BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Dylan Foster (MIT) DTSTART:20200925T150500Z DTEND:20200925T160500Z DTSTAMP:20260404T131146Z UID:sss/6 DESCRIPTION:Title: Separating Estimation from Decision Making in Contextual Bandits\nb y Dylan Foster (MIT) as part of Stochastics and Statistics Seminar Series\ n\n\nAbstract\nThe contextual bandit is a sequential decision making probl em in which a learner repeatedly selects an action (e.g.\, a news article to display) in response to a context (e.g.\, a user’s profile) and recei ves a reward\, but only for the action they selected. Beyond the classic e xplore-exploit tradeoff\, a fundamental challenge in contextual bandits is to develop algorithms that can leverage flexible function approximation t o model similarity between contexts\, yet have computational requirements comparable to classical supervised learning tasks such as classification a nd regression. To this end\, we provide the first universal and optimal re duction from contextual bandits to online regression. We show how to trans form any oracle for online regression with a given value function class in to an algorithm for contextual bandits with the induced policy class\, wit h no overhead in runtime or memory requirements. Conceptually\, our result s show that it is possible to provably separate estimation and decision ma king into separate algorithmic building blocks\, and that this can be effe ctive both in theory and in practice. Time permitting\, I will discuss ext ensions of these techniques to more challenging reinforcement learning pro blems.\n LOCATION:https://stable.researchseminars.org/talk/sss/6/ END:VEVENT END:VCALENDAR