BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Scaling inverse reinforcement learning for human-compatible AI - A
 dam Gleave\, UC Berkeley
DTSTART:20181023T160000Z
DTEND:20181023T173000Z
UID:TALK113629@talks.cam.ac.uk
CONTACT:Adrià Garriga Alonso
DESCRIPTION:Inverse reinforcement learning (IRL) is a technique for inferr
 ing human preferences from demonstrations of a target behaviour. Classical
  approaches make strong assumptions on human rationality\, are designed fo
 r only a single agent and do not scale to high-dimensional environments. I
 n this talk\, Adam Gleave will discuss recent work by himself and collabor
 ators scaling inverse reinforcement learning to video games\, demonstratio
 ns from multiple users with differing preferences\, and the very hard prob
 lem of learning from users with cognitive biases. The talk will be based o
 n "Inverse reinforcement learning for video games":https://gleave.me/paper
 s/2018-video-irl.pdf\, "Multi-task Maximum Causal Entropy Inverse Reinforc
 ement Learning":https://arxiv.org/abs/1805.08882 and "Inferring Reward Fun
 ctions from Demonstrators with Unknown Biases":https://openreview.net/foru
 m?id=rkgqCiRqKQ&noteId=rkgqCiRqKQ.\n\nAdam is a PhD student at UC Berkeley
  working with Stuart Russell in the Center for Human-Compatible AI. After 
 his talk\, we will have time to discuss with him how he started working in
  alignment\, what are the most promising approaches and ways of getting in
 volved\, and more.\n\nThere will *definitely* be snacks and pizza.
LOCATION: Cambridge University Engineering Department\, CBL Seminar room B
 E4-38.  See https://www.openstreetmap.org/#map=18/52.19804/0.11969
END:VEVENT
END:VCALENDAR