Optimized scheduling in Cyber–Physical Systems via constrained reinforcement learning

Online scheduling is a core tool in modern Cyber–Physical Systems (CPS), enabling real-time task assignment under dynamic resource availability and strict operational constraints. While reinforcement learning (RL) offers a principled approach to learning adaptive scheduling policies from interaction, standard RL methods often neglect hard safety and latency constraints that are critical in CPS applications. Reliable task scheduling under resource constraints is a core challenge in CPS, where real-time performance and safety must be ensured despite dynamic and uncertain environments. To bridge this gap, this paper proposes CoRLS, a constrained reinforcement learning framework for real-time scheduling in CPS. CoRLS formulates the task as a constrained Markov decision process and optimizes it using primal–dual updates and a Lagrangian relaxation scheme. To ensure safety and adaptability, it integrates a resource abstraction encoder and a safety filter for online feasibility. Evaluations on both synthetic and real-world CPS datasets demonstrate that CoRLS outperforms four competitive baselines, achieving average improvements of 24% in utility, 78% in violation rate, 60% in deadline adherence, and 11% in resource utilization. Sensitivity analysis further confirms its converged robustness across hyperparameter and environmental variations.