BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Cooperative Inverse Reinforcement Learning - Dylan Hadfield-Menell
 \, UC Berkeley
DTSTART:20170223T150000Z
DTEND:20170223T160000Z
UID:TALK71104@talks.cam.ac.uk
CONTACT:Adrian Weller
DESCRIPTION:In order to realize the benefits of artificial intelligence\, 
 we need to ensure that autonomous systems reliably pursue the objectives t
 hat their designers and users intend. This is the crux of the value alignm
 ent problem: jointly determining another agent's preferences and selecting
  actions in accordance with those preferences. Good strategies for value a
 lignment have implications for the utility and value of consumer robotics 
 and AI in the short term and the potential exeistential risk from superhum
 an intelligence in the long run. In this talk\, I will present a Cooperati
 ve Inverse Reinforcement Learning (CIRL)\, a novel mathematical framework 
 for value alignment. The core of the approach relies on framing the proble
 m as a game between two players (a human and a robot) with a shared object
 ive and asymmetric information about that objective. I will give an overvi
 ew of the model and present recent work that leverages this framework to a
 nalyze the incentives an agent has to defer to a human decision\, likely f
 ailure modes of misspecified preference models\, and ways to avoid acciden
 tal incentivizes for undesireable side effects.
LOCATION:CBL Room BE-438\, Department of Engineering
END:VEVENT
END:VCALENDAR