An Adaptive Approach for Infinitely Many-armed Bandits under Generalized Rotting Constraints

In this study, we consider the infinitely many-armed bandit problems in a rested rotting setting, where the mean reward of an arm may decrease with each pull, while otherwise, it remains unchanged. We explore two scenarios regarding the rotting of rewards: one in which the cumulative amount of rotting is bounded by $V_T$, referred to as the slow-rotting case, and the other in which the cumulative number of rotting instances is bounded by $S_T$, referred to as the abrupt-rotting case. To address the challenge posed by rotting rewards, we introduce an algorithm that utilizes UCB with an adaptive sliding window, designed to manage the bias and variance trade-off arising due to rotting rewards. Our proposed algorithm achieves tight regret bounds for both slow and abrupt rotting scenarios. Lastly, we demonstrate the performance of our algorithm using numerical experiments.

Paper

Similar papers

Peer review

Reviewer WSKL6/10 · confidence 4/52024-07-01

Summary

This paper studied an extension of multi-armed bandit problem by introducing infinitely many arms and generalized rotting constraints. It provides explicit regret lower bounds and proposes an algorithm with regret upper bound matching the lower bound when $\beta \geq 1$. It has been claimed that closing the gap between the lower and upper bounds when $\beta \in (0,1) $ remains an open problem.

Strengths

- Extending the multi-armed bandit problem to infinitely many arms with rotting mean rewards is feasible in many real-world applications. Solid theoretical analyses and empirical results are presented to justify the proposed solution. - The paper in general is well-written and easy to follow. Since I have not checked the supplementary material step by step, I can not guarantee the correctness of the proofs, but the theoretical results make sense to me.

Weaknesses

- The paper could be viewed as an extension of [1], and the impact of the paper might not be extremely significant. As mentioned in the paper, when $\beta \in (0,1)$, the proposed algorithm can not be proved near optimal at the current stage. The theoretical result could be more impactful if this issue can be solved. [1] Jung-hun Kim, Milan Vojnovic, and Se-Young Yun. Rotting infinitely many-armed bandits. ICML, pages 11229–11254. PMLR, 2022. - When the environment parameters $\beta, V_T,$ and $S_T$ are unknown, a significant amount of additional regret is generated with proposed Algorithm 2. - The paper could benefit from a more extensive experiment study. For example, the algorithm performances can be compared under different rotting processes (perhaps by varying the rotting rate and adding randomness).

Questions

The following questions are raised simply for discussion. There is no need to address them in the paper. - Why the rotting budget $V_T$ and $S_T$ are defined for all selected arms? I would imagine it can fit better with real-world problems to treat the rotting process for each arm independently. Saying if an arm is selected $N$ times, the rotting budget is $f(N)$ where $f$ is a nondecreasing function. - Corresponding to my second point in Weakness, is it possible to address the unknown parameters without incurring large additional regret? Can the methodology in the following paper [2] be applied in this paper to address the nonstationary rotting rewards? Can $\beta$ be estimated by selecting multiple new arms? [2] Chen, Yifang, et al. "A new algorithm for non-stationary contextual bandits: Efficient, optimal and parameter-free." Conference on Learning Theory. PMLR, 2019.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors adequately addressed the limitations.

Reviewer jk3J6/10 · confidence 4/52024-07-02

Summary

This paper considers the extension of the classic stochastic multi-armed bandit problem to the case where there is a) an infinite set of arms and b) rotting of the arm means. Specifically the rotting behaviour is of the 'rested bandit' variety, where the mean reward of an arm may fall as an immediate consequence of playing it, but the mean of an unplayed arm will not change, in contrast to the restless bandit. The infinite set of arms have i.i.d. mean rewards (no structural assumptions on the set of arms or existence of a reward function over them etc.) and reward observations are sub-Gaussian. This paper considers a more general rotting behaviour than Kim et al. (2022, ICML). In Kim et al. (2022) the magnitude of rotting in each round was bounded, in the present paper the more general setting where the total amount of rotting in T rounds is bounded is considered, and the somewhat different setting where the number of rounds in which rotting occurs is considered. These two settings are referred to as the 'slow' and 'abrupt' rotting cases respectively. An important parameter in the infinitely-many-armed bandit problem is $\beta>0$ which controls the probability of an arm being $\delta$-near optimal. All regret bounds in the paper for mean rewards in [0,1] and consider dependence on the number of rounds $T$ and parameter $\beta$. $\beta=1$ corresponds to a uniform distribution on arm means. The paper derives a sliding-window-UCB-based approach, which can be tuned to both the slow and abrupt rotting cases when the bounds on rotting behaviour are known, and an adaptive version which aims to learn the rotting behaviour when these parameters are not known. The paper shows tight (in terms of order) regret guarantees on these algorithms, and improved empirical performance over sensible competitors.

Strengths

The paper studies an interesting extension of prior work to allow for more general rotting constraints. This is likely to be of interest to the multi-armed bandit community and both the methodological and theoretical work present some novelty and careful algorithmic design that merit publication. I was pleased to see a treatment of the case where key problem parameters were not known, and that largely comparable bounds were achievable in this setting. Generally, the paper is written well and concisely, with useful clarifications made around the most important details and a clear relationship to the most closely related prior work.

Weaknesses

There are three main weaknesses which I would like to see commented on the rebuttal phase. I think the first is difficult to address fully in a rebuttal phase, but I would like to be reassured that potential issues do not limit the scope of the contribution. 1. All of the theoretical results are ultimately order results only, without identification of the constants in the appendices, and some being somewhat vaguely described as needing to be 'large enough'. The paper would be stronger if these constants (or upper bounds on them) could be identified, to guarantee for which values of $T$ the bounds are meaningful. It would be helpful if the bounds could thus be realised for the cases considered in the experiments and compared to the actual regret of Algorithms 1 and 2. 2. If I am not mistaken the experiments only consider $\beta=1$? It would be interesting to see if the convergence of the algorithm is replicated for various $\beta \neq 1$. 3. The motivation in terms of real applications is not especially strong. It is difficult to imagine a setting where the precise assumptions of the paper are realistic. In a recommendation setting, is it likely that there will be sufficient information to be confident that the click through rate of an item is always non-increasing, yet there is no contextual information available to supplement the model or exploit similar click through rates across similar items? In the clinical setting, is it likely that the rested assumption is realistic (i.e. any loss in efficacy is only determined by the actions of a single decision maker) alongside the immediacy of feedback and time horizons being of a scale such that the algorithms actually exhibit non-linear regret? If it is the case that the benefit of the paper is more fundamental - it presents an understanding of a simpler problem and its insights serve as a foundation to tackling these more complex real-world challenges, then perhaps it would be more helpful to make that clear?

Questions

1. Can you provide clarity on whether your constants in the regret analysis are identifiable, and suggest the values of these for some simple instance of the problem? How would the resultant bounds compare to the empirical performance of Algorithms 1 and 2? 2. What would the experimental results look like for $\beta \neq 1$? 3. Is Prop 2.3 missing some kind of additional condition, or clarity over what is meant by worst-case in line 453? It would seem to be that you could construct an example where $\rho_1=T+1$ and all other $\rho=0$ and one could achieve $\sqrt{T}$ regret while satisfying $\sum{t=1}^{T-1}\rho_t >T$? 4. Can you put the WUCB definition in an equation display and reference its equation number in the statement of Algorithm 1?

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

Yes

Reviewer ALsC6/10 · confidence 2/52024-07-11

Summary

Authors investigated an adaptive approach for the rotting bandits problem under the infinitely arms assumption.

Strengths

Authors introduced a new ucb like policy for the mentioned problem, additionally a lower bound analysis has been carried on.

Weaknesses

No clear weaknesses

Questions

I wonder how the regret bound behaves with respect to the effective rotting instead of the V_T or S_T. Assuming their values to be large and far from the real rotting, would it be possible to extend this analysis and solution to propose an adaptive result? Additionally, I wonder if also the regret lower bound would be tight in that case. Similarly, I was wondering how good these results would be for the infinitely many afrmed bandit with no rotting.

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

No clear limitations have been found.

Reviewer g8GP4/10 · confidence 4/52024-07-12

Summary

The paper studies infinite-armed bandits with rotting rewards. They show a lower bound on regret in terms of the total-variation and number of abrupt changepoints in the change in rewards. They also provide regret upper bounds: (1) a UCB-like algorithm tuned with knowledge of problem parameters gets optimal regret in some regimes and (2) a parameter-free bandit-over-bandit algorithm can attain optimal regret in some regimes when the level of non-stationarity is large.

Strengths

* The infinite-armed bandit problem is well motivated and the problem and results are clearly presented. * There are matching upper and lower bounds on regret, at least for the better-understood $\beta \geq 1$ setting.

Weaknesses

* The definition of $V_T,S_T$ are a bit unclear and may not be fully rigorous. They are bounds on the total amount and count of rotting, but the rotting $\rho_t$ depends on the chosen arm $\rho_t$ and is thus random (in fact, determined by an adaptive adversary according to Assumption 2.1). But, meanwhile, $V_T,S_T$ seem to be treated as deterministic constants in the whole paper given that they are, for example, used to tune Algorithm 1 to get the best regret bounds. In fact, it would not even make sense for $V_T,S_T$ to _appear_ in the expected regret bounds if they were random. So, the only way the results of this work can rigorously hold is if $V_T,S_T$ bound the respective random quantities $\sum_{t=1}^{T_1} \rho_t$ and $1 + \sum_{t=1}^{T-1} 1\{\rho_t\neq 0\}$ for all realizations of the randomness, which then means there are some missing assumptions for the results of this paper. For instance, even for a "nice" algorithm, $\sum_{t=1}^{T_1} \rho_t$ could be very large with some probability in which case $V_T$ is forced to be large as well because of a "bad realization". The only other way around this is to make further assumptions about the adversary/design of non-stationarity. In either case, there should be more discussion as the rigor of results presented versus the given assumptions seems unclear to me. * There are no optimal regret upper bounds without parameter knowledge. Also, the bandit-over-bandit result for parameter-free result requires further assumptions on the non-stationarity and so does not apply as generally as the other results of this paper.

Questions

Please refer to strengths and weaknesses above.

Rating

4

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

No broader impact concerns.

Authorsrebuttal2024-08-14

Official Comment by Authors

Dear Reviewer g8GP, Thank you again for taking the time to review our paper. We sincerely hope our responses have adequately addressed your questions and comments. If they have, we appreciate it if you could reconsider your evaluation. If you have any last-minute questions, please let us know. Sincerely, Authors

Reviewer g8GP2024-08-14

* **On $V_T/S_T$**: thank you for the clarification. It seems one assumes the adversary is not fully adaptive then and there are limitations on how much rotting it can force on the rewards. I would say wording this as a "adaptive adversary" can be a bit misleading then as it is really a constrained adversary. * **optimal regret upper bounds without parameter knowledge**: I would carefully explain in the presentation why the further Assumptions A.10 is needed. It seems like it is an artifact of the bandit-over-bandit approach as you cannot have the adversary respond in an overly strong way to the master's choice of base over $H$ rounds. Even under the broader above mentioned constrained adversary there seem to be no parameter-free results without this A.10. As such, I feel the scope of result remains limited in this work even while being the first to study this particular setting/parametrization.

Authorsrebuttal2024-08-14

Official Comment by Authors

Thank you for your comment. We would like to address a few points regarding your comments: **On $V_T$, $S_T$:** Based on your comment, we believe your initial concern regarding $V_T$ and $S_T$ has been resolved. However, we would like to provide additional explanation for clarity. We strongly believe that the adaptive adversary under the slow ($V_T$) or abrupt rotting constraint ($S_T$) is a natural and general assumption in our adversary rotting scenario. The adaptive adversary determines an arbitrary rotting rate at each time, immediately after the agent's action is determined, under either $V_T$ or $S_T$. From the values of $S_T$ and $V_T$, we can appropriately quantify the difficulty of our problems, as demonstrated by our regret lower bounds. Without such constraints (i.e., fully adaptive according to your comment), the adaptive adversary could easily make the problem trivial in the 'worst-case' scenario of our setting because whenever an algorithm finds a good arm, the adversary could cause that arm to rot and become a bad arm or even have a negative mean reward adaptively, without the constraints. We also note that similar quantities of $V_T$ and $S_T$ have also been considered in standard nonstationary bandit literature [7,19,4], where they are treated as determined quantities, not random variables, as in our setting. If necessary, we are more than happy to clarify this further in our final version to prevent any potential confusion. **Assumption A.10 for parameter-free:** The additional constraint in Assumption A.10 is required due to our adaptive adversary, which selects the rotting rate arbitrarily and adaptively in response to the selected action at each time, within the bandit-over-bandit framework. If we consider an oblivious adversary instead of an adaptive one, where the values of rotting rates $\rho_t$ are predetermined such that $\sum_t \rho_t\le V_T$ and $1+\sum_t 1(\rho_t\neq 0)\le S_T$ before the game begins, then this satisfies Assumption A.10 with $\rho_t=\varrho_t$. In other words, as we mentioned in Remark A.11, Assumption A.10 is a more general assumption than this oblivious adversary and that in [13]. As mentioned in lines 831~834, the well-known black-box framework proposed for addressing nonstationarity [25] is not applicable to this problem, and attaining the optimal regret bound under a parameter-free algorithm for all ranges of $V_T$ and $S_T$ remains an open problem. However, we respectfully disagree with your comment that the results remain limited. We again highlight that our study is the first work to examine $V_T$, $S_T$, and $\beta$ in the context of rotting bandits with infinite arms, which is fundamentally different from finite-armed bandits (details are described in lines 58–70). Notably, we achieve tight results through a novel approach when the parameters are known, and we establish regret lower bounds. We also examine the case of unknown parameters. We believe our work is crucial for the community.

Reviewer WSKL2024-08-08

Response to rebuttal

Thank you to the authors for their reply. I am writing to confirm that I've read the author's rebuttal and my scores remain unchanged.

Authorsrebuttal2024-08-08

Thank you for your comment.

Thank you for maintaining the positive rating of our paper. We will incorporate the discussion, including additional experimental results, into our final version. If there are any other questions, please let us know.

Reviewer jk3J2024-08-10

Response to Rebuttal

Thank you for your detailed reply to my comments. I appreciate the commitment to make modifications and am in agreement with your response. I have also read the other reviews and your responses and consider these to be suitably well addressed. As such I will retain my score and increase my confidence. (Justification: I agree with the remarks that finding results without constants and algorithms that can lead to further application-specific work are important fundamentals, but I feel for a score of 7+ the paper would have made more progress in one of these areas.)

Authorsrebuttal2024-08-10

Thank you for your comment.

Thank you for maintaining a positive rating on our paper and increasing your confidence score. We appreciate your detailed feedback. We will incorporate the discussion into our final version, including more details regarding constant factors, a detailed explanation of Prop 2.3, and additional experimental results.

Reviewer ALsC2024-08-12

Thank you to the authors for their reply. I am writing to confirm that I've read the author's rebuttal and my scores remain unchanged.

Authorsrebuttal2024-08-12

Thank you for your comment

Thank you for maintaining a positive rating on our paper. We will incorporate the discussion into our final version.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC