Vortragende: Prof. Dr. Tanya Braun (Uni Münster), Dr. Raphael Keusch (ETH Zürich)
Ort: Fraunhofer-Einrichtung für Individualisierte und Zellbasierte Medizintechnik IMTE, Mönkhofer Weg 239a, 23562 Lübeck
Organisatoren: DFKI (Ralf Möller), Fraunhofer IMTE (Philipp Rostalski)
Prof. Dr. Tanya Braun
Our day-to-day life is characterised by uncertainty that we as humans implicitly deal with. Probabilistic graphical models like Bayesian networks or Markov networks and factor graphs explicitly model these uncertain processes to allow for probabilistic inference. These models however do not have a notion of individuals and relations among them. And our world exists of things related to things like humans connected to humans through family, work, or friendship. Combining probabilistic models with a notion of relations opens up a whole new world of possibilities: on a technical level, it allows for tractable probabilistic inference. Beyond that, it provides exciting new inference problems like asking for the most probable source of an observation and enables to reach meta-goals such as privacy-preserving probabilistic inference. This talk covers how relations help solve known and new problems in probabilistic inference.
Composite NUV Priors for Applications in Constrained Control
Dr. Raphael Keusch
Normals with unknown variance (NUV) can represent many useful priors including Lp norms and other sparsifying priors, and they blend well with linear-Gaussian models and Gaussian message passing algorithms. In this talk, we elaborate on discretizing NUV priors and recently proposed NUV representations of half-space constraints and box constraints. We discuss the usefulness of such NUV representations with several exemplary applications in model predictive control. In such applications, the computations boil down to iterations of Kalman-type forward-backward recursions, with a complexity (per iteration) that is linear in the planning horizon. In consequence, this approach can handle long planning horizons, which distinguishes it from the prior art. For nonconvex constraints, this approach has no claim to optimality, but it is empirically very effective.