---------- Forwarded message ----------
From: Doina Precup <dprecup@cs.mcgill.ca>
Date: 2017-04-10 15:32 GMT-04:00
Subject: Marc Bellemare talk Wed Apr 12, 10:30am, MC103
To: labrl@cs.mcgill.ca
Cc: "Marc G. Bellemare" <marcgb@gmail.com>, Gen Fried <genevieve.fried@mail.mcgill.ca>, Yoshua Bengio <yoshua.umontreal@gmail.com>


Hi everyone,

Marc Bellemare, who has made great contributions to the theory and practice of reinforcement learning (including the ALE environment that we all use & love) will be visiting us Wed. The talk is at 10:30am, MC103. If you want to meet with Marc, please send email to Gen (cc’d), he will be around for the day.

Title: The role of density models in reinforcement learning

Abstract: Much of the theoretical foundations of reinforcement learning assume, or derive from, a tabular representation. In practical applications, however, the tabular representation is usually impractical and undesirable. The translation to practice therefore typically involves a regression step: a projection of the value function onto a tractable function class, for example a deep network. In performing this regression, we often lose many of the appealing properties of tabular representations, including measures of value uncertainty or the ability to learn from a few examples. In this talk I will argue that a particular kind of probabilistic generative models, density models, allow us to recover the benefits of the tabular representation, without sacrificing generalization. I will first revisit the Compress and Control approach, which uses density models (or in fact any sequential data compression algorithm) to model the value function. As an example, I will demonstrate an agent that learns to play Pong in 13 games. I will subsequently present our recent work on pseudo-counts, showing how to induce intrinsically motivated behaviour from a simple density model, and how this behaviour leads to state-of-the-art exploration in one of the hardest Atari 2600 games, Montezuma's Revenge.

Short bio

Marc G. Bellemare received his MSc from McGill University and Ph.D. from the University of Alberta, where he investigated the concept of domain-independent agents and led the design of the highly-successful Arcade Learning Environment. His research interests include reinforcement learning, online learning, information theory, lifelong learning, and randomized algorithms. He is currently a Senior Research Scientist at DeepMind.

http://www.marcgbellemare.info/static/index.html

Best,
Doina