BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Cambridge Ellis Unit Seminar Series - Dr Silvia Chiappa-  Asymptot
 ically Best Causal Effect Identification with Multi-Armed Bandits - Dr Sil
 via Chiappa
DTSTART:20220330T130000Z
DTEND:20220330T140000Z
UID:TALK171353@talks.cam.ac.uk
CONTACT:Kimberly Cole
DESCRIPTION:This talk presents a method for selecting from a set of causal
 -effect identification formulas the one with the lowest asymptotic varianc
 e\, in a sequential setting in which the investigator may alter the data c
 ollection mechanism in a data-dependent way with the aim of finding the fo
 rmula in as few samples as possible. We formalize this setting by using th
 e best-arm-identification bandit framework where the standard goal of lear
 ning the arm with the lowest loss is replaced with the goal of learning th
 e arm that will produce the best estimate. We introduce new tools for cons
 tructing finite-sample confidence bounds on estimates of the asymptotic va
 riance that account for the estimation of potentially complex nuisance fun
 ctions\, and adapt the best-arm-identification algorithms of LUCB and Succ
 essive Elimination to use these bounds. We validate our method by providin
 g upper bounds on the sample complexity and an empirical study on artifici
 ally generated data.
LOCATION:https://eng-cam.zoom.us/j/87668854064?pwd=czNKUVZkVmZQcnROMmVWTWh
 EcE50dz09
END:VEVENT
END:VCALENDAR