BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Misleading meta-objectives and hidden incentives for distributiona
 l shift - Paolo Bova (University of Cambridge)
DTSTART:20190508T160000Z
DTEND:20190508T180000Z
UID:TALK124699@talks.cam.ac.uk
CONTACT:Adrià Garriga Alonso
DESCRIPTION:This week:\n"Misleading meta-objectives and hidden incentives 
 for distributional shift." David Krueger\, Tegan Maharaj\, Shane Legg and 
 Jan Leike. ["Paper":https://drive.google.com/uc?export=download&id=1k93292
 JCoIHU0h6xVO3qmeRwLyOSlS4o] \n["BibTeX":https://drive.google.com/uc?export
 =download&id=1NAdOSvjJEzD0Ba2i6Mwfhtk6Ad77WsQ_]\n\nThe authors aim to show
  that Meta-Learning can create hidden incentives for agents to change thei
 r task rather than solving the task we tell them to. An example would be a
 n agent that predicts when someone wants coffee: after learning that the p
 erson has coffee in the morning they learn to wake them up when they try t
 o sleep in\, so following a seemingly suboptimal policy (wake up the human
 ) results in a better prediction. Their paper runs experiments to show tha
 t Meta-Learning agents with Population-Based Training (PBT) learn to exhib
 it non-myopic behaviour even when their reward is myopic. They also demons
 trate for these agents a method for eliminating this non-myopic behaviour 
 that they call Environment Swapping.\n\nAs always\, there will be free piz
 za. The first half hour is for stragglers to finish reading.\n\nInvite you
 r friends to join the mailing list (https://lists.cam.ac.uk/mailman/listin
 fo/eng-safe-ai)\, the Facebook group (https://www.facebook.com/groups/1070
 763633063871) or the talks.cam page (https://talks.cam.ac.uk/show/index/80
 932). Details about the next meeting\, the week’s topic and other events
  will be advertised in these places.\n
LOCATION:Engineering Department\, CBL Seminar room BE4-38
END:VEVENT
END:VCALENDAR