Learning to Choose Under a Fixed Budget: Bandits, Adaptivity, and Large Deviations

Abstract

In many decision problems, we must choose the best option after only a limited number of trials. Examples include A/B testing, online advertising, recommendation systems, and career exploration. This talk studies such questions through the lens of fixed-budget best-arm identification in stochastic bandits. I will first revisit the basic two-option setting, where the value of adaptivity is more subtle than it may appear. I will then move to the multi-armed setting, where adaptive elimination strategies become natural. Finally, I will explain how large-deviation ideas provide tools for analyzing adaptive sampling algorithms and for understanding both the promise and the limitations of adaptivity under a fixed budget.

線上視訊請點選連結