A1227
Title: Stability and inference in adaptive experiments
Authors: Koulik Khamaru - Rutgers University (United States) [presenting]
Abstract: Modern decision-making increasingly relies on adaptive experimentation, particularly in settings such as A/B testing, multi-armed bandits, and reinforcement learning. While these methods enable more efficient learning and allocation of resources, they fundamentally challenge traditional statistical inference. Classical i.i.d.-based tools often break down under adaptive data collection, resulting in biased estimators and misleading confidence intervals. The aim is to offer an overview of statistical inference in these adaptive environments. The pitfalls of naive inference are highlighted through concrete examples, and the concept of stability is introduced, originally formulated in a past study, as a unifying principle for valid inference under adaptivity. In particular, it is demonstrated how algorithms like the upper confidence bound (UCB) achieve stability, enabling the application of classical inferential tools despite the lack of independence. Key illustrations include the empirical mean in a stochastic multi-armed bandit and the contextual bandit problem, both supported by a central limit theorem.
