Graph-Augmented MAPPO for Dynamic Task Offloading in Edge Networks

In multi-access edge computing (MEC) environments, optimizing task offloading and communication-aware resource allocation is challenging due to dynamic topologies, fluctuating channels, and heterogeneous link capacities. This letter proposes a Graph-Augmented Multi-Agent Proximal Policy Optimization (GA-MAPPO) framework under a centralized training and decentralized execution (CTDE) paradigm. A Master Agent cooperative architecture integrates a Graph Attention Network (GAT) to model dynamic communication dependencies and bandwidth variations among edge nodes. The hybrid action space enables joint optimization of task offloading and transmission resource allocation. Experiments based on mobility and workload traces show that GA-MAPPO improves task completion rate and reduces latency compared with representative baselines, demonstrating its adaptability in dynamic MEC and Internet of Vehicles (IoV) scenarios.