Tom Lefebvre (Ghent University)

Hybrid machine learning and domain engineering knowledge for real physical machines


This talk is about taking machine learning out of simulations and bringing it to the real physical world. To integrate ML algorithms into real machines requires them to be robust, explainable and trustworthy. This calls for a proper combination of these algorithms with state-of-the-art/use methodologies, ultimately resulting in a hybrid form of domain engineering knowledge and machine learning. We’ll show how supervised learning can be combined with classical physics-based models for accurately identifying system behavior and how synergies can be found between reinforcement learning and nonlinear optimal control to provide control actions.

Tom Lefebvre obtained the MSc degree and the PhD degree in electromechanical, automation and control engineering from Ghent University in 2015 and 2019, respectively. His doctoral thesis was entitled “Approximate numerical solution strategies for dynamic optimization and optimal control problems”. Currently, he’s a post-doctoral research assistant with the same research group. His main research interests include efficient numerical methods for (stochastic) optimal control, trajectory optimization, stochastic strategies for static and dynamic optimization problems and expansion-based uncertainty quantification.

Tom Lefebvre

Ghent University