Deep Learning-Based Precoder Design for Network Massive MIMO Transmission

We investigate the linear precoding for sum-rate maximization in network massive multiple-input multiple-output (MIMO) transmission, where the cooperative transmission by all base stations (BSs) enhances the capacity, reliability, and robustness. To address the growing complexity of traditional iterative algorithms in large-scale systems, we leverage the weighted minimum mean square error (WMMSE) solution and show that the precoding vectors can be fully reconstructed from a set of low-dimensional parameters. By exploiting the structure and relationship of these parameters, we reformulate the original problem in a reduced-dimensional space while preserving equivalence to the original solution. Deep learning techniques are employed to solve this reformulated problem, where equivalent scaling of the variables facilitates pre-processing for training and further reduces the dimension of the learning input. A neural network is trained on the refined low-dimensional objectives with a tailored loss, allowing the precoding vectors to be directly calculated from its output. As demonstrated by numerical results, the proposed deep learning-based precoder performs well with considerably reduced online processing complexity.