Enough for Whittle: Improved, Near-Optimal Algorithms for Restless Bandits. In Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems (AAMAS '23). International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC, 1294–1302.

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

4 excellent

Limitations

Yes.

Reviewer TkxD6/10 · confidence 4/52023-07-05

Summary

This paper studied, for both the discrete-time and continuous-time settings, the average-reward restless bandit problem without the uniform global attractor property. They proposed the simulation-based policies Follow-the-Virtual-Advice (FTVA) and FTVA-CT, which converts the optimal single-armed policy into the N-armed constrained policy by simulating a virtual process that follows the single-armed policy, and then by letting the real process act closely to the virtual process but satisfying the constraints. The policies provide vanishing performance loss.

Strengths

Novelties: In this reviewer's knowledge, this paper is the first paper to provide an algorithm for the average reward constraint RB problem without uniform global attractor property, which is a property that is hard to verify. The proposed FTVA policy is quite straightforward and also easy to understand. The idea of using virtual processes in this problem seems original to me. Presentation: The paper is well-written, with precise language and clear structure. The paper is clear and easy to understand. And there are a lot of intuitive explanations. Significance: The paper provides the first algorithm for the average reward constraint RB problem without uniform global attractor property.

Weaknesses

The following assumption is very strong and important to this policy: the author assumed that all the arms with the same state and action have the same reward. And this policy seems very hard to extend to the case where the arms have different reward functions. (More comments in Questions.) The paper feels incomplete, which prevents a clear conclusion at the end. The simulation is not sufficient, the author only showed the average reward comparison for different policies, but I would also want to know how is the action look like for different policies or is there any intuition in it. The paper presentation is too high-level and has missing details. The author claimed at first, but didn’t explain later, how the FTVA policy can be implemented in a distributed manner. The author claimed the algorithm can run at a linear computational cost but didn’t explain how, later.

Questions

In the setup, the author assumed that all the arms with the same state and action have the same reward. With this assumption, the problem (with the \alpha N constraint) is greatly simplified. This assumption makes me feel that even with the \alpha N constraint, those arms are still acting very independently, not affecting each other too much. And this policy seems very hard to extend to the case where the arms have different reward functions. If possible, instead of only the average reward comparison, I am also curious about how the FTVA action policy look like compared to other non-asymptotically optimal policies in the example. And is there any significant different actions on some special states that makes the policy be optimal? In the beginning, the author said that FTVA policy can be implemented in a distributed manner for arms, could you add more about it after the algorithm descriptions? Figure1 is missing one label.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

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

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

There is no negative societal impact. Discussion on limitations is not extensive.

Reviewer k12y8/10 · confidence 2/52023-07-07

Summary

This work studies the restless bandit problem. In this setting the agent controls a set of $N$ arms modelled as Markov Decision Processes with two actions (active, inactive). The MDPs are coupled with a constraint that forces the agent to maintain a fixed arm activation rate. The authors propose a novel algorithm for this problem that does not use the commonly used uniform global attractor property (UGAP). The key insight of the approach is to transform a learned single arm policy to a $N$ armed policy. While this may lead to a discrepancy between the learned policy and what is actually played because of the constraint, the proposed algorithm aims to bring these two states closer together. The authors provide convergence guarantees and rates for both discrete and continuous time variants of the algorithm.

Strengths

To the best of my knowledge, the proposed algorithm is novel. The authors also presented very clearly the setup of the problem which is especially important given that many members of the community may not be familiar. While I did not follow all the technical details in this work, the synchronization assumption seems like a significantly weaker assumption than UGAP, indicating this a very strong contribution. On top of that, the way the algorithm solves the problem is itself highly non trivial.

Weaknesses

I think clarifying a little bit more the importance of synchronization when we are using the constraint would be very helpful. I will try to outline my understanding in case it is helpful in improving the presentation: * In Section 4.1, based on the stated assumptions it is clear that achieving synchronization once is sufficient to achieve synchronization forever. But for the restless bandit setting it seems that the real and virtual states synchronize and desynchronize in a loop. Indeed the states start synchronized and some arms may immediately get desynchronized (become bad in the terminology of Theorem 1). * I understand that the the SA assumption helps us understand desynchronization is resolved in finite time. But if synchronization lasts only a single round, then the arms will still be bad most of the time. So SA alone is not sufficient. * Based on the proof sketch of Theorem 1, it seems like desynchronization is inherently rare for large $N$ so this is why SA is sufficient. Intuitively, SA is only needed to bound the effect of a single good-bad-good arm loop. If my understanding is indeed correct, it would be helpful to make this more explicit somewhere in the text, preferably in the context of the example of Section 3.3.

Questions

See above for some suggestions.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed 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

All limitations are addressed.

Reviewer JXov6/10 · confidence 2/52023-07-27

Summary

This paper studies the problem of optimizing a policy for restless bandits in the infinite horizon case. A new algorithm is proposed that can convert a single-arm policy into an N-arm policy using virtual trajectories of the states and actions of the single-arm policy. The main idea is to generate virtual actions using the single-arm policy and then attempt to match these virtual actions with real actions that match the budget constraint. The algorithm is shown to have suboptimality that degrades with N. Both discrete time and continuous time versions are given. One of the main contributions is that this paper forgoes the UGAP condition. This is an emergency review. I am not familiar enough with this literature to definitively judge the significance and novelty.

Strengths

The paper appears to tackle an important problem in the literature broadly (restless bandits) and provides a potential solution to optimize policies with vanishing suboptimality as N increases. Achieving this without the UGAP condition that appears in prior work is interesting. Overall the paper is well written and generally easy to understand. The examples and organization of the paper help convey these results.

Weaknesses

The main weakness appears to be that the entire premise hinges on the synchronization assumption that the real states and virtual states will eventually match, if the same actions are played, after which point they will remain identical. This seems reasonable in some settings, but a more clear contrast with conditions in other papers would be helpful. In general, it would be helpful to better understand the significance of this change of assumptions and how this might translate into practice.

Questions

It is mentioned that Algorithm 1 should “steer” the trajectories towards the virtual trajectories but still respect the budget constraints. However, in Line 4, the selection of the smaller set to respect the budget does not appear to filter arms in any particular way. What prevents the same arms from being activated while starving others and thus causing the virtual trajectory to differ from the real one by a lot? Suggestion: the paper could benefit from further discussion of the last theorem and a conclusion.

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

4 excellent

Contribution

3 good

Limitations

No negative societal impact, but it would be helpful to have more discussion of limitations.

Area Chair qMx22023-08-16

Contrasts between discounted and average reward setting

The authors have claimed that the related work the reviewer mentions is for a different technical setting. I can corroborate this statement. Therefore, can the reviewer please consider the contribution of this work in light of these developments?

Reviewer JXov2023-08-19

Reply

Thank you for the thorough response. I will update my score.

Reviewer TkxD2023-08-19

I thank the authors for a thorough and satisfactory response. I will update my rating.

Reviewer k12y2023-08-19

Thanks for the clarification

I would like to thank the authors for the clarification!

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC