AI & Fundamentals

A tour of distributional reinforcement learning - Marc G. Bellemare, Research Scientist, Google Brain, Mila

Please register for this event here.

Abstract:

In reinforcement learning, decisions are typically made by maximizing the agent's expected sum of future rewards, or expected return. Where the reward encodes an immediate notion of utility, the return corresponds to the long-term value of the agent's decisions. Reinforcement learning algorithms are concerned with estimating the expected return from sample interactions, and making optimal decisions on the basis of these estimates. Distributional reinforcement learning extends these ideas to the realm of probability distributions and studies the return as a probabilistic object. In doing so, we find a wealth of new algorithmic tools, ranging from the abstract to the truly applied. This talk will survey the core principles behind distributional reinforcement learning, algorithmic implementations, and its recent successes in applications such as video game-playing and robotics.

Bio:

Marc G. Bellemare leads the reinforcement learning efforts at Google Research in Montreal and holds a Canada CIFAR AI Chair at the Quebec Artificial Intelligence Institute (Mila). He received his Ph.D. from the University of Alberta, where he developed the highly-successful Arcade Learning Environment benchmark. From 2013 to 2017 he held the position of research scientist at DeepMind in London, UK, where he made major contributions to deep reinforcement learning, in particular pioneering the distributional method. Marc G. Bellemare is also a CIFAR Learning in Machines & Brains Fellow and an adjunct professor at McGill University.

Please register for this event here.