An $\varepsilon$-Best-Arm Identification Algorithm for Fixed-Confidence and Beyond

We propose EB-TC$\varepsilon$, a novel sampling rule for $\varepsilon$-best arm identification in stochastic bandits. It is the first instance of Top Two algorithm analyzed for approximate best arm identification. EB-TC$\varepsilon$ is an *anytime* sampling rule that can therefore be employed without modification for fixed confidence or fixed budget identification (without prior knowledge of the budget). We provide three types of theoretical guarantees for EB-TC$\varepsilon$. First, we prove bounds on its expected sample complexity in the fixed confidence setting, notably showing its asymptotic optimality in combination with an adaptive tuning of its exploration parameter. We complement these findings with upper bounds on its probability of error at any time and for any error parameter, which further yield upper bounds on its simple regret at any time. Finally, we show through numerical simulations that EB-TC$\varepsilon$ performs favorably compared to existing algorithms, in different settings.

Paper

Similar papers

Peer review

Reviewer kKhf6/10 · confidence 3/52023-07-04

Summary

The paper gives a new epsilon-best arm identification algorithm which is simple and anytime, just amounting to balancing between the empirical best arm and roughly speaking, the arm with the best chance to be superior to it by epsilon. Its performance is analyzed mainly for fixed confidence, and it also attains decent (not particularly optimal) bounds for fixed budget. The results assume a unique best arm but allow repeated mean values among the other arms.

Strengths

The fixed confidence bounds are very good (sharp for Gaussians) and the algorithm is nice and simple. The epsilon parameter can be mismatched, though this degrades the bounds a bit.

Weaknesses

No new minimax regret bounds are obtained. As I understand it the paper is basically giving a simpler algorithm with some nice guarantees. Which is fine but not the most revolutionary.

Questions

Maybe ``\beta-optimal'' could be defined more carefully, and the ratio bound of T_{\eps} vs T_{\eps,\beta} could be stated near the definitions. Just before line 200, it would be nice to cite a few of these other works on Top Two algorithms.

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

N/A

Reviewer PH2Y6/10 · confidence 3/52023-07-06

Summary

This work proposed the EB-TC algorithm for $\varepsilon$-best arm identification in stochastic bandits. EB-TC is a BAI algorithm that does not require the knowledge of confidence or horizon, and hence it can be applied in both fixed-confidence and fixed-budget settings. Beyond theoretical guarantees, this work also provided numerical experiments to evaluate the performance of EB-TC. ========================== Thanks for the response. I would like to keep the score.

Strengths

1. The paper is generally easy to follow. 2. EB-TC seems to be a nice BAI algorithm that does not need to know confidence or horizon as priori knowledge, which can be applied in both fixed-confidence and fixed-budget settings. 3. EB-TC can be viewed as a generalization of the Top Two Sampling rule which was designed in an earlier work. 4. The paper provides a clear comparison to existing works.

Weaknesses

As discussed in Section 2.1, the EB-TC algorithm may not perform well to identify the optimal arm. Can this issue be solved?

Questions

As mentioned in the Weakness section: what is the key obstacle such that EB-TC cannot work well for the identification of the optimal arm? I appreciate some explanation of this point. A (possibly) suboptimal theoretical bound is even better.

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

NA

Reviewer KTru7/10 · confidence 2/52023-07-06

Summary

This paper proposes an anytime $\epsilon$-best-arm identification algorithm for multi-armed bandits that can be employed in both fixed-confidence and fixed-budget setting. The algorithm is analyzed in asymptotic fixed-confidence, non-asymptotic fixed-confidence and fixed-budget settings. Meanwhile, the paper also provides empirical results in many different settings.

Strengths

- The proposed algorithm is anytime in nature and thus widely applicable. - The analysis for fixed $\beta$ is thorough and covers all asymptotic fixed-confidence, non-asymptotic fixed-confidence and fixed-budget settings. - The empirical experiments are extensive.

Weaknesses

- The analysis for the IDS variant is limited. - This may be just due to my misunderstanding, but it seems the expected stopping time of the proposed algorithm with finite confidence $\delta>0$ will explode for BAI ($\epsilon=0$) problem. ### Suggestions on Writing - It may be better to also give the formula of $\bar{\beta}_{n+1}(i, j)$ under the IDS variant in the algorithm summary. - Do we simply have $T_{\epsilon_0, \beta}(\mu)=\min_{i\in\mathcal{I}\_{\epsilon_0}(\mu)}T_{\epsilon_0, \beta}(\mu, i)$ in the statement of Theorem 1? It seems this quantity is not formally defined in the main paper. - It seems the statement of Theorem 1 does not mention that the result for IDS variant only holds for Gaussian distribution. - If possible, it may be better to present results in the main paper more in terms of big-$\mathcal{O}$ instead of exact formula.

Questions

- How does the proposed algorithm perform, compared with uniform sampling, SH and SR, in fixed-budget BAI ($\epsilon=0$) problem? - Based on Theorem 1, the proposed algorithm is asymptotically optimal for 0-BAI when taking $\epsilon_0=0$. However, based on Theorem 2, it seems the expected stopping time just explode when taking $\epsilon_0=0$ with positive $\delta>0$. Is this a contradiction? Did I misunderstand anything? How does the algorithm perform in 0-BAI with finite $\delta>0$?

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

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

4 excellent

Presentation

2 fair

Contribution

4 excellent

Limitations

The limitations are addressed well in this paper.

Reviewer 6z6m7/10 · confidence 3/52023-07-13

Summary

The paper considers the epsilon-BAI problem in the multi-armed bandit setting with sub-gaussian arm distributions. It proposes a top2 algorithm for the problem that works across different performace criteria: fixed-confidence, fixed-budget, as well as simple-regret. The proposed algorithm is asymptotically optimal (as delta-> 0) for the fixed confidence setting for gaussian bandits, where delta is the bound on probability of errors. The paper also provides finite delta analysis of the upper bound on the sample complexity in the fixed confidence setting. Another novel feature of the proposed algorithm is that it can be used in the fixed-budget setting, without a prior knowledge of the budget. The theoretical results are accompanied with extensive simulations and numerical studies.

Strengths

I believe that the results of this work are novel. The proposed sampling rule, when coupled with either stopping rule or recommendation rule (or both) works across different BAI settings: fixed-confidence, fixed-budget, simple regret. The paper also discusses well and brings out the comparison with existing works and techniques for each of these problems. I especially like the discussions that are included after each result. I believe that the literature review is also thorough. The extensive numerical studies undertaken definitely add value to the theoretical work.

Weaknesses

I believe that the paper is dense and too long for a conference submission with many details in the appendix. But this may be because of the space constraints. Plots in Figures 6, 10, and 12 are very light. Also the legends are too small to read for most of the plots. In general, it will help to spread out the plots for readability. Since the plots are already in appendix, space isn't an issue.

Questions

1. Could you clarify the definition of simple regret in the epsilon-BAI framework? In line 47, should there also be an indicator on chosen arm not being epsilon-best? In line 97-98, is Delta_{\hat{i}_n} missing in the integration? 2. The lower bound in Lemma 1 is specifically for Gaussian bandits. How does it compare with sub-Gaussian bandits? Is there some monotonicity? It will be good to add a discussion around this. If not, including an example (if known) could help. 3. In Figure 1 (EP-TC_eps algorithm with fixed or IDS) line 3, it will be good to refer to the equation for updating \bar{beta}_{n+1}(i,j) in the main text. 4. It will be helpful if the discussion in lines 150-151 could be elaborated. What are the K(K-1) different tracking rules? Are these the different choices of beta that are maintained for each leader/challenger pair? 5. How should one choose epsilon_0 in practice? 6. In lines 201-203, is that a requirement for asymp. optimality, or is it a requirement for proofs to work? Can one show that if number of optimal arms is greater than 1, then the algorithm doesn't converge (may be b/c of some discontinuity)? 7. Lines 235-237: Would using any-delta methods to construct UCB and LCB indexes in LUCB help in this regard? 8. In Lemma 9, Line 530, could you clarify what j is in mu_eps(i)_j ? 9. Is Lemma 12 a standard result? Adding a citation or proof would help. 10. What are the challenges in generalizing the proofs to beyond gaussian arms to exponential families? What specific properties of Gaussians are used in analysis that may not hold for exponential families? Can one pin down to the level that if we can prove this, this, and this about KL (for example) or other functions in exponential families, then we should get the bound? Minor: 1. Line 207 satisfies these ... 2. Line 587: eeror --> error

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

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

Na

Reviewer B3zV5/10 · confidence 2/52023-07-14

Summary

In this paper, the authors proposed EB-TC_epsilon, a novel sampling rule for epsilon-best arm identification. While the proposed algorithm is analyzed for approximate best arm identification, it is also an anytime sampling rule that can be used for fixed confidence, fixed budget, or anytime settings. The algorithm is analyzed for those settings and is empirically evaluated with numerical simulations.

Strengths

1. The proposed algorithm is simple and easy to implement. Without modification, it works for approximate best arm identification, fixed confidence, fixed budget, or anytime settings. 2. The proposed algorithm is theoretically analyzed and empirically evaluated with numerical simulations. The results show that the proposed algorithm can achieve competitive performance compared to other methods.

Weaknesses

1. The presentation of this paper can be improved. I found it hard to follow this paper in its current form, and I can not discuss its contribution with confidence. (a) The main results and theoretical discussions are mixed with related works. While it is nice to relate the proposed algorithms with previous works, it is hard to see the difference/improvement without clearly presenting previous results -- many of the previous algorithms are only cited by name and without further explanation. (b) Many of the high-level discussions and insights are also mixed with mathematical derivations, which makes it hard to follow. It would be nice to focus more on the main results and insights and present the details in the appendix. 2. The algorithm presentation is not well organized. The algorithm is introduced in line 120 and a crucial quantity $\bar \beta_{n+1}$ is not defined until line 140. It would be nice to define all the quantities before or when introducing the algorithm.

Questions

What is the motivation for proposing one algorithm that works for approximate best arm identification and works, without modification, for fixed confidence, fixed budget, or anytime settings? There are other algorithms proposed specifically for different settings, and one might choose an algorithm based on the setting. It would be nice to see some discussion on this.

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

3 good

Limitations

Yes.

Reviewer KTru2023-08-11

Response

Thank you very much for your rebuttal and most of my concerns have been well-addressed! As a follow-up question, it seems there is a hidden barrier between $\epsilon$-BAI problem and BAI problem such that the results of $\epsilon$-BAI problem cannot be naturally applied to BAI problem by simply setting $\epsilon=0$, which sounds quite counter-intuitive. Is that possible to explain this phenomenon from a high-level perspective? Do you think this hidden barrier between $\epsilon$-BAI and BAI is fundamental? Thanks!

Authorsrebuttal2023-08-11

In the asymptotic fixed-confidence regime and for a fixed $\beta$, EB-TC$_\epsilon$ permits a smooth interpolation between the $\varepsilon$-BAI and the BAI problem, as the algorithm is asymptotically $\beta$-optimal in both cases, provided that all means are distincts (Theorem 1 and Appendix H in [21]). However, in non asymptotic regimes, we believe that there might indeed be some barrier between BAI and $\epsilon$-BAI. In the finite-confidence regime, in both $\varepsilon$-BAI and even in BAI, it is still an open problem to derive a tight lower bound in the finite-confidence regime. Recent works [7, 36] have shown that the sample complexity in BAI is affected by strong moderate confidence terms (independent of $\delta$). While similar terms probably have their counterparts in $\varepsilon$-BAI, our intuition suggests that having $\varepsilon > 0$ may act as a regularizer preventing those terms to ``explode’’ (as they would have upper bounds scaling in $O(K/\epsilon^2)$).

Reviewer 6z6m2023-08-11

Response to rebuttal by authors

Thank you for your response. I acknowledge reading the entire thread of reviews and corresponding rebuttals. For now, I don't have further questions and I'll keep my score unchanged.

Reviewer kKhf2023-08-14

Thanks for the reply. By "minimax" I just meant worst-case rather than Bayesian, sorry for the confusion. I agree that the contribution is nice and worthwhile and that simplicity is nice to have. My intent was just to clarify that if this were a super simple algorithm which also beat the state of the art significantly, it would merit an even higher score. I'll keep my score unchanged for now and don't have further questions.

Reviewer B3zV2023-08-20

Response

Thank you for the response and the clarification on the motivation. I have raised my score to 5. Please consider adding this discussion on motivation in future revisions.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC