Showing 1–1 of 1 results for author: Chassang, S

Batched bandit problems

Authors: Vianney Perchet, Philippe Rigollet, Sylvain Chassang, Erik Snowberg

Abstract: Motivated by practical applications, chiefly clinical trials, we study the regret achievable for stochastic bandits under the constraint that the employed policy must split trials into a small number of batches. We propose a simple policy, and show that a very small number of batches gives close to minimax optimal regret bounds. As a byproduct, we derive optimal policies with low switching cost fo… ▽ More Motivated by practical applications, chiefly clinical trials, we study the regret achievable for stochastic bandits under the constraint that the employed policy must split trials into a small number of batches. We propose a simple policy, and show that a very small number of batches gives close to minimax optimal regret bounds. As a byproduct, we derive optimal policies with low switching cost for stochastic bandits. △ Less

Submitted 29 March, 2016; v1 submitted 2 May, 2015; originally announced May 2015.

Comments: Published at http://dx.doi.org/10.1214/15-AOS1381 in the Annals of Statistics (http://www.imstat.org/aos/) by the Institute of Mathematical Statistics (http://www.imstat.org)

Report number: IMS-AOS-AOS1381

Journal ref: Annals of Statistics 2016, Vol. 44, No. 2, 660-681