Task-Prediction-Augmented Multi-Agent Collaborative Offloading in Space-Air-Ground Integrated Network

摘要

The space-air-ground integrated network (SAGIN) can provide a promising architecture for computation-intensive mobile applications through the complementary advantages of unmanned aerial vehicles (UAVs) and low Earth orbit (LEO) satellites. However, efficient task offloading remains challenging due to the strong coupling among UAV trajectories, user association, and resource allocation, especially with dynamic and uncertain task workloads. In this paper, we aim to minimize the average task completion delay by enabling UAVs and satellites to cooperatively assist users in executing complex computational tasks. We formulate the joint optimization problem as a decentralized partially observable Markov decision process (DEC-POMDP), where task workload prediction is incorporated to enhance foresight and coordination among agents. Then, a task-prediction-augmented multi-agent collaborative offloading (TAMACO) framework is proposed, which integrates task-prediction-augmented multi-agent proximal policy optimization (TA-MAPPO) for UAV trajectory control. Meanwhile, a prediction-augmented coalition formation game (PA-CFG) algorithm embedded into TA-MAPPO is proposed to solve the joint problem of user association and computing resource allocation. Simulation results demonstrate that the proposed TAMACO framework achieves up to 19.3% reduction in average task completion delay compared to non-predictive baselines, validating its effectiveness for real-time task offloading in dynamic SAGIN.