Differentiable programming for modeling and control of dynamical systems
In this talk, we will present a differentiable programming perspective on optimal control of dynamical systems. 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 scientific machine learning. We show that DPC allows for domain-aware learning of differentiable models for dynamical systems that can be used for subsequent offline policy optimization subject to nonlinear constraints. Empirically we demonstrate the DPC's scalability, data efficiency, and constraints handling in simulation case studies, including building control and dynamic economic dispatch problems. Furthermore, we experimentally demonstrate DPC's computational and memory efficiency in embedded implementation on a laboratory device with nonlinear dynamics serving as a proof of concept for control as a service setup.
Short bio: Jan is a senior data scientist and the 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, Slovakia, 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, including building control and power systems optimization.
Speaker will be in-person!
Place: Math, 501 & Zoom: https://arizona.zoom.us/j/81337180102