Learning User Scheduling and Hybrid Precoding with Sequential Graph Neural Network

Learning-based methods have been developed for user scheduling and precoding in multi-antenna systems, among which most existing studies learned the two policies separately. In this paper, we strive to learn jointly optimized user scheduling and hybrid precoding policy with graph neural network (GNN), which has emerged as a powerful tool for optimizing resource allocation thanks to its potential in generalizability to problem scales. We find that the GNN for selecting users simultaneously does not perform well, due to a same-feature same-action phenomenon. To alleviate its adverse impact, we propose a sequential GNN (SGNN) architecture, which is a cascade of preprocessor, scheduler consisting of multiple sub-schedulers, and precoder modules. To assist SGNN in learning favorable scheduling policy, we add two model-based inputs into the preprocessor. To help reduce multiuser interference and allow generalizability to problem scales, we integrate attention mechanism into the GNN for precoding. Simulation results show that the joint scheduling and precoding policy learned by the proposed SGNN achieves higher sum-rate than separately optimized scheduling and precoding by numerical algorithms with much shorter running time, and is generalizable to the numbers of users and antennas.