ALeSCo Research Unit: Active Learning for Systems and Control - Data Informativity, Uncertainty, and Guarantees

In recent years, data- and learning-based approaches for the control of (partially) unknown dynamic systems in highly complex and uncertain environments have become increasingly important. Machine learning can be used here, for example, to generate a system model or to directly learn a control system. While standard applications of machine learning usually lead to static problems, many additional challenges arise when using learning-based concepts for dynamic systems. For example, safety guarantees are of crucial importance in applications such as autonomous driving, human-machine interaction or energy systems. Such guarantees are generally not available for conventional machine learning methods, so that novel methods and the use of structured approaches from control engineering and systems theory are required. The main goal of the research group is the development of fundamentally new approaches to active learning for dynamic systems and their control, in which the learning process is continuously influenced. This is in contrast to the majority of existing work on learning-based control, where offline or online data is used for learning without active components. Such active learning strategies are necessary to ensure safe, performant and data-efficient operation of complex and uncertain systems.

To this end, we investigate what, when and how active learning should take place. The focus is on what needs to be learned depending on the specific problem structure (such as model generation or controller design) and when active learning should take place, e.g. if the information content of the current data is insufficient, if the model uncertainty is too high or if the control quality is insufficient. Finally, different methods of active learning are developed. We investigate different learning techniques (neural networks, Gaussian processes and Koopman methods) as well as different implicit and explicit mechanisms for active learning. For active learning, the information content of the data plays a decisive role. In principle, the active learning components should allow the extraction of the richest possible information from the collected data. Within the research group, we develop and investigate different information measures for data and their use for active learning. The newly developed methods ensure the safe and reliable operation of dynamic systems in a mathematically rigorous way. This includes, for example, guarantees regarding error bounds, achieved learning rates or stability and robustness of the closed control loop as well as fulfillment of constraints. Finally, efficient numerical algorithms are to be designed that allow the successful application of the investigated methods to benchmark systems from robotics and energy technology.

ALeSCo's principal investigators are expert control researchers from all over Germany: 

• Matthias Müller and Victor Lopez from LU Hannover
• Moritz Diehl from U Freiburg
• Sandra Hirche from TU Munich
• Timm Faulwasser from TU Hamburg
• Karl Worthmann from TU Ilmenau
• Mercator Fellow Armin Lederer from NU Singapore