CPN-HRL :a hierarchical deep reinforcement learning approach for priority–aware task scheduling in CPN enabled by cloud–edge–end environments

A Computing Power Network (CPN) is a cloud–edge–end integrated infrastructure that interconnects heterogeneous computing resources to support latency-sensitive and volatile workloads. A core challenge in CPN task scheduling is to jointly determine task prioritization and computational node selection under dynamic arrivals, resource heterogeneity, and multi-objective conflicts. In this paper, we study dynamic task scheduling in a CPN and aim to (i) minimize average task completion time, (ii) maximize scheduling success rate under deadline constraints, and (iii) improve global resource utilization. We propose CPN-HRL, a hierarchical deep reinforcement learning framework that decomposes scheduling into two coordinated subproblems: a high-level agent (LSTM-PPO) generates global priority weights for queued tasks, and a low-level agent (GAT-PPO) performs topology-aware node selection for each task in the ranked sequence. To further guarantee timeliness for latency-critical tasks, we introduce a two-queue mechanism with an Urgent Queue and a Main Queue, together with a dynamic reordering trigger. Simulation results on multiple real-world topologies show that, under our experimental settings, CPN-HRL reduces average completion time and maintains high scheduling success rate under high loads, while improving overall resource utilization.