Gaussian-Width Bounds for Offline-to-Online Linear Bandits
Read the original on arXiv Machine Learning →The paper investigates how offline data can improve online learning in stochastic linear bandits, focusing on the case where only a subset of feature directions is poorly covered. It shows that previous regret bounds of ≥O(√{dkT}) stem from loose reward prediction error control and derives tighter bounds using concentration of Lipschitz functions of Gaussian vectors. The authors propose a phased elimination algorithm achieving regret ≥O(k√T) and establish matching lower bounds for both partial and uniform coverage, while also addressing pure exploration with offline data and providing matching bounds for simple regret.
Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.