Online Scheduling for Multi-Access in Space-Air-Ground Integrated Networks Using iGNNSeq-Enhanced Reinforcement Learning

Abstract

Space-air-ground integrated networks (SAGINs) have become a key area of research, representing a cutting-edge architectural concept for sixth-generation (6G) networks aimed at achieving ubiquitous connectivity. This paper focuses on the challenge of online access scheduling across multiple platforms. Specifically, when multiple users randomly enter the network and request access through high-altitude platform (HAP) relays, the system must make real-time decisions on whether to grant or deny access, while considering the connections between HAPs and low-Earth orbit (LEO) satellites. By explicitly modelingthe data collection process and the workload associated with transmission tasks, we formulate the multi-access scheduling challenge as a mixed-integer programming (MIP) problem, which presents high computational complexity. Given the practical system's lack of prior knowledge about user attributes, such as arrival times and channel conditions, we propose a deep reinforcement learning (DRL)-based model for real-time decision-making. To enhance feature extraction within the DRL architecture, we introduce a graph neural network (GNN)-based encoder, and employ a sequence-to-sequence (Seq2Seq) module to manage the exponential increase in action space resulting from simultaneous decisions on multiple access requests. Extensive simulations demonstrate that the integrated GNN and Seq2Seq (iGNNSeq)-enhanced RL achieves performance comparable to offline solutions while significantly reducing computational complexity.