Joint Optimization of User Scheduling and Beamforming Design Based on Deep Learning

This letter considers the joint optimization of user scheduling and beamforming design with the target of maximizing the sum-rate of a multi-user MISO system. The resulting optimization problem includes the non-convex objective function, the discrete constraint, and the mutually coupled constraints, thus is challenging to tackle with. In order to overcome the shortcomings of traditional optimization schemes, a deep learning (DL) framework based on graph convolutional neural network and cross-attention mechanism (GCN-CA) is proposed. This framework first leverages GCN to capture the network topology information and extract features from the channel state information (CSI) to predict the transmit power. Then, the CA mechanism is adopted to further extract the features between the CSI and the transmit power vector outputted by GCN, so as to predict the sparse transmit power vector and obtain the user scheduling strategy. Finally, by combining the analytical formula of the beamforming vector and the sparse transmit power vector of the final output of the GCN-CA framework, the joint optimization of the beamforming design and user scheduling is realized. Simulation results demonstrate that, compared to traditional optimization algorithms and pure GCN solution, the proposed GCN-CA framework exhibits superior convergence performance and lower training complexity, and also has good generalizability in dynamic networks.