Graph Neural Network based Beamforming in D2D Wireless Networks

An unsupervised graph neural network (GNN) approach is proposed to solve the beamforming design problem in device-to-device (D2D) wireless networks. Instead of directly learning the beamforming, the GNN is utilized to learn primal power and dual variables, and then a beamforming recovery module is applied to convert them to the beamforming. In this way, the overall problem dimension is decreased by a factor of the number of antennas. Additionally, the proposed GNN approach is potential to be generalized to different system settings without retraining when the number of antennas remains unchanged. Simulation results demonstrate that the proposed GNN based beamforming approach achieves superior performance with 10 times fewer samples than the benchmarks, and the running time is reduced down to millisecond-level for 50 pairs of D2D users which is promising for practical applications in D2D wireless networks.