Dinesh Krishnamoorthy

Department of Mechanical Engineering, Eindhoven University of Technology (TU/e)

Tuesday, October 22, 2024, 11:00 - 11:59

Room 01-012, Georges-Köhler-Allee 102, Freiburg 79110, Germany

Reducing the computational complexity of Model Predictive Control (MPC) is critical for enabling real-time control in various engineering applications, particularly in distributed MPC and mixed-integer MPC. Machine learning offers attractive tools that can be leveraged for complexity reduction in MPC. A key aspect in the broad field of “Learning-based MPC” is what is being learned and how. This talk presents essential methodologies, advantages, and challenges in integrating learning into MPC for real-time decision-making, especially under the context of distributed MPC and mixed integer MPC. Specifically, we examine frameworks that learn the value function to build myopic policies with probabilistic closed-loop guarantees. Additionally, we introduce efficient data generation strategies through sensitivity-based data augmentation techniques, showcasing their role in building learning-based MPC controllers.