Finite-Time Regret Analysis of Retry-Aware Bandits

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Tong, Bingkui, Komiyama, Junpei, Nishimori, Soichiro, Parmas, Paavo
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918533353439232
author Tong, Bingkui
Komiyama, Junpei
Nishimori, Soichiro
Parmas, Paavo
author_facet Tong, Bingkui
Komiyama, Junpei
Nishimori, Soichiro
Parmas, Paavo
contents We study a stochastic bandit algorithm motivated by retry-aware objectives that value the best outcome among multiple attempts, such as pass@$k$ and max@$k$. Given a posterior over arm values, ReMax chooses a sampling distribution that maximizes the posterior expected maximum reward over $M$ virtual draws. Although this objective was introduced in reinforcement learning as an exploration mechanism under uncertainty, its regret properties in bandit problems have remained unclear. For Gaussian rewards and the first nontrivial case $M=2$, we characterize the optimal ReMax distribution through an expected-improvement balance condition and prove the first sublinear regret bound for ReMax. Our analysis separates the usual saturation behavior of suboptimal arms from a ReMax-specific underestimation effect, in which the optimal arm may be sampled too rarely after an unfavorable estimate. This explains why ReMax can be more exploitative than Thompson sampling (TS) and why its regret analysis is technically delicate. Experiments support this picture: ReMax often outperforms KL-UCB and Thompson sampling under mild underestimation, while posterior-variance scaling empirically mitigates severe underestimation.
format Preprint
id arxiv_https___arxiv_org_abs_2605_20854
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Finite-Time Regret Analysis of Retry-Aware Bandits
Tong, Bingkui
Komiyama, Junpei
Nishimori, Soichiro
Parmas, Paavo
Machine Learning
We study a stochastic bandit algorithm motivated by retry-aware objectives that value the best outcome among multiple attempts, such as pass@$k$ and max@$k$. Given a posterior over arm values, ReMax chooses a sampling distribution that maximizes the posterior expected maximum reward over $M$ virtual draws. Although this objective was introduced in reinforcement learning as an exploration mechanism under uncertainty, its regret properties in bandit problems have remained unclear. For Gaussian rewards and the first nontrivial case $M=2$, we characterize the optimal ReMax distribution through an expected-improvement balance condition and prove the first sublinear regret bound for ReMax. Our analysis separates the usual saturation behavior of suboptimal arms from a ReMax-specific underestimation effect, in which the optimal arm may be sampled too rarely after an unfavorable estimate. This explains why ReMax can be more exploitative than Thompson sampling (TS) and why its regret analysis is technically delicate. Experiments support this picture: ReMax often outperforms KL-UCB and Thompson sampling under mild underestimation, while posterior-variance scaling empirically mitigates severe underestimation.
title Finite-Time Regret Analysis of Retry-Aware Bandits
topic Machine Learning
url https://arxiv.org/abs/2605.20854