ESIC Seminar Drgona April 30

Differentiable Programming for Data-driven Modeling, Optimization, and Control
This talk will present a different programming perspective for physics-informed machine learning (PIML) of dynamical system models, learning to optimize, and learning to control methods. We will discuss the opportunity to develop a unified PIML framework by leveraging the conceptual similarities between these distinct approaches. Specifically, we introduce differentiable predictive control (DPC) as a sampling-based learning to control method that integrates the principles of parametric model predictive control (MPC) with physics-informed neural networks (PINNs). We also show how to use recent developments in control barrier functions and neural Lyapunov functions to obtain online performance guarantees for learning-based control policies. We demonstrate the performance of these PIML methods in a range of simulation case studies, including modeling of networked dynamical systems, robotics, building control, and dynamic economic dispatch problem in power systems.