Adaptive Algorithms for Relaxed Pareto Set Identification

In this paper we revisit the fixed-confidence identification of the Pareto optimal set in a multi-objective multi-armed bandit model. As the sample complexity to identify the exact Pareto set can be very large, a relaxation allowing to output some additional near-optimal arms has been studied. In this work we also tackle alternative relaxations that allow instead to identify a relevant subset of the Pareto set. Notably, we propose a single sampling strategy, called Adaptive Pareto Exploration, that can be used in conjunction with different stopping rules to take into account different relaxations of the Pareto Set Identification problem. We analyze the sample complexity of these different combinations, quantifying in particular the reduction in sample complexity that occurs when one seeks to identify at most $k$ Pareto optimal arms. We showcase the good practical performance of Adaptive Pareto Exploration on a real-world scenario, in which we adaptively explore several vaccination strategies against Covid-19 in order to find the optimal ones when multiple immunogenicity criteria are taken into account.

Paper

References (29)

Scroll for more · 17 remaining

Similar papers

Peer review

Reviewer rUsc6/10 · confidence 2/52023-07-01

Summary

This paper studies a less common setting that arms are multi-variate distributions. In the expiration period, the agent seeks to identify the arm which is Pareto optimal. This problem is formulated as a fixed-confidence identification problem. The most novel setting in that, the fixed confidence identification problem s extended for multi-objective settings. This paper proposed a novel sampling rule called Adaptive Pareto Exploration. The central idea is to identify the two contentious arms and sample both or one of them. The paper also gives the theoretical guarantee that the proposed method can recommend a correct set whp after most a given number of iterations.

Strengths

This paper proposes a principled method and provides theoretical guarantees.

Weaknesses

It seems trival to me to extend the single objective best arm identification problem to the mo setting.

Questions

1. Can the proposed method identify the full (relaxed) Pareto front?

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

3 good

Contribution

3 good

Limitations

No.

Reviewer yuHu6/10 · confidence 3/52023-07-07

Summary

The primary objective of the paper is to address the challenge of identifying a set of arms, consisting of at most $k$ arms, where each arm is either Pareto optimal or close to Pareto optimal. The paper explains the concept of Pareto optimality within the context of bandit problems. To tackle this problem, the paper presents the $\epsilon_1$-APE-$k$ (Adaptive Pareto Exploration) algorithm and establishes an upper bound on the sample complexity. Furthermore, empirical evidence is provided to demonstrate the effectiveness of the proposed algorithm.

Strengths

One strength of the paper is its focus on addressing the problem of Pareto optimality in bandit literature, which is an underexplored area. The paper introduces a novel problem setup and clearly defines the goals for different scenarios related to Pareto optimal identification, including the identification of Pareto optimal sets, near Pareto optimal sets, and at-most $k$ near Pareto optimal sets. The paper excels in providing a comprehensive discussion on sample complexity upper bounds, highlighting the theoretical aspects of the problem. It also establishes connections to the existing literature on Pareto optimality in bandits, demonstrating a solid understanding of the research landscape. Furthermore, the paper offers ample experimental evidence to support its claims, including experiments conducted on real-world datasets. This empirical validation strengthens the credibility of the proposed approaches and enhances the practical relevance of the research.

Weaknesses

One weakness of the paper is the absence of references to real-world applications that could benefit from the proposed framework. For instance, discussing potential applications such as A/B testing in clinical trials would enhance the practical relevance and broader impact of the research. Another weakness is the lack of discussion of lower bounds in the main paper. Given that the problem setup for Pareto set identification is relatively new, it would be valuable to explore more properties of the lower bounds to gain insights into the tightness of the derived upper bounds. This would provide a more comprehensive understanding of the problem and its inherent complexities. The paper lacks a thorough discussion on the computational complexity of the $\epsilon_1$-APE-$k$ algorithm. Considering the importance of computational efficiency in practical applications, it would be beneficial to address

Questions

One potential weakness is the lack of information regarding scalability issues encountered when running large-scale experiments using the $\epsilon_1$-APE-$k$ algorithm. It would be valuable to understand whether the algorithm faces any challenges in handling larger datasets or more complex scenarios. Additionally, providing references on the scale of parameters used in A/B testing for applications beyond COVID datasets would further strengthen the practicality of the proposed approach. Another point to consider is the limited discussion on the broader scope of utilizing Pareto sets in various applications apart from clinical trials. Exploring and discussing other potential domains where Pareto sets could be beneficial would enhance the paper's impact and shed light on additional practical applications.

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

No limitations or potential impact of their work discussed

Reviewer J7Kd5/10 · confidence 2/52023-07-09

Summary

This paper extends the best arm bandit problem to a multi-objective version to identify the arms that are in the Pareto front.

Strengths

This problem is new and the author provides a clear practice motivation of this problem. The theory study seems valid and complete.

Weaknesses

The math notation seems a bit over-complicated and I was wondering can you simplify it? Overall, extending the best arm bandit problem into multi-object setting seems straightforward, which limits the novelty of this paper. The experiment result looks a bit hand-wavy. I understand that there might not be any algorithm specifically designed for such problem, but I was wondering how the proposed method compares with some naively modified baselines?

Questions

In the definition of Pareto front, why we don't consider the variance of the distribution?

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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

See above

Reviewer 9Lr46/10 · confidence 3/52023-07-10

Summary

This paper studies a relaxed problem of Pareto set identification, in which the learner is required to identify a subset of the optimal arms. A single sampling strategy APE is proposed and then combined with various stopping rules to realize different relaxations of the original problem. In theory, this paper derives the sample complexity of these combinations. The proposed method is also validated empirically on a real-world scenario of Covid-19 vaccination.

Strengths

1. Relaxed Pareto set identification is an important problem for multi-objective MAB. The proposed new framework may inspire new researches in this field, if properly justified. 2. The proposed sampling strategy sounds novel, and the analysis is technically sound. 3. The experiment on selecting vaccination strategies against Covid-19 is interesting.

Weaknesses

1. One of the main contributions of this paper is to propose a new problem of $\epsilon$-PSI-$k$ and provide an analysis framework for this problem. However, I think the motivation of $\epsilon$-PSI-$k$ should be explained in more detail. I am not sure if we need this new formulation as $(\epsilon_1,\epsilon_2)$-PSI has already dealt with the sparsity issue. 2. Some detailed explanations of the intuition of the proposed sampling strategy (and why it outperforms previous strategies in principle) may help better understand the novelty in the algorithm design.

Questions

See the weaknesses.

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

Limitations have been properly discussed.

Reviewer rUsc2023-08-17

reply

I have read the response. And since I am not an expert in this field. I am not confident with my score.

Reviewer J7Kd2023-08-18

Thanks.

Thanks for the rebuttal. I keep my original score.

Reviewer yuHu2023-08-19

I would like to thank the authors for their detailed responses, they helped me clear up some of my queries.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC