Workshop @DYCOPS: Nonlinear model predictive control for anything

Neural networks are a very promising tool to implement complex control algorithms. They are extremely fast to evaluate, rendering real-time control of challenging applications possible. Their deployment on embedded hardware is simple, due the user-defined memory size, which allows to adjust the architecture on the memory requirements, as well as the simple evaluation of the neural network controllers, which just involves nonlinear and linear function evaluations, that can be implemented on almost any hardware. Furthermore, almost any arbitrary control algorithm on any scale can be represented by a neural network because of their general expressiveness. However, finding a suitable parametrization is very challenging and existing methods such as reinforcement learning struggle due to data inefficiencies and limited use of knowledge about the control task.

To alleviate these challenges, it is possible to exploit synergies between nonlinear model predictive control (NMPC) and neural networks that try to imitate the NMPC behavior. This approach provides a very promising baseline for neural network control, as it includes physical knowledge about the controlled system and allows for application of tools from control theory and optimization to achieve optimal performance under consideration of constraints and uncertainties.

The objective of this workshop is to provide a hands-on overview from neural network controller design to deployment on real-world applications using robust NMPC as a baseline to be imitated. To this end, this workshop will first give an overview of the large body of current research on neural network-based control based on nonlinear MPC. A short hands-on introduction to robust NMPC and its implementation using open-source tools will be presented. A tutorial on supervised and unsupervised imitation learning approaches to synthesize the neural network controller from the robust NMPC baseline will be covered, also including practical considerations of efficient data-sampling strategies and an introduction on probabilistic safety and performance guarantees. Finally, this workshop will showcase the deployment on embedded hardware and application to real-world problems.

As a result, we expect that the attendants to this workshop have an overview of the possibilities of using neural networks to implement approximate MPC controllers, as well as gain detailed insights on the most important parts for its safe design and efficient implementation. This outcome will enable the deployment of (approximate) NMPC on virtually any system, since computational restrictions are not any more relevant, even in case where uncertainties and therefore complex robust NMPC schemes need to be considered.

Joshua Adamek

Joshua Adamek graduated with a M.Sc. degree in electrical engineering from TU Dortmund University in 2022. Since October 2022, he is a PhD student at the chair of process automation system, the group of Prof. Lucia. His research interests lies in combining machine learning and control. He is currently working on the DFG research project Opti-FAAS which deals with optimization of Cloud scheduling tasks.

Affiliation: TU Dortmund University

E-Mail: joshua.adamek@tu-dortmund.de 

Joel Paulson

Joel A. Paulson is currently the H.C. “Slip” Slider Associate Professor of Chemical and Biomolecular Engineering at The Ohio State University (OSU) where he is also a core faculty member of the Sustainability Institute and an affiliate of the Translational Data Analytics Institute. He will be starting as the Battist Associate Professor in the Chemical and Biological Engineering Department at the University of Wisconsin-Madison in August of 2025. He has received several awards including the Best Application Paper Prize from the 2020 IFAC World Congress, the NSF CAREER Award, and the Lumley Research Award. His research interests are in data-driven optimization, physics-informed machine learning, and model predictive control. Methods developed by the Paulson group are being applied to a variety of next-generation biochemical systems including continuous pharmaceutical manufacturing, chemical looping combustion, sustainable battery systems, and non-equilibrium plasma jets.

Affiliations: The Ohio State University, University of Wisconsin-Madison

E-Mail: paulson.82osuedu

Ali Mesbah

Ali Mesbah is Associate Professor of Chemical and Biomolecular Engineering at the University of California at Berkeley. Dr. Mesbah is a Senior Member of the IEEE and AIChE. His research interests include learning-based analysis and control of uncertain systems, with applications to materials processing and manufacturing systems.

Affiliation: University of California, Berkeley

E-Mail: mesbahberkeleyedu

Sergio Lucia

Sergio Lucia is a Full Professor at TU Dortmund University, Germany. He received his M.Sc. degree in electrical engineering from the University of Zaragoza, Spain, in 2010, and the Dr. Ing. degree in optimization and automatic control from the TU Dortmund University, Germany, in 2014. He joined the Otto-von-Guericke Universität Magdeburg and visited the Massachusetts Institute of Technology as a Postdoctoral Fellow. He was an Assistant Professor at TU Berlin between 2017 and 2020. Since 2020, he has been a Professor at TU Dortmund University and head of the Chair of Process Automation Systems. His research interests include decision-making under uncertainty, distributed control, as well as the interplay between machine learning techniques and control theory.

Affiliation: TU Dortmund University

E-Mail: sergio.luciatu-dortmundde