Online Multi - Access Scheduling in Space-Air-Ground Integrated Networks Using Graph Neural Network - Enhanced Reinforcement Learning

Space-air-ground integrated networks (SAG IN) are currently a focal point of research, representing an emerging architectural concept for sixth-generation (6G) networks. Given the limited resources and varying connection patterns of the platform, the cross-platform access design is challenging. In this paper, we explore the scenario where multiple users enter the network simultaneously and request access to the target high altitude platforms (HAPs). The system must make real-time decisions regarding concurrent requests, all while considering the connection of HAPs to satellites. By explicitly modeling the data collection process and the workload of transmission tasks, the optimal access scheduling algorithm is investigated by maximizing the total network data size while maintaining a balance in workload allocation across HAPs. This problem is first formulated as a mixed-integer programming (MIP) problem, which has high computational complexity and requires prior knowledge. Therefore, a deep reinforcement learning (DRL)-based model with a graph neural network (GNN) and a sequence-to-sequence (Seq2Seq) module for practical online problems is proposed. Extensive simulation results validate the effectiveness of the proposed method for handling multi-user requests with significantly reduced computational complexity.