DFG: Active Learning for Systems and Control
ALeSCo is a collaborative research project on Active Learning for Systems and Control by the German Research Foundation“ (DFG).
Motivation
Recent advances in data science and machine learning have transformed numerous domains and are now poised to reshape automatic control, a priority highlighted in the IEEE Control Systems Society’s “Road Map 2030.” The proposed research unit aims to devise fundamentally new approaches to active learning for dynamical systems and their control, where the learning process is deliberately triggered rather than passively relying on pre-collected data. Central questions concern what to learn (system models, controllers), when to learn (balancing exploration and exploitation), and how to learn (selecting a proper inductive bias). To achive that, data-informativity is treated as a unifying thread: measures such as persistency of excitation, uncertainty reduction, or risk minimization will be analyzed and related to building a holistic framework.
Ensuring safety and reliability is imperative; therefore, the methods will provide guarantees on learning rates, error bounds, closed-loop stability, robustness, and constraint satisfaction. Because active learning operates online, the consortium will also create computationally efficient algorithms that exploit problem structure.
Research Topics
For 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 and Karl Worthmann from TU Ilmenau as well Mercator Fellow Armin Lederer from NU Singapore.
ALeSCo covers topics from data-informativity, over modeling with neural networks, Gaussian Processes, or evolution operators to numerical optimization and benchmarking. It is aimed at exploring commonalities, synergies, and develop algorithms with the whole control pipeline in mind. In this context ITR will lead the research on
P4: Safe active learning control with Gaussian processes
The main goal of this project is developing a novel active learning method for Gaussian process (GP) models in learning-based control. For this purpose, novel control-oriented data informativity measures are developed and employed for exploration with guarantees from a Bayesian perspective.
And, in cooperation with TU Hamburg, ITR will develop a novel benchmark for active learning in control
P6: Benchmarks for Active Learning in Systems and Control
This project develops new, challenging open problems and benchmark suites for transparent evaluation of active learning methods tailored to dynamic systems, closing the gap in control-specific benchmarks. The benchmark covers energy systems and robotics and includes scalable tasks, datasets, uncertainty descriptions, and comparison metrics, enabling long-term, consistent method assessment.