Supervised Learning for Controlled Dynamical System Learning
We develop a framework for reducing the identification of controlled dynamical systems to solving a small set of supervised learning problems. We do this by adapting the two-stage regression framework proposed in (Hefny et. al. 2015) to controlled systems, which are more subtle than uncontrolled systems since they require a state representation that tolerates changes in the action policy. We then use the proposed framework to develop a non-parametric controlled system identification method that approximates the Hilbert-Space Embedding of a PSR (HSE-PSR) using random Fourier features, resulting in significant gains in learning speed. We also propose an iterative procedure for improving model parameters given an initial estimate. We report promising results on multiple experiments.
READ FULL TEXT