BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20250506T201734EDT-8854VFMrhs@132.216.98.100
DTSTAMP:20250507T001734Z
DESCRIPTION:Eric B. Laber\, PhD\n\nJames B. Duke Distinguished Professor\n	D
 ept of Statistical Sciences and\n	Biostatistics and Bioinformatics |\n	Duke 
 University\n\nWHEN: Wednesday\, October 9\, 2024\, from 3:30 to 4:30 p.m.
 \n	WHERE: Hybrid | 2001 51³Ô¹ÏÍøCollege Avenue\, Room 1201\; Zoom\n	NOTE: Eri
 c Laber will be presenting in-person\n\nAbstract\n\nContextual bandit mode
 ls are a primary tool for sequential decision making with applications ran
 ging from clinical trials to e-commerce. While there are numerous bandit a
 lgorithms which achieve optimal regret bounds and show strong performance 
 on benchmark problems\, algorithm selection and tuning in any given applic
 ation remains a major open problem. We propose the Bayesian Basket of Band
 its (B3)\, a meta-learning algorithm which automatically ensembles a set (
 basket) of candidate algorithms to produce an algorithm which dominates al
 l those in the basket. The method works by treating the evolution of a ban
 dit algorithm as a Markov decision process in which the states are posteri
 or distributions over model parameters and subsequently applying approxima
 te Bayesian dynamic programming to learn an optimal ensemble. We derive bo
 th Bayesian and frequentist convergence results for the cumulative discoun
 ted utility. In simulation experiments\, the proposed method provides lowe
 r regret than state-of-the-art algorithms including Thompson Sampling\, up
 per confidence bound methods\, and Information-Directed sampling.\n\nSpeak
 er bio\n\nPlease visit: https://laber-labs.com\n
DTSTART:20241009T193000Z
DTEND:20241009T203000Z
SUMMARY:Bayesian Basket of Bandits
URL:/epi-biostat-occh/channels/event/bayesian-basket-b
 andits-359849
END:VEVENT
END:VCALENDAR