Variance-sensitive Thompson sampling for generalised linear bandits, revisited
Title: Revisiting Variance-Sensitive Thompson Sampling for Generalized Linear Bandits
Abstract: This paper establishes a variance-sensitive regret bound for Thompson sampling within the context of stochastic generalized linear bandits. The proposed analysis relies on a warm-up phase, after which the Gaussian Poincaré inequality is employed to manage regret. This approach circumvents the limitations where prior optimism-based analytical methods fail. While eliminating the warm-up period while maintaining identical variance-sensitive scaling is still an open problem, the authors suggest that this challenge is nontrivial.
Source: arXiv Generated at: 2026-06-02 00:00:00 UTC
