Learning End-to-End Hybrid Precoding for Multi-User mmWave Mobile System With GNNs

Learning End-to-End Hybrid Precoding for Multi-User mmWave Mobile System With GNNs

Hybrid precoding is an efficient technique for achieving high rates at a low cost in millimeter wave (mmWave) multi-antenna systems. Many research efforts have explored the use of deep learning to optimize hybrid precoding, particularly in static channel scenarios. However, in mobile communication s...

Full description

Saved in:
Bibliographic Details
Main Authors: Ruiming Wang, Chenyang Yang, Shengqian Han, Jiajun Wu, Shuangfeng Han, Xiaoyun Wang
Format: Article
Language:English
Published: IEEE 2024-01-01
Series:IEEE Transactions on Machine Learning in Communications and Networking
Subjects:
Hybrid precoding
end-to-end learning
channel prediction
channel estimation
graph neural network
Online Access:https://ieeexplore.ieee.org/document/10577095/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Hybrid precoding is an efficient technique for achieving high rates at a low cost in millimeter wave (mmWave) multi-antenna systems. Many research efforts have explored the use of deep learning to optimize hybrid precoding, particularly in static channel scenarios. However, in mobile communication systems, the performance of mmWave communication severely degrades due to the channel aging effect. Furthermore, the learned precoding policy should be adaptable to dynamic environments, such as variations in the number of active users, to avoid the need for re-training. In this paper, resorting to the proactive optimization approach, we propose an end-to-end learning method to learn the downlink multi-user analog and digital hybrid precoders directly from the received uplink sounding reference signals, without explicit channel estimation and prediction. We take into account the frame structure used in practical cellular systems and design a parallel proactive optimization network (P-PONet) to concurrently learn hybrid precoding for multiple downlink subframes. The P-PONet consists of several graph neural networks, which enable the generalizability across different system scales. Simulation results show that the proposed P-PONet outperforms existing methods in terms of sum-rate performance and sounding overhead, and is generalizable to various system configurations.
ISSN:2831-316X

Similar Items