Christian Schilling

(Department of Computer Science, RPTU)
hosted by Department of Computer Science

"Safety verification of cyber-physical systems with learned components"

Cyber-physical systems (CPS) consist of digital (cyber) components embedded in an analog (physical) environment. CPS are widely deployed in our interconnected world, from medical devices to robots and cars. The digital component typically takes the role of a controller that steers the physical system to a desired behavior. Due to the complexity of interacting analog and digital components, constructing such controllers is inherently challenging. In practice, machine learning has emerged as one approach to their construction. Since CPS often appear in safety- or mission-critical contexts, they must not exhibit undesirable behaviors. While traditionally, the task of safety assurance fell to the engineers who designed the controller, learned controllers are black boxes without any guarantee about their worst-case behavior. In this talk, we will present recent developments of automatic and formal analysis tools for CPS, with a particular focus on learned components, and demonstrate how these tools exhaustively analyze the system and formally prove the absence of undesirable behavior. This gives engineers a tool to certify the correctness of CPS, thus unleashing the full potential of machine-learned controllers in safety-critical contexts. We also present our vision to embed these algorithms in a larger context for safe-by-design machine learning.

Time:	Thursday, 11.01.2024, 15:30
Place:

Termin als Datei downloaden und in den Kalender importieren.