This talk will show the use of differentiable programming (DP) for domain-aware learning of differentiable models for dynamical systems. We will show how the differentiability of these models can be utilized for direct policy optimization. For that purpose, we introduce differentiable predictive control (DPC) as a data-driven model-based policy optimization method that systematically integrates the principles of classical model predictive control (MPC) with differentiable programming. We also show how to use recent developments in control barrier functions and neural Lyapunov functions to obtain online performance guarantees for data-driven control policies. Empirically we demonstrate the performance of these new DP-based methods in a range of simulation case studies, including modeling of networked dynamical systems, building control, and dynamic economic dispatch. Furthermore, we demonstrate real-time deployment of DPC on an embedded device serving as a proof of concept for control as a service setup.

Jan is a senior data scientist and principal investigator in the Physics and Computational Sciences Division at Pacific Northwest National Laboratory (PNNL). Jan has a PhD in Control Engineering from the Slovak University of Technology in Bratislava, and before joining PNNL, he was a postdoc at the mechanical engineering department at Katholieke Universiteit (KU) Leuven in Belgium. His current research is focused on differentiable programming for scientific machine learning, constrained optimization, and model-based optimal control with applications in the energy sector.

Personal webpage