PPO-STGNN Uses Reinforcement Learning for Cloud-Edge DAG Scheduling

Summary

PPO-STGNN is a reinforcement-learning algorithm for scheduling computation-intensive directed acyclic graph (DAG) tasks in collaborative cloud-edge-end environments. The setting is difficult because cloud, edge, and end nodes differ in computing capacity, network bandwidth, and energy consumption, while task dependencies make scheduling NP-hard. The paper argues that traditional heuristics and conventional reinforcement-learning methods do not adequately represent the spatio-temporal changes in available resources. PPO-STGNN combines proximal policy optimization with spatio-temporal graph neural networks. The STGNN extracts features from both the DAG task topology and the physical resource graph spanning cloud, edge, and end nodes. PPO then optimizes the scheduling policy to reduce makespan and schedule length ratio while improving CPU and memory load balance. The method also uses multi-teacher behavior cloning to pretrain the policy and accelerate convergence. Experiments reported in the abstract show substantially better load balancing while retaining low completion time, supporting the method's use for dynamic, heterogeneous DAG scheduling scenarios.