Distributional Reinforcement Learning

Marc G. Bellemare(Google Brain) - Google Brain

Jan. 18, 2019, 2:30 p.m. - Jan. 18, 2019, 3:30 p.m.

Trottier 2120


Abstract:

In reinforcement learning, one typically predicts the expected sum of future rewards, or return. Distributional reinforcement learning takes this idea a step further by predicting the full distribution of the random return. By now, distributional reinforcement learning is well-established, achieving state-of-the-art performance in a number of reinforcement learning benchmarks. This talk reviews the main ideas and formalisms underlying the method, provides an overview of practical algorithms and empirical evidence that distributional predictions are useful, and presents recent results that shed light on exactly why the method works so well.

Speaker Bio:

Marc G. Bellemare is a research scientist at Google Brain in Montreal. He is also adjunct professor at McGill University, a CIFAR Learning in Machines & Brain Fellow, and holds a Canada CIFAR AI chair at the Mila. He received his Ph.D. from the University of Alberta where he studied the concept of domain-independent agents and built the highly-successful Arcade Learning Environment, the platform for AI research on Atari 2600 games. He is known for his work on reinforcement learning, including approximate exploration, representation learning, and the distributional method.