Tailored neural networks for learning-based predictive control

• Applicant: Professor Dr. Moritz Schulze Darup
• Subject Area: Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 556837821

Machine learning (ML) has been used in automatic control for decades, with recent advancements in deep learning bringing more effective and versatile methodologies, particularly in data-driven model predictive control (MPC). Various ML paradigms have been deployed to enhance MPC applications, including supervised learning for offline control policy approximations and reinforcement learning for online optimization. Despite these advancements, practical control applications impose strict requirements such as real-time capability, computational efficiency, and satisfaction of process constraints. To address these challenges, this project focuses on the specification of tailored neural networks (NN) for the approximation of predictive control policies or corresponding optimal value functions. It builds on the observation that special NNs with rectified linear units (ReLU) or maxout activations possess the same piecewise affine (PWA) structure as, for example, MPC laws for linear systems. Furthermore, variants of such NNs also allow the description of convex piecewise quadratic (PWQ) functions, which, e.g., appear in the form of optimal value functions in MPC. The project aims to develop novel methods that exploit the observed relationships more efficiently in order to significantly enhance learning-based control. Central building blocks are (i) novel decompositions of maxout neurons, (ii) tailored affine liftings of PWQ functions, and (iii) extensive numerical benchmarks. A novel bridge to (iv) learning-based solver iterations concludes the project.

• DFG Programme: Research Grants