“Learning from a bandit learner”
Martedì 23 Giugno 2026, ore 17:00 - Aula 2AB45 - Michela Petriconi (University of Tübingen)
Abstract
Bandit problems provide a simple yet powerful model for decision-making under uncertainty, with applications in recommender systems, A/B testing, clinical trials, and online optimization. In this talk, I will introduce the basics of bandit problems and standard learning algorithms, before presenting a new research direction: inverse bandits. Instead of asking how a bandit learner should act in an unknown environment, inverse bandits ask what can be inferred about the environment from the learner’s behavior. I will show that, under suitable conditions, this behavior can reveal the learner’s beliefs about the problem it is solving.
