Reducing Overall Path Latency in Automotive Logical Execution Time Scheduling via Reinforcement Learning
Risheng Xu, Marvin Kühl, Hermann von Hasseln, Dirk Nowotka · 2023
The Logical Execution Time paradigm is a promising approach for achieving time-deterministic communication on multi-core CPUs. Task scheduling under this paradigm is a variant of the Multi-Row Facility Layout Problem, which is known to be NP-hard. In this paper, we propose using reinforcement learning to reduce the overall path latency among all scheduled runnables while adhering to other constraints, such as schedulability, load balance, and data contention control. The neural networks, also known as agents, are trained on a real-world automotive powertrain project. We compare two schedules generated by the agents to the current one and one produced by a genetic algorithm. The agent trained with the Proximal Policy Optimization algorithm demonstrated the best performance. Additionally, we investigate the generalization ability of the agents against software updates, and the results show that our agents are well-generalized.