BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Talks.cam//talks.cam.ac.uk// X-WR-CALNAME:Talks.cam BEGIN:VEVENT SUMMARY:BSU Seminar: "\;Efficient Sequential Experimentation: Bridging Model-Based Reinforcement Learning and Bayesian Optimal Experimental Desi gn"\; - Alberto Caron\, The Alan Turing Institute DTSTART:20250204T140000Z DTEND:20250204T150000Z UID:TALK227854@talks.cam.ac.uk CONTACT:Alison Quenault DESCRIPTION:Sequential Experimental Design (SED) aims to adaptively collec t data that most efficiently reduce uncertainty in a statistical model of interest in a multi-stage context\, where each stage’s choice depends on the knowledge gained from previous ones. Although this is a classic task in many applied disciplines\, recent advances in Reinforcement Learning (R L) offer interesting new perspectives for approaching this challenge. In t his talk\, I will first discuss how the Bayes-Adaptive Markov Decision Pro cesses (BAMDPs) framework provides a principled way to balance exploration and exploitation in RL when key parameters are uncertain\, and how it nat urally lends itself to model multi-stage SED through a parametrized “sam pling policy” that selects actions/experiments based on the current cont ext. Building on ideas from Model-Based Reinforcement Learning (MBRL)\, I will then present a method for picking design choices in SED by “looking ahead” and maximizing a Cumulative Expected Information Gain (C-EIG) ob jective over H-steps\, which can be approximated in the context of MBRL-ba sed planning via a measure of disagreement among an ensemble of probabilis tic dynamics models. By explicitly modelling how each design choice affect s downstream data-collection\, this approach avoids myopic behaviour and e nables the selection of actions/experiments that are most informative in e xpectation not just immediately\, but across an entire H-steps sequence of trials. LOCATION:Large Seminar Room\, East Forvie Building\, Forvie Site Robinson Way Cambridge CB2 0SR. END:VEVENT END:VCALENDAR