Timezone: »

Sergei Levine: Distribution Matching and Mutual Information in Reinforcement Learning
Sergey Levine

Fri Jun 14 03:30 PM -- 04:10 PM (PDT) @

Conventionally, reinforcement learning is considered to be a framework for optimization: the aim for standard reinforcement learning algorithms is to recover an optimal or near-optimal policy that maximizes the reward over time. However, when considering more advanced reinforcement learning problems, from inverse reinforcement learning to unsupervised and hierarchical reinforcement learning, we often encounter settings where it is desirable to learn policies that match target distributions over trajectories or states, covering all modes, or else to simply learn collections of behaviors that are as broad and varied as possible. Information theory and probabilistic inference offer is a powerful set of tools for developing algorithms for these kinds of distribution matching problems. In this talk, I will outline methods that combine reinforcement learning, inference, and information theory to learn policies that match target distributions and acquire diverse behaviors, and discuss the applications of such methods for a variety of problems in artificial intelligence and robotics.

Author Information

Sergey Levine (UC Berkeley)
Sergey Levine

Sergey Levine received a BS and MS in Computer Science from Stanford University in 2009, and a Ph.D. in Computer Science from Stanford University in 2014. He joined the faculty of the Department of Electrical Engineering and Computer Sciences at UC Berkeley in fall 2016. His work focuses on machine learning for decision making and control, with an emphasis on deep learning and reinforcement learning algorithms. Applications of his work include autonomous robots and vehicles, as well as computer vision and graphics. His research includes developing algorithms for end-to-end training of deep neural network policies that combine perception and control, scalable algorithms for inverse reinforcement learning, deep reinforcement learning algorithms, and more.

More from the Same Authors