Balancing Uncertainty and Determinism: Safely Integrating Machine Learning in Real-Time Systems

• Applicant: Dr.-Ing. Junjie Shi
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 572586998

Machine learning (ML) algorithms are increasingly integrated into time-sensitive systems such as autonomous vehicles, industrial controllers, and healthcare devices. These applications require not only intelligent decision-making but also strict guarantees on timing, reliability, and safety. However, most ML models are computationally intensive and probabilistic by nature, making them difficult to reconcile with the deterministic requirements of real-time and embedded systems. In addition, deploying ML across distributed embedded platforms introduces new challenges related to strict data privacy requirements and dynamic heterogeneity, including variations in hardware capabilities, resource availability, and continuously evolving data distributions at the local level. This project aims to develop principled methods for the safe and efficient integration of machine learning into embedded real-time systems. It addresses key challenges at the intersection of timing predictability, adaptive model deployment, and intelligent system behavior. Specifically, the project pursues the following three core objectives: (1) Ensuring the timing predictability of ML tasks, by developing analytical models and architecture-aware optimization techniques to derive tight worst-case execution time (WCET) and worst-case response time (WCRT) bounds. This includes model restructuring, task partitioning, and memory-aware deployment on multicore platforms. (2) Enabling adaptive, resource-efficient, and privacy-preserving deployment of ML models to address the challenges of dynamically evolving distributed embedded systems. This involves developing federated learning frameworks that accommodate heterogeneous hardware, constrained communication, and dynamically evolving, non-independent and identically distributed (non-IID) data distributions. (3) Leveraging ML techniques to enhance the performance and analysis of real-time systems, including accelerating schedulability testing, generating counterexamples for infeasibility detection, and applying learning-based approaches to support adaptive scheduling and system optimization—all while maintaining formal correctness and timing guarantees. Together, these contributions will support the development of intelligent, dependable, and timing-safe cyber-physical systems capable of learning and adapting under stringent operational constraints.

• DFG Programme: WBP Fellowship
• International Connection: Singapore
• Host: Professorin Tulika Mitra, Ph.D.