ADMM-SLPNet: A Model-Driven Deep Learning Framework for Symbol-Level Precoding

Constructive interference (CI)-based symbol-level precoding (SLP) is an emerging downlink transmission technique for multi-antenna communications systems, and its low-complexity implementations are of practical importance. In this paper, we propose an interpretable model-driven deep learning framework to accelerate the processing of SLP. Specifically, the network topology is carefully designed by unrolling a parallelizable algorithm based on the proximal Jacobian alternating direction method of multipliers (PJ-ADMM), attaining parallel and distributed architecture. Moreover, the parameters of the iterative PJ-ADMM algorithm are untied to parameterize the network. By incorporating the problem-domain knowledge into the loss function, an unsupervised learning strategy is further proposed to discriminatively train the learnable parameters using unlabeled training data. Simulation results demonstrate significant efficiency improvement of the proposed ADMM-SLPNet over benchmark schemes.