Seminar: Bandit Theory and its applications to Active Learning and Stratified Monte-Carlo

SpeakerAlexandra Carpentier
DateMonday, 20 Feb 2012
Time13:00 - 14:00
Location1.20, Malet Place Engineering Building
Event seriesCSML Joint Seminar Series

Bandit theory is a powerful yet simple setting for formalising the exploration/exploitation dilemma. This setting can be used to formalise and solve in a very efficient way the problem of active learning in the context e.g. of (stratified) functional regression or also the problem of efficient stratified Monte-Carlo. The idea is that each stratum of the space is an arm of the bandit. In the case of fixed strata, it is possible to do almost as well (in the regret sense) as what an oracle knowing the function would be able to do. It is also possible to do some clever choice of the strata, either in a minimax sense, and that for different regularities of the function, or even in an active sense, i.e. select adaptively an (almost) optimal partitioning of the space.

iCalendar csml_id_12.ics