Optimal Batched Best Arm Identification

We study the batched best arm identification (BBAI) problem, where the learner's goal is to identify the best arm while switching the policy as less as possible. In particular, we aim to find the best arm with probability $1-δ$ for some small constant $δ>0$ while minimizing both the sample complexity (total number of arm pulls) and the batch complexity (total number of batches). We propose the three-batch best arm identification (Tri-BBAI) algorithm, which is the first batched algorithm that achieves the optimal sample complexity in the asymptotic setting (i.e., $δ\rightarrow 0$) and runs only in at most $3$ batches. Based on Tri-BBAI, we further propose the almost optimal batched best arm identification (Opt-BBAI) algorithm, which is the first algorithm that achieves the near-optimal sample and batch complexity in the non-asymptotic setting (i.e., $δ>0$ is arbitrarily fixed), while enjoying the same batch and sample complexity as Tri-BBAI when $δ$ tends to zero. Moreover, in the non-asymptotic setting, the complexity of previous batch algorithms is usually conditioned on the event that the best arm is returned (with a probability of at least $1-δ$), which is potentially unbounded in cases where a sub-optimal arm is returned. In contrast, the complexity of Opt-BBAI does not rely on such an event. This is achieved through a novel procedure that we design for checking whether the best arm is eliminated, which is of independent interest.

Paper

References (45)

Scroll for more · 33 remaining

Similar papers

Peer review

Reviewer Uhz86/10 · confidence 4/52024-07-01

Summary

This paper studies the Batched Best Arm Identification (BBAI) problem in multi-armed bandits. The goal is to design an efficient algorithm that correctly finds the arm with the highest mean with probability $\ge 1 - \delta$, while minimizing: (1) the sample complexity, defined as the total number of arm pulls; (2) the "batch complexity", defined as the number of rounds in which the algorithm requests arm pulls. (Formally, the algorithm specifies the number of times it pulls each arm in the $r$-th round after observing all the outcomes in the previous $r-1$ rounds.) The main results of this work are: - An algorithm (termed Tri-BBAI) that achieves an asymptotically (i.e., as $\delta \to 0^{+}$) optimal sample complexity in three batches in expectation. - Another algorithm (termed Opt-BBAI) that achieves: (1) near-optimal sample and batch complexities in the non-asymptotic setting; (2) the same guarantees as Tri-BBAI in the asymptotic setting.

Strengths

- This work studies a fairly natural problem. While the setup is not new, I liked that the authors explored certain perspectives that are different from prior work, namely: (1) asymptotic optimality of sample complexity; (2) focusing on the expected number of rounds, rather than treating it as a hard constraint. - The results are pretty strong: The algorithms achieve asymptotically optimal sample complexities within a (small) constant number of rounds (in expectation). - The presentation is clear in general. I found the main paper fairly easy to follow.

Weaknesses

I think a major weakness of the work is that its main result is fairly intuitive, and arguably, its proof does not give new insights to the design of pure exploration algorithms. Here is why: For simplicity, suppose that we only have two arms with means $1/2+\epsilon$ and $1/2-\epsilon$, and the parameter $\epsilon > 0$ is *unknown*. In this case, the "natural" algorithm is to make geometrically decreasing guesses on $\epsilon$: $\epsilon_1, \epsilon_2, \ldots$, where each $\epsilon_k = 2^{-k}$. At guess $\epsilon_k$, we would pull each arm $(1/\epsilon_k)^2$ times to get an $O(\epsilon_k)$-approximation of the means. If the means are clearly separated, we get the answer; otherwise, we continue with smaller guesses. Clearly, this approach requires many rounds ($\Omega(\log(1/\epsilon))$ rounds) of adaptivity. On this instance, very roughly speaking, the authors' approach is to start with the guess, say, $\epsilon_k = 1/\log^{1/3}(1/\delta)$. For each fixed $\epsilon$, when $\delta$ is small enough, the guess would be smaller than the actual $\epsilon$, and the algorithm wins the game in $O(1)$ rounds. Also, this first round only takes $(1/\epsilon_k)^2 = \log^{2/3}(1/\delta) \ll \log(1/\delta)$ samples, so this will not affect the asymptotic behavior as $\delta \to 0^{+}$. In general, the Tri-BBAI algorithm uses a round-robin strategy as an inefficient exploration round. The length of this round is chosen as a function of $\delta$, so that the asymptotic behavior is not affected. Therefore, for each fixed instance, there exists some $\delta_0 > 0$ such that whenever $\delta < \delta_0$, this inefficient exploration succeeds with a good probability, and may guide the algorithm to sample in an asymptotically optimal way in the rest. As a result, this asymptotic optimality might only hold for extremely small values of $\delta$: From Line 492, the analysis needs $1/\log\log(1/\delta) = \epsilon \le \Delta_2$ to go through; in other words, $1/\delta$ needs to be **doubly exponential** in $1/\Delta_2$. Admittedly, this weakness was addressed by the other algorithm Opt-BBAI, which achieves a sample complexity bound for finite $\delta$ as well. However, it should be noted that the complexity contains a $\sum_{i=2}^{n}(1/\Delta_i)^2\log n$ term, which could be higher than the optimal sample complexity by a $\log n$ factor. In contrast, the state-of-the-art bounds for (non-batched) BAI (e.g., [Karnin-Koren-Somekh, ICML'13][Jamieson-Malloy-Nowak-Bubeck, COLT'14][Chen-Li, arXiv'15][Chen-Li-Qiao, COLT'17]) are tight up to a doubly-logarithmic factor. Despite the weakness mentioned above, I think this submission presents some solid work and a nice observation (Lines 72-73), namely, the need for many rounds of adaptivity only arises when $\delta$ is small (or moderate), and goes away when we focus on the asymptotic regime. Therefore, I lean towards accepting the paper.

Questions

I don't have specific questions for the authors.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Limitations have been adequately addressed.

Reviewer vkzZ3/10 · confidence 4/52024-07-07

Summary

The paper "Optimal Batched Best Arm Identification" introduces Tri-BBAI and Opt-BBAI algorithms to identify the best arm in multi-armed bandit settings. Tri-BBAI achieves optimal sample complexity with only three batches on average as the confidence parameter $\delta$ approaches zero. Opt-BBAI extends Tri-BBAI for finite-confidence settings, providing near-optimal sample and batch complexities for finite $\delta$.

Strengths

The paper is generally well-written. The theoretical analysis looks solid.

Weaknesses

First, I don't think the regime when $\delta$ approaches $0$ is interesting. In a typical scenario, we want to set the confidence relatively small, and it doesn't make a lot of sense to consider the algorithm's performance on extremely small values. The paper doesn't fairly compare previous results. It mentions the work "Collaborative top distribution identifications with limited interaction," and work "Optimal streaming algorithms for multi-armed bandits", but it doesn't include it in the table, which can lead to the impression that Opt-BBAI is the first algorithm to achieve $O(\log \frac{1}{\Delta_{2}})$ batch complexity. However, many algorithms have already achieved similar results. Moreover, these algorithms are not included in experiments. Significant work should be done to implement a fair comparison and properly acknowledge other works.

Questions

N/A

Rating

3

Confidence

4

Soundness

2

Presentation

2

Contribution

1

Limitations

N/A

Reviewer 3W3Z7/10 · confidence 3/52024-07-13

Summary

This paper considers the problem of BAI in the fixed confidence setting, in the context of finding the best arm in as few batches as possible. That is, instead of observing the reward of each arm pulled in turn, the learner makes multiple pulls in a single batch and only observes all rewards after the completion of said batch. The first result of the paper is to show that it is possible to achieve asymptotically optimal sample complexity, with only a constant number of batches. Specifically the authors propose the Tri-BBAI algorithm, which is shown to have asymptotically optimal sample complexity, PAC guarantees and requiring at most 3 batches in expectation, Theorems 3.1, 3.2, 3.3 respectively. The authors then describe a second algorithm Opt-BBAI which has near optimal finite confidence guarantees as well as being asymptotically optimal.

Strengths

Best arm identification under fixed confidence is well studied problem and running algorithms in batches has clear practical relevance. Thus, achieving both asymptotically optimal sample and constant batch complexity is a nice result.

Weaknesses

When considering finite confidence guarantees the authors could also compare with the recent work "An ε-Best-Arm Identification Algorithm for Fixed-Confidence and Beyond", Jourdan, Degenne, Kaufmann.

Questions

Do the authors have an idea as to why the Tri-BBAI has poor experimental performace against Track and Stop?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

No concern.

Reviewer YCPj5/10 · confidence 2/52024-07-13

Summary

The paper presents two novel algorithms for the batched best arm identification (BBAI) problem. The first is the Tri-BBAI algorithm, which employs three batches with the expectation of achieving the asymptotic optimal sample complexity. Based on Tri-BBAI, the authors conceived the Opt-BBAI algorithm, which achieves the near-optimal sample and batch complexity when δ is finite and it does not tend to 0. Opt-BBAI enjoys the same sample and batch complexity of Tri-BBAI for the asymptotic setting

Strengths

- Tri-BBAI is the first batched algorithm to achieve optimal sample complexity in an asymptotic setting - Opt-BBAI is the first batched algorithm to achieve near-optimal sample and batch complexity in non-asymptotic setting - Both algorithms are supported by valid theoretical analysis - Batched solutions may bring benefits to real-world scenarios that cannot rely on sequential methods, like Track-and-Stop algorithms

Weaknesses

- The paper dedicates a significant portion to the introduction and related works, only beginning to describe methodologies and algorithms on page 5. I suggest reducing the introductory section to allow more space for discussing comparisons with existing algorithms (Table 2) and for elaborating on the experimental results. - The paragraph on the notation is partially useful, since it presents some notations that are never utilized in the main paper, but only in the appendix and may be employed to describe other unclear symbols. - I found some statements a bit misleading. For example, at line 251 the period ”In this section, we introduce Opt-BBAI, which can attain the optimal sample and batch complexity in both asymptotic and non-asymptotic settings adaptively [...]” is in contrast with lines 255-256 where, always regarding Opt-BBAI, is correctly said that ”we can achieve asymptotic optimality and near non-asymptotic optimality adaptively[...]. - The claim at line 714 ”the sample complexity of Tri-BBAI and Opt-BBAI [...] is at most 2.6 times greater than Track and Stop when δ is very small” is not true, since in the second experiment for δ = 1 × 10−10 the ratio is evidently larger.

Questions

- Reward distributions are assumed to belong to a single one-parameter exponential family, a common choice in literature. Are there any limitations in choosing different reward distributions that could affect the effectiveness of the presented approach? - I have some doubts about the necessity of presenting both Tri-BBAI and Opt-BBAI in the paper. If Opt-BBAI achieves the same performances of Tri-BBAI in the asymptotic settings, why is it necessary to describe Tri-BBAI as well, considering that Opt-BBAI also addresses non-asymptotic settings? Could you explain and give some intuition about the need to present Tri-BBAI? - The choice of exactly three batches in the Tri-BBAI algorithm is not entirely clear. Additionally, I would like to ask the authors to provide some intuition regarding the optimal batch complexity of 2 for both Tri-BBAI and Opt-BBAI (shown in Table 1) compared to the structure of Tri-BBAI algorithm which employs at most 3 batches. What is the intuition behind these results? Why was a 2-batch algorithm not considered from the beginning? - Experiments have been carried on crafted environments. Since authors argued about the benefits that their work would bring in real-world settings, I would expect to find an experiment on a such real-world scenario.

Rating

5

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Authorsrebuttal2024-08-12

Dear Reviewer vkzZ, Thanks again for reviewing our paper. As the end of the discussion period approaches, we would like to know whether our responses have addressed your concerns. If there are any additional questions or areas that require clarification, please do not hesitate to let us know. We highly value your perspective and, if you find our responses satisfactory, would be grateful if you would consider raising your score for our paper. Thank you.

Reviewer YCPj2024-08-13

Thanks for the answers to my concerns. They are satisfactory.

Authorsrebuttal2024-08-14

Thank you for your reply. We would like to summarize our responses to your review as follows: 1. You raised concerns that the asymptotic setting (as $\delta$ approaches 0) is not interesting. 2. You noted the omission of two related works (‘Collaborative Top Distribution Identification with Limited Interaction’ and ‘Optimal Streaming Algorithms for Multi-Armed Bandits’) from our comparison table. In our previous response, we addressed these points, which we believe are relatively minor and do not detract from the overall significance of our technical contributions. Specifically: * We demonstrated that the asymptotic setting is not only widely studied in the literature but also holds practical significance (see A1 for more details). * We emphasized our substantial contributions to the non-asymptotic setting (finite $\delta$), as discussed in both our paper and the previous response. * We clarified that 'Optimal Streaming Algorithms for Multi-Armed Bandits' is not designed for the batched bandit setting, and that we have already cited the main results of 'Collaborative Top Distribution Identification with Limited Interaction' in Lines 151 and 152 of our paper. * We committed to add more discussions on these papers and add related results in our comparison table and we have conducted additional experiments in the rebuttal which will be included in the revised manuscript. We hope this summary clarifies our position and the steps we have taken to address your concerns. Best regards, Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC