Bandit problems

24.05.2018, 16:15
Forschungsseminar Wahrscheinlichkeitstheorie

Pierre Ménard

First Talk

Title: An introduction to best arm identification in the bandit problems

Abstract: In this talk we will introduce the bandit setting and in particular the problem of best arm identification in the fixed confidence setting.
We will show how to characterize the complexity of these problems and present an algorithm that reaches asymptotically this complexity.
Most of the results that will be presented are extract from Optimal Best Arm Identification with Fixed Confidence, Garivier and Kaufmann, 2016.

Second Talk
Title: Threshold bandit for dose-ranging

Abstract: We will present a variant of the best arm identification in the bandit problem: the threshold bandit problem. It is a way to model the dose-ranging problem for phase I
of clinical trials. As in the fist talk we will give the complexity of these problems and present an algorithm to solve them. This talk is based on the paper: Thresholding Bandit for Dose-ranging: The Impact of Monotonicity, Garivier et al., 2017.