Estimation, Learning, and Control of Autonomous Systems

Abstract:

Machine learning as a concept has been widely researched over the last decades and is currently being used in several industrial implementations, such as in speech recognition and image analysis. However, machine learning has so far had relatively limited success in the control community. One reason for this is that in control there is typically a need for a, possibly uncertain, dynamic model of the controlled system in order to reason about properties such as performance and convergence. In recent years, there has been increased focus on overcoming this bottleneck. This talk addresses the joint state estimation and learning problem for nonlinear dynamical systems, where a focus is to be able to use the estimation and learning for real-time control. Specifically, we will discuss how Bayesian state inference and parameter learning can be combined with model-predictive control (MPC) to provide a method for vehicle control that in real-time propagates information about the available tire friction to the MPC. On a similar note, we will discuss a recent approach for real-time joint state estimation and learning of dynamical systems. The method leverages a recently developed reduced-rank formulation of Gaussian-process state-space models (GP-SSMs), and results in a recursive formulation for updating the sufficient statistics associated with the GP-SSM by exploiting marginalization and conjugate priors. The method provides probabilistic confidence---not only of the state estimate, but also of the model that was used to determine the state estimate. We argue that these confidence measures are directly suitable for incorporating into control.

Biography:Karl Berntorp received the M.Sc. degree in Engineering Physics in 2009 and the Ph.D. degree in Automatic Control in 2014, from Lund University, Lund, Sweden. In 2008 he was a visiting researcher at Daimler AG in Sindelfingen, Germany. In 2014 he joined Mitsubishi Electric Research Laboratories in Cambridge, MA. His research is on statistical signal processing, Bayesian inference, sensor fusion, and optimization-based control, with applications to automotive, aerospace, transportation, navigation, and communication systems. His work includes design and implementation of nonlinear estimation, constrained control, motion-planning, and learning algorithms. Dr. Berntorp is the author of more than 65 peer-reviewed papers in journals and conferences and 10 granted patents.