BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:EE Distinguished Lecturer Seminar: Towards Safe Reinforcement Lear
 ning
DTSTART:20190412T131500
DTEND:20190412T141500
DTSTAMP:20260511T162540Z
UID:dcf7ac2f905847c97bf3166621854da61f77bf155d7f98d5a40f2ecd
CATEGORIES:Conferences - Seminars
DESCRIPTION:Andreas Krause is a Professor of Computer Science at ETH Zuric
 h\, where he leads the Learning & Adaptive Systems Group. He also serves a
 s Academic Co-Director of the Swiss Data Science Center. Before that he wa
 s an Assistant Professor of Computer Science at Caltech. He received his P
 h.D. in Computer Science from Carnegie Mellon University (2008) and his Di
 plom in Computer Science and Mathematics from the Technical University of 
 Munich\, Germany (2004). He is a Microsoft Research Faculty Fellow and a K
 avli Frontiers Fellow of the US National Academy of Sciences. He received 
 ERC Starting Investigator and ERC Consolidator grants\, the Deutscher Must
 ererkennungspreis\, an NSF CAREER award\, the Okawa Foundation Research Gr
 ant recognizing top young researchers in telecommunications as well as the
  ETH Golden Owl teaching award. His research on machine learning and adapt
 ive systems has received awards at several premier conferences and journal
 s. Andreas Krause served as Program Chair for ICML 2018\, and is serving a
 s Action Editor for the Journal of Machine Learning Research.\nAbstract: M
 ore and more machine learning systems make data-driven decisions in the re
 al work\, in increasingly higher-stakes applications.  This has caused su
 bstantial interest in reinforcement learning -- the field of learning to m
 ake decisions from data -- which has seen stunning recent empirical breakt
 hroughs. At its heart is the challenge of trading exploration -- collectin
 g data for learning better models -- and exploitation -- using the estimat
 e to make decisions.  In many applications\, however\, exploration is a p
 otentially dangerous proposition\, as it requires experimenting with actio
 ns that have unknown consequences.  Hence\, most prior work has confined 
 exploration to simulated environments.  In this talk\, I will present our
  work towards rigorously reasoning about safety of exploration in reinforc
 ement learning.  I will discuss a model-free approach\, where we seek to 
 optimize an unknown reward function subject to unknown constraints.  Both
  reward and constraints are revealed through noisy experiments\, and safet
 y requires that no infeasible action is chosen at any point. I will also d
 iscuss model-based approaches\, where we learn about system dynamics throu
 gh exploration\, yet need to verify safety of the estimated policy.  Our 
 approaches use Bayesian inference over the objective\, constraints and dyn
 amics\, and -- under some regularity conditions -- are guaranteed to be bo
 th safe and complete\, i.e.\, converge to a natural notion of reachable op
 timum.  I will also show experiments on safely tuning cyber-physical syst
 ems in a data-driven manner.\n 
LOCATION:ELA 2 https://plan.epfl.ch/?room==ELA%202
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR