FOR 5785:  Active Learning for Systems and Control (ALeSCo) -- Data Informativity, Uncertainty, and Guarantees

• Subject Area: Computer Science, Systems and Electrical Engineering
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 535860958

In recent years, data- and learning-based approaches have gained rapidly increasing attention in systems and control. Such methods are of interest when (partially) unknown dynamical systems are operated in highly complex and uncertain environments. Machine learning can then be used, e.g., to acquire a system model or to directly learn a feedback controller. Compared to standard applications of machine learning that typically lead to static problems, many additional challenges arise when using learning-based concepts in the context of dynamical systems. For example, safety guarantees are crucial in many applications such as autonomous driving, human-robot interaction, energy systems, etc. Traditional machine learning techniques typically fail to provide such certificates, necessitating the development of novel methods and leveraging structured systems-and-control approaches towards provable guarantees. The main goal of this research unit is the development of fundamentally new approaches towards active learning for dynamical systems and their control, where the involved learning process is actively incentivized or triggered. This is in contrast to much of the existing work on learning-based control, where offline or online data are used in a learning framework without active components. Such active learning strategies are of paramount importance for ensuring safe and high-performance operation of complex and uncertain systems, while being as data-efficient as possible. To this end, this research unit considers the questions of what, when, and how to actively learn. In particular, we study what to actively learn for different problems such as model learning and control design. Furthermore, we analyze when active learning should take place, e.g., when the current data is not informative enough, model uncertainty is too high, or when control performance is not satisfactory. Finally, different methods are designed to determine how to actively learn. We examine both different learning techniques (namely neural networks, Gaussian processes, and Koopman-based methods), as well as different mechanisms for active learning such as implicit or explicit ones. In this context, the informativity of data points plays a crucial role. In general, the active learning components should be such that as much information as possible is retrieved from the collected data. Within this research unit, we develop and study different measures of data informativity and leverage them for active learning. Moreover, in order to ensure safe and reliable operation of dynamical systems, the developed methods admit provable guarantees. This includes, e.g., a quantification of error bounds or guarantees on the obtained learning rates and on closed-loop stability, robustness and constraint satisfaction. Finally, we aim to develop efficient numerical algorithms that allow the successful application of the developed methods in benchmark examples from robotics and energy systems.

• DFG Programme: Research Units

Projects

• Active learning for stochastic economic model predictive control - nonlinear optimization formulations and numerical methods (Applicant Diehl, Moritz )
• Active learning in the Koopman framework (Applicant Worthmann, Karl )
• Benchmarks for Active Learning in Systems and Control (Applicants Faulwasser, Timm ;  Hirche, Sandra )
• Coordination Funds (Applicant Müller, Matthias )
• Neural network training via persistence of excitation (Applicant Müller, Matthias )
• Neural ODE training via stochastic control and uncertainty quantification (Applicant Faulwasser, Timm )
• Safe active learning control with Gaussian processes (Applicant Hirche, Sandra )

• Spokesperson: Professor Dr.-Ing. Matthias Müller