Priority-Based Load Balancing With Multi-Agent Deep Reinforcement Learning for Space-Air-Ground Integrated Network Slicing

Space-air-ground integrated network (SAGIN) slicing has been studied for supporting diverse applications, which consists of the terrestrial layer (TL) deployed with base stations (BS), the aerial layer (AL) deployed with unmanned aerial vehicles (UAV), as well as the space layer (SL) deployed with low earth orbit (LEO) satellites. The capacity of each SAGIN component is limited, and efficient and synergic load balancing has not been fully considered yet in the exiting literature. For this motivation, we originally propose a priority-based load balancing scheme for SAGIN slicing, where the AL and SL are merged into one layer, namely non-TL (NTL). Firstly, three typical slices (i.e., high-throughput, low-delay, and wide-coverage slices) are built under the same physical SAGIN. Then, a priority-based cross-layer load balancing approach is introduced, where the users will have the priority to access the terrestrial BS, and different slices have different offloading priorities. More specifically, the overloaded BS can offload the users of low-priority slices to the NTL preferentially. Furthermore, the throughput, delay, and coverage of the corresponding slices are jointly optimized by formulating a multi-objective optimization problem (MOOP). In addition, due to the independence and priority relationship of TL and NTL, the above MOOP is decoupled into two sub-MOOPs. Finally, we customize a two-layer multi-agent deep deterministic policy gradient (MADDPG) algorithm for solving the two sub-problems, which firstly optimizes the user-BS association and resource allocation at the TL, then it determines the UAVs’ position deployment, users-UAV/LEO satellite association, and resource allocation at the NTL. The reported simulation results show the advantages of our proposed LB scheme and show that our proposed algorithm outperforms the benchmarkers.