End-to-end Beamforming Design Based on Pilot in Multiuser Multi-Input-Single-Output System

This paper proposes a deep learning-based end-to-end system for the beamforming problem of time-division duplexing (TDD) multiuser multiple input single output (MUMISO) system under power constraints. In this paper, an end-to-end system network based on residual block embedded graph neural network (RB-GNN) is designed, which can directly parameterize the mapping from the pilots received at the base station (BS) to the beamforming matrix. RB-GNN can effectively utilize graph data structures with high-dimensional characteristics in wireless networks, and it has outstanding generalization when there is a change in user count by sharing parameters among them. The results indicate that the proposed network can obtain excellent sum-rate performance while short pilots and has good generalization ability.