Provably Efficient Multi-Objective Bandit Algorithms under Preference-Centric Customization
Keywords: multi-objective multi-arm bandit, bandit optimization, preference-centric learning
TL;DR: We consider a preference-aware multi-objective multi-arm bandit framework, with tailored algorithms proposed under different preference structrures that enjoy sub-linear regrets.
Abstract: Existing multi-objective multi-armed bandit (MO-MAB) approaches mainly focus on achieving Pareto optimality. However, a Pareto optimal arm that receives a high score from one user may lead to a low score from another, since in real-world scenarios, users often have diverse preferences across different objectives. Instead, these preferences should inform *customized learning*, a factor usually neglected in prior research. To address this need, we study a *preference-aware* MO-MAB framework in the presence of explicit user preferences, where each user’s overall-reward is modeled as the inner product of user preference and arm reward. This new framework shifts the focus from merely achieving Pareto optimality to further optimizing within the Pareto front under preference-centric customization. To the best of our knowledge, this is the first theoretical exploration of customized MO-MAB optimization based on explicit user preferences. This framework introduces new and unique challenges for algorithm design for customized optimization. To
address these challenges, we incorporate *preference estimation* and *preference-aware optimization* as key mechanisms for preference adaptation, and develop new analytical techniques to rigorously account for the impact of preference estimation errors on overall performance. Under this framework, we consider three preference structures inspired by practical applications, with tailored algorithms that are proven to achieve near-optimal regret, and show good numerical performance.
Primary Area: learning theory
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4667
Loading