Research Seminar on AI: Causality in Learning-based Control
Dienstag, 21.09.2021, 16.00 Uhr
Causality is a fundamental notion in science. It tells us, how different processes influence each other. In engineering, a causal analysis can reveal what effects different input signals we apply to a machine will have. Especially for large-scale systems, such as the power grid or power plants, deriving such causal chains from first principles is a difficult problem and even small mistakes can lead to catastrophic events. Causality becomes even more important when we consider learning-based control. Standard approaches for learning-based control typically learn control policies or physical models from data through a correlation-based analysis. Thus, machine learning models may confuse cause and effect. When such algorithms are deployed in autonomous systems that act in the real world, e.g., autonomous vehicles or mobile robots, confusing cause and effect can again lead to severe accidents. In this talk, we discuss examples that demonstrate the importance of introducing a proper notion of causality in learning-based control. We then present an algorithm that enables autonomous systems to automatically reason about the causes of their actions. The algorithm provides theoretical guarantees and we demonstrate its applicability on a real-world robot manipulator and a simulated tank system.
Dominik Baumann is a postdoctoral researcher at the Institute for Data Science in Mechanical Engineering (DSME) at RWTH Aachen University. Before joining DSME, he studied electrical engineering at TU Dresden where he graduated in 2016. Afterwards, he became a joint PhD student with the Max Planck Institute for Intelligent Systems and KTH Stockholm. He defended his PhD thesis in 2020. Dominik's research focuses on resource-efficient control of (typically wireless) cyber-physical systems and automatic causal inference.