Jakob Matthias Stammler, M.Sc.
- Wissenschaftlicher Mitarbeiter
- Exam date: October 6, 2026
- Group: Prof. Becker
- Room:
CS 30.10 - Phone: +49 721 608-41322
- stammler ∂does-not-exist.kit edu
Engesserstr. 5
76131 Karlsruhe
Summary of the Dissertation
This dissertation examines the challenges posed by modern, increasingly centralized vehicle architectures, which are characterized by software-defined vehicles, hardware abstraction layers, and the growing use of machine learning. To this end, a holistic scheduling approach is developed that combines classical periodic control tasks, machine learning-based applications, and dynamic operating modes within a secure time-triggered scheduling method.
The approach combines a correct-by-construction proof for safe execution with anytime neural networks and a mode-change mechanism that enables flexible responses to errors, anomalies, or changes in system state. Evaluation using real-world use cases and synthetic task sets shows that the scheduler efficiently handles even complex scenarios and is therefore suitable for practical automotive systems.
Monitoring of safety security-relevant incidents
Developing autonomous acting vehicles requires a working and reliable collaboration between functions, (sub-)systems, and services. These collaborations are not limited to the vehicle itself but continue well into the development cycle. A continuously changing landscape of cyber security risks and the fulfillment of functionality (SOTIF) require dynamic monitoring of a wide range of vehicle sensor data. Deciding which data should be monitored and saved, how this can be achieved and the integration into the next development cycle is an open question that gets many different answers frequently.
Making stochastic components certifiable
Machine learning components are becoming increasingly widespread in a wide range of contexts and achieve remarkable performance. Because of their stochastic nature, their use isn’t deterministic and they often act like a “black box”, which makes certification difficult. Creating frameworks and models to simulate or predict their behavior could yield a more controllable and formal approach of combining the high performance of ML components with the predictability of proven conventional components. These models shouldn’t be limited to ML components but are usable in different scenarios.
Graceful Degradation strategies
The failing of a non-critical subsystem shouldn’t result in a complete failure of a whole mission. Often redundant hardware components are used to reduce the mean time between failure. This at least doubles the installed hardware and makes the whole system more expensive. If the redundant hardware isn’t needed it is idle and takes up unnecessary space and power. To mitigate this, non-critical tasks could be executed on these systems, that might be canceled on demand. Creating sophisticated models and frameworks that predict the behavior of these systems is necessary for predictability and legal certification.
| title | type |
|---|---|
| Automated and split Binding of Software Components to Execution Units on the Basis of Safety and Security Premises | Masterarbeit |
| Automated Context Adaption of Iterative Software in Constraint Execution Environments | Masterarbeit |
| Automated AI-based Placement of Early Exits in Deep Neural Networks | Masterarbeit |