Multistage Bandits
Abstract
Motivated by practical applications, chiefly clinical trials and web design optimization, we study the regret achievable for stochastic multi-armed bandits under the constraint that the employed policy must function in a small number of stages. Our results show that a very small number of stages gives already close to minimax optimal regret bounds and we also evaluate the number of trials in each stage. [Joint work with: S. Chassang, V. Perchet and E. Snowberg]
Background
Philippe Rigollet received his Ph.D. in mathematics from University Paris 6 where he studied in the Laboratoire de Probabilites under the supervision of Alexandre Tsybakov. He then moved to Georgia Tech as a Post-Doc working with Vladimir Koltchinskii in the School of Mathematics. In 2008, he joined the faculty in the department of Operations Research and Financial Engineering at Princeton University as an assistant professor. He received a Berkeley-France fund in 2006 and an NSF CAREER award in 2011.
Research Interests
Philippe has developed new tools for the theory of aggregation, which allows a better understanding of finite sample properties of stochastic optimization and sparse prediction procedures for example. This research is at the intersection of Statistics, Machine Learning and Optimization. More recently, Philippe has been interested in understanding the statistical limitations of learning under computational constraints.