Multi-Fidelity Multi-Armed Bandits Revisited

We study the multi-fidelity multi-armed bandit (MF-MAB), an extension of the canonical multi-armed bandit (MAB) problem. MF-MAB allows each arm to be pulled with different costs (fidelities) and observation accuracy. We study both the best arm identification with fixed confidence (BAI) and the regret minimization objectives. For BAI, we present (a) a cost complexity lower bound, (b) an algorithmic framework with two alternative fidelity selection procedures, and (c) both procedures' cost complexity upper bounds. From both cost complexity bounds of MF-MAB, one can recover the standard sample complexity bounds of the classic (single-fidelity) MAB. For regret minimization of MF-MAB, we propose a new regret definition, prove its problem-independent regret lower bound $\Omega(K^{1/3}\Lambda^{2/3})$ and problem-dependent lower bound $\Omega(K\log \Lambda)$, where $K$ is the number of arms and $\Lambda$ is the decision budget in terms of cost, and devise an elimination-based algorithm whose worst-cost regret upper bound matches its corresponding lower bound up to some logarithmic terms and, whose problem-dependent bound matches its corresponding lower bound in terms of $\Lambda$.

Paper

References (32)

Scroll for more · 20 remaining

Similar papers

Peer review

Reviewer 3tC76/10 · confidence 3/52023-07-05

Summary

The authors consider the problem of multi-fidelity multi-armed bandits under fixed-confidence BAI and regret-minimization objectives. The best-arm identification algorithm is based on lower-upper confidence bound and associated cost-complexity. A novel definition of regret is introduced which captures the fact that fidelities influence the accuracy of affected rewards. Lower and upper bounds are provided for both problem cases.

Strengths

This paper provides a novel outlook at the multi-fidelity MAB problem introduced by Kandaswamy et. al. The BAI algorithm is an index-based algorithm with a new UCB based procedure to determine the optimal fidelity to be sampled.

Weaknesses

-

Questions

-

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

-

Reviewer zHKa7/10 · confidence 4/52023-07-07

Summary

This paper studies the multi-fidelity multi-armed bandits problem where each arm has different fidelities and observation accuracy. There are many existing works in this problem. This paper studies both best arm identification with fixed confidence and regret minimization. There are some novel theoretical results in both problems.

Strengths

1. This paper studies an important problem and overall presentation is good so it’s an easy-to-follow paper. 2. This paper claims to be the first work to study the BAI task under the MF-MAB model. 3. It’s great to see many application-oriented discussion after each theoretical analysis. 4. The tightness of theoretical results is checked; see Remark 3.6, 4.4, 4.5. 5. Compared with Kandasamy et al. [19], the cost complexity definition in this paper makes more sense to me.

Weaknesses

1. Although this paper claims to be the first work to study the BAI task under the MF-MAB model, I do have a question on comparison to Kandasamy et al. [18] as discussed in line 76-82. Yes, simple regret minimization in continuous domain is equivalent to BAI in discrete domain, but after discretization of continuous domain how is the algorithm in this paper compared to Kandasamy et al. [18]? 2. In Remark 3.8, I don’t think $\mu_1$ and $\mu_2$ can be easily obtained in real-world applications. Yes, a perfect classification means $\mu=1.0$ but that’s because of the output range. For the 2nd-best arm, its \mu is hard to get. 3. A minor comment. I do like application-oriented discussion but why do they appear after each technical stuff? I think first being motived by applications, stating the problem, and presenting the results makes more sense to me. 4. A minor suggest. Instead of denoting best arm as arm 1 and 2nd-best arm as arm 2, $a_1$ and $a_2$ look better.

Questions

1. In line 225, it’s good to compare explore-a vs. explore-b but what’s the criterion to select one of them to perform best? 2. There are some technical terms that need more explanation. (2.1) In line 109, why the error upper bounds can be revealed to the learning agent? Could you give a real-world example and explain what are they in this example? (2.2) In line 131, are v^(m)_k different according to different algorithms? If yes, why can they appear in lower bound? (2.3) In line 172, \beta seems like something similar as in GP-UCB paper. Could you explain why \beta is defined as that? Pointing it to somewhere in the proof also works. (2.4) In line 182, how “half” is chosen? I guess any constant between 0 and 1 works, right? If a constant C is used instead of half, how does it affect theoretical results?

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Limitations are discussed in last section.

Reviewer Rfmo4/10 · confidence 2/52023-07-11

Summary

The paper studies the multi-fidelity multi-armed bandit problem where each arm can be pulled at different fidelities, providing better or worse estimate of the true mean, at a different cost. The paper studies best arm identification with fixed confidence, provides a lower bound on the cost complexity, and an algorithm with cost complexity upper bounds. In addition, the paper proposes a new regret definition which finds multiple applications. Upper and lower bounds on the proposed regret are proved for the instance dependent and instance independent cases.

Strengths

- The paper studies a practical problem that has various application. - It is well-written and easy to follow. - The proposed definition of regret looks natural to study and has multiple applications. - The paper provides upper and lower bounds that are tight, in some cases, for both best arm identification and regret minimization.

Weaknesses

My main concerns are the following: - The proposed algorithm for BAI requires the estimates $\tilde{\mu}_1^{(M)}, \tilde{\mu}_2^{(M)}$ which is not a realistic assumption. Even though the author argue that this is possible in some applications, it is still unrealistic for many other applications. These assumptions make the problem theoretically and practically less interesting. - The cost upper bound for BAI depends on $\tilde{m}^\star_k$, which depend on $\tilde{\mu}_1^{(M)}, \tilde{\mu}_2^{(M)}$ that are not guaranteed to be tight estimates. It also has extra terms that are not matched in the lower bound. - The problem dependent regret upper bound is not tight in most cases.

Questions

- Can the proposed regret be upper bounded by $\lambda^{(M)}$ times simple regret considered in the literature? How does the proposed upper bound compare to this? - Please see 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

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

2 fair

Limitations

Yes.

Authorsrebuttal2023-08-21

This is a gentle reminder for the reviewer that the author-reviewer discussion will end in less than 15 hours. In case you have any further questions, please feel free to ask; we are happy to answer them.

Reviewer QHcL5/10 · confidence 2/52023-07-21

Summary

This paper studys the problem of Multi-Fidelity Multi-Armed Bandits (MF-MAB) where each arm can be pulled at different fidelity with different rewards and costs. The main contribution of this paper includes derive the cost complxity lower bound for best arm idenfication with fixed confidence and a new definition of regret for regret minimization.

Strengths

The theorical results about the cost complexity lower bound is interesting and the proof of each theorem is comprehensive.

Weaknesses

- In Algorithm 1, the best arm and second best arm is selected according to the up confidence bound (UCB). - Why use the UCB given that UCB of an arm is larger than that of another arm does not necessary mean that the first arm is better? How about use LCB? - Since the UCB and LCB both depends on the confidence radius, is it possible that the algorithm will come to an earlier decision to pick the arm that is not well explored? Here, I am assuming that if the arm is not well expored the confidence radius will be large thus the UCB will also be large. - It would be better to have some toy examples/experiments to show the effectiveness of the theorical results?

Questions

As in Weakness section

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

3 good

Contribution

3 good

Limitations

Yes, the authors addressed the limitations.

Authorsrebuttal2023-08-20

New experiment updated in general comment

This is a gentle reminder to the reviewer that we added [a new simulation in the general comments](https://openreview.net/forum?id=oi45JlpSOT&noteId=qMzrbHt8FZ). It compares our two exploration policies to one baseline and illustrates when our exploration policies (utilizing low fidelity) outperform the baseline (not utilizing low fidelity).

Reviewer ZToP7/10 · confidence 3/52023-07-26

Summary

In this work authors revisit the multi-fidelity multi armed bandits and provide results i.e., upper and lower bounds for cost complexity for best arm identification with fixed confidence objective using the Lower-upper confidence bound framework. Further, this work introduces 3 procedures for finding optimal fidelity where the comparable upper bound results are given for 1st and 2nd procedure. Finally, the regret minimization objective is discusses briefly, introducing new definition of regret which differs from the definition introduced in the prior work.

Strengths

The paper has made a significant effort to keep clear problem framework and analysis. The proposed method is novel for the best arm identification objective. Further, a detailed theoretical analysis of LUCB framework for both best arm identification objective and regret minimization objective, and comparable upper and lower bounds on the cost complexity and regret respectively are theoretically sound and well presented.

Weaknesses

This work assumes the reward distributions are on bounded support which may not be the case always. Further the knowledge of upper and lower bounds of 1st and 2nd arm respectively is assumed and when these bounds are too loose then the upper bounds on the cost complexity would not be comparable to lower bounds. The assumption 3.3 on which all the results are dependent may not hold always. The paper is missing the experimental evaluation of the proposed framework and comparison with the existing methods. Lastly, some times paper is hard to follow when 1st and 2nd arms are mentioned as to when is the ground truth being referred and when is the indexing of arms being referred.

Questions

Is the assumption 3.3 expected to hold in general?

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer rkrs3/10 · confidence 4/52023-07-27

Summary

This paper studies the the multi-fidelity multi-armed bandit setting where each arm is associated with a cost (fidelity) and observed reward. In this setting when the learner pulls an arm $k \in \mathcal{K}:=\{1, \ldots, K\}$ at fidelity $m \in \mathcal{M}:=\{1, \ldots, M\}$, it pays a cost of $\lambda^{(m)}$ and observes a reward $X_k^{(m)}$ whose mean $\mu_k^{(m)}$ is not too far away from the arm's true reward mean $\mu_k$. This paper studies both the best arm identification and regret minimization setting. For the best arm setting, they propose a LUCB algorithm where the key novelty/challenge (different from standard MAB) is to decide the fidelities for exploring the critical arms. This fidelity selection faces an accuracy-cost trade-off, i.e., higher fidelity (accuracy) but suffering higher cost, or lower fidelity (accuracy) but enjoying lower cost. Hence, this trade-off is different from the common exploration-exploitation trade-off in classic bandits. To address this trade-off in fidelity selections, they propose a UCB-type policy which is called upon by the LUCB algorithm that finds the optimal fidelity $m_k^*$ in (3) for each arm $k$ (EXPLORE-A) and an explore-then-commit policy stopping at a good fidelity that is at least half as good as the optimal one (EXPLORE-B). In theorem 3.4 they show that the sample complexity upper bound matches the lower bound upto constant factors. they also study the regret minimization setting.

Strengths

1) The best arm identification in multi-fidelity multi-armed bandit setting is novel and interesting setting. 2) The proposed solution seems novel and is theoretically analyzed. 3) The sample complexity upper bound matches the lower bound under certain assumptions.

Weaknesses

1) You assume that the reward distribution and the mean $\mu_k^{(m)}$ of each arm $k$ at fidelity $m$ are unknown, while the costs $\lambda^{(m)}$ 's and the error upper bounds $\zeta^{(m)}$ 's are known to the learning agent. This setting assumption seems very contrived. Is this a standard assumption? Can you point out some references? Can you give some motivating examples where this setting assumption makes sense. 2) The writing needs to substantially improve. There are too many results packed into the paper, without discussion on a specific result in depth. For example one of the main contribution is the regret bound, however, the regret bound theorem is moved to appendix D entirely. Also the regret minimization algorithm is not presented in the main paper and is presented in Appendix D. The novelty of the arm elimination algorithm is not clear to me. Is the regret bound of the regret minimization algorithm tight? 3) This is mainly a theory paper, yet the technical novelty of your approach is not clear to me. What are the technical challenges in proving Theorem 3.4? I think this needs a detailed discussion. For example the [Kandasamy et al.](https://papers.nips.cc/paper_files/paper/2016/file/2ba596643cbbbc20318224181fa46b28-Paper.pdf) paper tackles the multi-fidelity setting for regret minimization and [Kauffmann et al](http://proceedings.mlr.press/v30/Kaufmann13.pdf) studies the LUCB algorithm for single fidelity. My guess is that the proof of thorem 3.4 must use these papers. Can the authors briefly discuss the proof technique. 4) There is no experiment in the main paper. Yet, multi-fidelity paper do substantial experiments on their setting (See [Kandasamy et al.](https://papers.nips.cc/paper_files/paper/2016/file/2ba596643cbbbc20318224181fa46b28-Paper.pdf) ). What are the baselines in your experiments? 5) There is a summation over $m\in\mathcal{M}$ in the Explore-B sample complexity bound, which is not present in Explore-A bound. It is said that when $\tilde{G}=O(M \tilde{H})$, the cost complexity upper bound of EXPLORE-A is less than that of EXPLORE-B. However, if $\tilde{G}=O(M \tilde{H})$ then EXPLORE-A bound is far away from the lower bound. Can you clarify this? Also what are the implications of $\tilde{G}=O(M \tilde{H})$ and what settings have these? In short, when do we prefer EXPLORE-A than EXPLORE-B?

Questions

See weakness section.

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

See weakness section.

Reviewer rkrs2023-08-17

Further clarification questions

I thank the authors for their response. I have a few clarification questions which will help me understand the paper: - Thank you for clarifying the motivation of the paper. - I still feel that the presentation of the paper is unsatisfactory where half the result, including the regret algorithm, regret theorem is in the appendix. Regarding novelty of the regret, I have a few queries. - I am concerned about the regret definition in eq (9). There are two issues. Why is the definition only discounting $\mu^1$ by $\lambda^{1}$ and not $\mu\_{I_t}$ by also $\lambda^{I_t}$? Why the $N$ is required and what happens if you simply sum over $t=1$ to $\Lambda$? - The algorithmic novelty of MF-MAB in Appendix D is still not clear to me. It seems just a kernelized version of UCB-Improved of [Auer et al (2010)](). Pull each arm in the active set equal number of times at highest fidelity and then eliminate sub-optimal arms at that are below the gap $2^{-p + 1}$. - The factor $\epsilon$ is not described in the algorithm description in Appendix D.2 (from lines 617-622). Is it a problem dependent parameter, a lower bound to the gaps? I see that it is showing up in the Theorem D.3 as $1/\epsilon^2$. Does this mean that you need to know a lower bound to the gaps for running MF-MAB? - This queries again bring me to one of the key points I am raising. Algorithm design is one of the key contribution of this paper. You *cannot* put the algorithm in Appendix, hide away key design choices and discuss result in remarks. - Experiments: No further experiments are provided by the authors. UCB revisited: Improved regret bounds for the stochastic multi-armed bandit problem Peter Auer 2010

Authorsrebuttal2023-08-18

We highly appreciate the reviewer for spending time evaluating this work. > I am concerned about the regret definition in eq (9). There are two issues. - > Why is the definition only discounting $\mu_1$ by $\lambda^{(1)}$ and not $\mu_{I_t}$ by also $\lambda^{(I_t)}$? **Answer:** We are *not* discounting reward by fidelity. No matter which fidelity is leveraged to pull the arm, the obtained observations are counted as rewards without any scaling. Please refer to Remark 4.1 for the application motivation of not discounting rewards by fidelity. Since there is no reward scaling due to different fidelity, in the second term $\mathbb{E}\left[ \sum_{t=1}^N \mu_{I_t}^{(M)} \right]$ of Eq. (9), the $\mu_{I_t}^{(M)}$ is not discounted. The first term $\frac{\Lambda}{\lambda^{(1)}}\mu_1$ of Eq. (9) signifies that the optimal strategy for maximizing cumulative rewards involves consistently pulling the optimal arm $\mu_1$ with the lowest fidelity $\lambda^{(1)}$, where the expected pulling times are $\frac{\Lambda}{\lambda^{(1)}}$, and the $\lambda^{(1)}$ is not for discounting the reward mean $\mu_1^{(M)}$. - > Why the $N$ is required and what happens if you simply sum over $t=1$ to $\Lambda$? **Answer:** $N$ represents the overall count of decision rounds (time slots) within the algorithm under consideration, and $\Lambda$ stands for the total budget allocation for said algorithm. In light of the fact that the second term in the regret definition (e.g., Eq. (9)) is the accumulation of rewards across all time slots, one needs to perform the summation across the range from $t=1$ to $t=N$. As $t$ serves as the index for time slots and possesses distinct units from the budget $\Lambda$, it is inappropriate to execute a summation over the interval $t=1$ to $t=\Lambda$. > The algorithmic novelty of MF-MAB in Appendix D is still not clear to me. It seems just a kernelized version of UCB-Improved of Auer et al (2010). Pull each arm in the active set equal number of times at highest fidelity and then eliminate sub-optimal arms at that are below the gap $2^{-p + 1}$. **Answer:** The primary innovation in the algorithm's design lies in its two-stage approach: a high-fidelity exploration phase (Lines 3-7 in Algorithm 4) followed by a low-fidelity exploitation phase (Line 8 in Algorithm 4). As acknowledged in the paper, the design of this algorithm, particularly its exploration stage, bears resemblance to the work of Auer et al. (2010). However, our paper extends beyond this algorithmic design and centers on demonstrating that the algorithm can attain tight regret upper bounds (both problem-dependent and -independent) for the novel regret definition in Eq. (9) for the MF-MAB model. This novel MF-MAB regret definition is suitable to model many practical application scenarios in real-world applications, e.g., advertisement distribution as described in the paper and appreciated by the other reviewers. The two-stage algorithm provides a practical strategy for such scenarios. The proposed tight regret guarantees of the two-stage algorithm support the viability of adopting such a design in practical settings. > The factor $\epsilon$ is not described in the algorithm description in Appendix D.2 (from lines 617-622). Is it a problem dependent parameter, a lower bound to the gaps? I see that it is showing up in the Theorem D.3 as $1/\epsilon^2$. Does this mean that you need to know a lower bound to the gaps for running MF-MAB? **Answer:** No, $\epsilon$ is not gap dependent and one can run the algorithm with any $\epsilon > 0$. The regret upper bound in Theorem D.3. (Eq. (21)) holds for any $\epsilon > 0$. Theorem D.4. shows that setting $\epsilon = (K\log \Lambda / \Lambda)^{1/3}$ guarantees the problem-independent (worst case) regret upper bound for any MF-MAB instance. > This queries again bring me to one of the key points I am raising. Algorithm design is one of the key contribution of this paper. You cannot put the algorithm in Appendix, hide away key design choices and discuss result in remarks. **Answer:** We thank the reviewer for suggesting a better presentation of this paper. Although we agree that the algorithm design (Algorithm 4) is a contribution to this paper, it is only one of the key contributions (among other algorithm designs and theoretical results). Especially as this is a theoretical paper, it is also very important to present the theoretical results. We indeed made hard decisions on what to keep in the paper in order to achieve a balance. As the camera-ready version allows an additional page, we promise to move the algorithm design and the detailed theorems in Appendix D into the main paper in the final version.

Authorsrebuttal2023-08-20

New experiment updated in general comment

This is a gentle reminder to the reviewer that we added [a new simulation in the general comments](https://openreview.net/forum?id=oi45JlpSOT&noteId=qMzrbHt8FZ). It compares our two exploration policies to one baseline and illustrates when our exploration policies (utilizing low fidelity) outperform the baseline (not utilizing low fidelity).

Reviewer rkrs2023-08-21

Response to authors

Thanks for your response and I have noted the simulation as well. We will discuss these results and your responses in the reviewer-area chair discussion period.

Reviewer ZToP2023-08-18

I thank the authors for the response. The reviewer has no further questions.

Reviewer zHKa2023-08-18

Thanks for rebuttal

Authors' rebuttal addresses most of my concerns and I really encourage authors to polish the paper presentation by incorporating our discuss in review and rebuttal in the next version. I'm happy to increase my rating from 6 to 7.

Authorsrebuttal2023-08-19

New Experiment: Empirical Comparison to Baseline

We highly appreciate the reviewers and area chairs for spending time evaluating this paper. In this post, we present a new numerical simulation that compares our two exploration policies, $\texttt{Explore-A}$ and $\texttt{Explore-B}$, with a $\texttt{Baseline}$ that is always selecting the highest fidelity to explore arms. All three policies are employed under the LUCB framework (Algorithm 1). We note that, for the MF-MAB model, there is no prior algorithm (other baselines) on the best arm identification with fixed confidence setting (see Section 1.2's second paragraph for related work discussion). This simulation aims to illustrate when our exploration policies $\texttt{Explore-A}$ and $\texttt{Explore-B}$ (utilize low-fidelity) are better than $\texttt{Baseline}$ (not utilize low-fidelity). **Experiment Setup.** We consider a MF-MAB model consisting of $K=5$ arms and $M=5$ fidelities. The reward means are set as $\mu_{k}^{(m)} = 0.1 (k+4) + 0.01(m-5)$, where, for example, $\mu_5^{(5)} = 0.9$ is the true reward mean of optimal arm and its low-fidelity reward means are $\{0.89, 0.88, 0.87, 0.86\}$ accordingly, etc. The error upper bound in $\zeta^{(m)}$ are $\\{0.04, 0.03, 0.02, 0.01, 0\\}$ for $m\in\\{1,2,3,4,5\\}$ respectively. We set the cost of fidelity $m\in\\{1,2,3,4\\}$ as $\lambda^{(m)} = m$, and consider four different cases $I\in \\{1,2,3,4\\}$ of the cost of highest fidelity $\lambda^{(5)}=5I$. That is, from Case 1 to Case 4, the cost of the highest fidelity $\lambda^{(5)}$ increases from $5$ to $20$. We set the confidence parameter $\delta=0.1$ and run each experiment for 100 trials. Averaged over these $100$ trials of each experiment, we present the empirical cost complexity in the table as follows, | Cost Complexity ($\times 10^4$) | Case 1 ($\lambda^{(5)}=5$) | Case 2 ($\lambda^{(5)}=10$) | Case 3 ($\lambda^{(5)}=15$) | Case 4 ($\lambda^{(5)}=20$) | | -------------------- | ------ | ------ | ------ | ------ | | $\texttt{Explore-A}$ | $20.5$ | $20.0$ | $17.4$ | $17.4$ | | $\texttt{Explore-B}$ | $36.1$ | $47.9$ | $59.5$ | $69.4$ | | $\texttt{Baseline}$ | $21.4$ | $42.0$ |$64.6$ |$86.2$ | **Discussion of Simulation Results.** In all four cases, the cost complexities of $\texttt{Explore-A}$ are similar. Especially, when the cost of the highest fidelity $\lambda^{(5)}$ becomes larger, it is easier for $\texttt{Explore-A}$ to find out the optimal fidelity (e.g., the highest fidelity is obviously not the optimal one due to its high cost) and, hence, $\texttt{Explore-A}$ enjoys a slightly lower cost complexity. The cost complexities of both $\texttt{Explore-B}$ and $\texttt{Baseline}$ increase along these four cases. When the cost of the highest fidelity is expensive, e.g., in Cases 3 and 4, both $\texttt{Explore-A}$ and $\texttt{Explore-B}$ outperform $\texttt{Baseline}$. This highlights the advantage of utilizing low fidelity to explore arms while avoiding using the expensive high fidelity. When the cost of the highest fidelity is cheap, e.g., in Case 1, the cost complexity of $\texttt{Explore-A}$ is similar to $\texttt{Baseline}$, while $\texttt{Explore-B}$ is worse than $\texttt{Baseline}$. This suggests that when the cost of the highest fidelity is similar to that of lower fidelity, utilizing the high-fidelity to explore arms (i.e., $\texttt{Baseline}$) may be preferable. In this comparison, $\texttt{Explore-A}$ always outperforms $\texttt{Explore-B}$. For another scenario where $\texttt{Explore-B}$ is better than $\texttt{Explore-B}$, please refer to Figure 1(b) in our paper. Last but not least, we emphasize that this paper is a theoretical work and our focus is on devising algorithms and proving their theoretical guarantees. Our numerical experiments so far (including Figures 1(a) and 1(b) in the paper and the additional one reported here) achieve the purpose of providing the basic support and verification of the results in the theorems. We will include the new comparison in the above table to the final version of this paper as well. We also plan to further validate our algorithms with more experiments on larger datasets.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC