Task Offloading and Resource Allocation for Space-Air-Ground Integrated Network

Abstract

Space–air–ground integrated networks (SAGIN) are a key enabler of 6G, as they can federate space, aerial, and terrestrial segments to deliver seamless global connectivity and distributed computing. However, the strong heterogeneity across domains, the mobility-induced time-varying topology, and the coupled latency–energy requirements make task offloading and resource allocation (TORA) substantially more difficult than in terrestrial networks. Existing DRL- or GNN-based TORA schemes typically rely on static and/or homogeneous graph representations, so they cannot accurately track dynamic inter-domain relations and thus yield suboptimal decisions in realistic SAGIN environments. To tackle this issue, we propose a dynamic heterogeneous graph neural network (DHGNN) driven reinforcement-learning framework that (i) models SAGIN as a time-varying heterogeneous graph to capture user–HAPS–LEO interactions, and (ii) performs multi-objective optimization to jointly reduce end-to-end latency and inter-node power consumption. The dynamic adjacency is periodically updated according to node mobility, channel quality, and resource availability, and the learned graph embeddings are fed to the policy network for fine-grained TORA decisions. Simulation results show that the proposed DHGNN achieves 282.08% higher cumulative reward than the static homogeneous-graph (SHOM) baseline, reduces the inference latency by more than 35% compared with the static heterogeneous (SHEM) scheme and by over 60% compared with SHOM, and attains 10–15% energy savings over static counterparts. Moreover, when integrated with the proposed dynamic function-switching mechanism (DFSM), the framework further cuts inference delay by 26.7%–30.8% in dynamic networks and reduces power consumption by up to 42.1%, demonstrating its effectiveness and scalability for intelligent resource management in realistic 6G SAGIN deployments.Space–air–ground integrated networks (SAGIN) are a key enabler of 6G, as they can federate space, aerial, and terrestrial segments to deliver seamless global connectivity and distributed computing. However, the strong heterogeneity across domains, the mobility-induced time-varying topology, and the coupled latency–energy requirements make task offloading and resource allocation (TORA) substantially more difficult than in terrestrial networks. Existing DRL- or GNN-based TORA schemes typically rely on static and/or homogeneous graph representations, so they cannot accurately track dynamic inter-domain relations and thus yield suboptimal decisions in realistic SAGIN environments. To tackle this issue, we propose a dynamic heterogeneous graph neural network (DHGNN) driven reinforcement-learning framework that (i) models SAGIN as a time-varying heterogeneous graph to capture user–HAPS–LEO interactions, and (ii) performs multi-objective optimization to jointly reduce end-to-end latency and inter-node power consumption. The dynamic adjacency is periodically updated according to node mobility, channel quality, and resource availability, and the learned graph embeddings are fed to the policy network for fine-grained TORA decisions. Simulation results show that the proposed DHGNN achieves 282.08% higher cumulative reward than the static homogeneous-graph (SHOM) baseline, reduces the inference latency by more than 35% compared with the static heterogeneous (SHEM) scheme and by over 60% compared with SHOM, and attains 10–15% energy savings over static counterparts. Moreover, when integrated with the proposed dynamic function-switching mechanism (DFSM), the framework further cuts inference delay by 26.7%–30.8% in dynamic networks and reduces power consumption by up to 42.1%, demonstrating its effectiveness and scalability for intelligent resource management in realistic 6G SAGIN deployments.