Datenschutzerklärung|Data Privacy

A. Borusan

20.01.2020, 5p.m./c.t. HHI lecture hall, Einsteinufer, ground floor.: "Shift your Expectations - Covariate Shift in Reinforcement Learning" (Prof. Dr. Alexander Smola - Director, Amazon Web Services)

In this talk I will give an overview over covariate shift, its correction and how it applies to reinforcement learning algorithms. RL is particularly sensitive to this phenomenon since a change in distribution is directly the result of an updated policy. That is, any update step in RL immediately results in a shift and it requires careful correcting. I will give an example of how this can lead to significant improvement in the performance of policy gradient algorithms, simply by making better use of the off-policy replay buffer. A second application illustrates how this applies to meta-learning, yielding a principled alternative to MetaRL heuristics. This is joint work with Rasool Fakoor, Pratik Chaudhari, and Stefano Soatto.