Learning-Based NOMA-Enabled Queue-Aware Task Offloading and AAV 3D Trajectory Planning for SAGIN

Space-air-ground integrated network (SAGIN) is a viable solution to serve users in such areas lacking ground base stations, in which case autonomous aerial vehicles (AAVs) provide massive access and satellites provide backhaul transmission. To improve spectrum utilization and throughput, non-orthogonal multiple access (NOMA) is used to reuse channels. However, different trajectories of AAVs and computing task assignment may lead to various communication delay and energy consumption. Additionally, regarding the high dynamic nature of SAGIN, it often falls into the dilemma of information uncertainty and curse of dimensionality. To overcome the above challenges, a hierarchical network model is designed, where AAVs and satellites collaborate to process users' offloaded computing tasks. Then, a joint AAV trajectory design, task offloading, task assignment and computing resource allocation problem for minimizing system cost is then proposed. Due to the coupling between queue delay constraints and decision-making, Lyapunov optimization is applied to split the issue into three subproblems. Firstly, a AAV trajectory design and task offloading algorithm based on multi-agent twin delayed deep deterministic policy gradient (MATD3) is designed to deal with the curse of dimension issue. Then, the problem of joint task assignment and AAV computing resource allocation is solved via CVX due to that it is convex. Finally, a low-complexity Greedy-based Satellite Computing Resource Allocation (GSCRA) algorithm is proposed to obtain the minimum system cost. Simulations show the superior performance of our algorithm compared with benchmarks in reducing the system cost.