Location and Venue: Saturday, 10.10.2015, 11.40-12.30 Informatikzentrum, RWTH Aachen.
Representation Learning for Control
Dr. Joschka Boedecker
(University of Freiburg, Germany)
Defining features for control learning tasks with high-dimensional inputs is challenging. A balance needs to be found between compressing the state-dimensionality for fast convergence of the learning algorithm on the one hand, and retaining enough information about the full state of the system on the other. Learning features for these tasks automatically from data has received increasing interest lately. In this talk, I will review some of these approaches, highlighting our recent "Embed to Control" method which learns a latent representation and a locally linear transition model that facilitates optimal control by design.
