UNIVERSITY OF HERTFORDSHIRE
	    COMPUTER SCIENCE RESEARCH COLLOQUIUM

			   presents


   "Learning Probabilistic Models of Atari 2600 Games"

              	   Dr. Joel Veness
         (DeepMind Technologies Ltd., London, UK and
	        University of Alberta, Canada)

                    8 May 2013  (Wednesday)
			  1 pm -2 pm

		  Hatfield, College Lane Campus
		 * * Lecture Theatre LC108  * *

		 Everyone is Welcome to Attend

		 Refreshments will be available

Abstract:

In this talk, I will discuss some recent information theoretic techniques
for learning probabilistic models of large reinforcement learning
environments. In particular, I will describe some work that attempts to
learn models of arbitrary Atari 2600 games from raw video data. Along
the way, I will discuss a number of tricks for doing efficient,
exact Bayesian model averaging over various kinds of combinatorial
spaces. These techniques are efficient, come with strong competitive
(regret) guarantees, and allow us to build surprisingly good models of
quite a number of simple arcade games. I will conclude by discussing the
limitations of our current model, and outline some of my recent efforts
to improve it.

---------------------------------------------------
Hertfordshire Computer Science Research Colloquium
http://cs-colloq.stca.herts.ac.uk