Deep reinforcement learning and an integral
Speaker: Thomas Schön Professor at Uppsala Universitet
Thomas B. Schön is Professor of the Chair of Automatic Control in the Department of Information Technology at Uppsala University. He received the PhD degree in Automatic Control in Feb. 2006, the MSc degree in Applied Physics and Electrical Engineering in Sep. 2001, the BSc degree in Business Administration and Economics in Jan. 2001, all from Linköping University. He has held visiting positions with the University of Cambridge (UK), the University of Newcastle (Australia) and Universidad Técnica Federico Santa María (Valparaíso, Chile). He is a Senior member of the IEEE and an Associate Editor of Automatica. He received the Arnberg prize by the Royal Swedish Academy of Sciences in 2016. He was awarded the Automatica Best Paper Prize in 2014, and in 2013 he received the best PhD thesis award by The European Association for Signal Processing (EURASIP). He received the best teacher award at the Institute of Technology, Linköping University in 2009.
The amount of data being collected from a vast range of systems is increasing at a rapid pace, and it is critical that our ability to reason and act based on this data keeps pace. Mathematical models are a key enabling technology: they provide a compact and interpretable representation of the data capturing the most important features of the underlying system. I will try to provide some insights into the current deep learning trend/hype via some work that we have done recently on data-efficient reinforcement learning in continuous state-action spaces using rather high-dimensional observations. We consider the so-called pixels-to-torques problem, where an agent learns a closed-loop control policy (“torques”) from pixel information only. We introduce a data-efficient, model-based reinforcement learning algorithm that learns such a closed-loop policy directly from pixel information. The key ingredient is a deep dynamical model for learning a low-dimensional feature embedding of images jointly with a predictive model in this low-dimensional feature space. As a separate topic I will towards the end show how we can render photorealistic images by solving a particular integral using a relatively new trick, that is useful in quite a few settings.