Robust Batched Bandits

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Guo, Yunwen, Shu, Yunlun, Zhuo, Gongyi, Wang, Tianyu
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908903891009536
author Guo, Yunwen
Shu, Yunlun
Zhuo, Gongyi
Wang, Tianyu
author_facet Guo, Yunwen
Shu, Yunlun
Zhuo, Gongyi
Wang, Tianyu
contents The batched multi-armed bandit (MAB) problem, in which rewards are collected in batches, is crucial for applications such as clinical trials. Existing research predominantly assumes light-tailed reward distributions, yet many real-world scenarios, including clinical outcomes, exhibit heavy-tailed characteristics. This paper bridges this gap by proposing robust batched bandit algorithms designed for heavy-tailed rewards, within both finite-arm and Lipschitz-continuous settings. We reveal a surprising phenomenon: in the instance-independent regime, as well as in the Lipschitz setting, heavier-tailed rewards necessitate a smaller number of batches to achieve near-optimal regret. In stark contrast, for the instance-dependent setting, the required number of batches to attain near-optimal regret remains invariant with respect to tail heaviness.
format Preprint
id arxiv_https___arxiv_org_abs_2510_03798
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Robust Batched Bandits
Guo, Yunwen
Shu, Yunlun
Zhuo, Gongyi
Wang, Tianyu
Machine Learning
The batched multi-armed bandit (MAB) problem, in which rewards are collected in batches, is crucial for applications such as clinical trials. Existing research predominantly assumes light-tailed reward distributions, yet many real-world scenarios, including clinical outcomes, exhibit heavy-tailed characteristics. This paper bridges this gap by proposing robust batched bandit algorithms designed for heavy-tailed rewards, within both finite-arm and Lipschitz-continuous settings. We reveal a surprising phenomenon: in the instance-independent regime, as well as in the Lipschitz setting, heavier-tailed rewards necessitate a smaller number of batches to achieve near-optimal regret. In stark contrast, for the instance-dependent setting, the required number of batches to attain near-optimal regret remains invariant with respect to tail heaviness.
title Robust Batched Bandits
topic Machine Learning
url https://arxiv.org/abs/2510.03798