A Task Prediction-Augmented Hierarchical Offloading Method for Space-Air-Ground Integrated Networks

Abstract

Space-air-ground integrated networks (SAGIN) combine low Earth orbit satellites, unmanned aerial vehicles and ground devices into an efficient converged architecture for computation-intensive mobile applications. However, the strong coupling among UAV trajectory control, task offloading and resource allocation, together with dynamic and uncertain task workloads, makes it challenging to offload tasks efficiently while balancing delay and energy. This paper minimizes a weighted cost of task completion delay and UAV flight energy, models the problem as a decentralized partially observable Markov decision process, and proposes a prediction-augmented multi-agent proximal policy optimization algorithm (PA-MAPPO). A lightweight task workload prediction module is introduced into the multi-agent reinforcement learning framework to strengthen forward-looking decisions among agents, jointly optimizing UAV trajectories, task offloading and computing resource allocation in dynamic SAGIN. Simulation results show that the proposed algorithm effectively lowers the overall system cost and balances average task delay against UAV flight energy.

Translated from the Chinese abstract.