This is a reminder that on Thursday august 31 we have Chloé Rouyer in the Statistics and Machine Learning Thematic Seminar. There is also a change in topic.
Thursday august 31, 16:00-17:00
In person, at the University of Amsterdam
Location: Science park 904 A1.24
Title: A near-optimal best-of-both-worlds algorithm for online learning with feedback graphs
Abstract: In this work, we consider an online learning problem that interpolates between bandits and full information problems: extra observations are granted depending on the action played and the position of that action in a feedback graph, which is known to the learner. While this problem has been widely studied against both adversarial and stochastic environments separately, aiming to obtain best-of-both-worlds guarantees is a challenging problem that has only been studied recently. We propose an algorithm that achieves near-optimal guarantees in both the adversarial and the stochastic regimes simultaneously. Our results are also computationally efficient and can naturally adapt to sequences of feedback graphs that change over time.