Date - Heure / Date - Hour
Date(s) - 12/01/2017
11h00 - 12h00
Emplacement / Location
Amphi Breguet, Building Breguet
Adversarial multi-armed bandits: an online-learning problem with robustness guarantees.
In this talk we address the so-called “adversarial” multi-armed bandit problem, where no assumption is made on the sequence of payoffs received by the learner. This worst-case viewpoint provides robustness guarantees, together with online techniques for automatic parameter tuning. We will first introduce the classical bandit setting (with stochastic payoffs) before describing the adversarial problem in details. We will finally present some recent results published in the paper Refined Lower Bounds for Adversarial Bandits.
Sébastien Gerchinovitz, Institut de Mathématiques de Toulouse, Université Paul Sabatier.