Boris Houska

Shanghai Tech, China

Monday, August 12, 2019, 11:00 - 12:00

Room 01-012, Georges-Köhler-Allee 102, Freiburg 79110, Germany

The first part of this talk presents a moment based approach for supervised statistical learning with applications in system identification and control. For this aim, we introduce a novel class of generating functions for analyzing the moments of the posterior distribution of Bayesian updates. These functions enable us to develop computational algorithms that can learn general nonlinear models from streaming data by maintaining a sequence of moments via Bayesian inference and generalized unscented propagation without ever approximating the underlying probability distributions directly.

A second part of the talk introduces a novel open-source software, named MBL-Toolbox, which implements a generic tool for moment based learning (MBL) and which is scheduled to be released publicly in the near future. We provide a crash course on how to use our current beta-version of this new tool discussing a number of tutorial problems for Bayesian learning and nonlinear dynamic system identification. Here, we focus on developing an intuition of what the advantages and disadvantages of the Bayesian viewpoint are compared to traditional least-squares based parameter estimation techniques, as, for example, used in modern moving horizon estimators for nonlinear systems.