A Fast Deep Unfolding Learning Framework for Robust MU-MIMO Downlink Precoding

This paper reformulates a worst-case sum-rate maximization problem for optimizing robust multi-user multiple-input multiple-output (MU-MIMO) downlink precoding under realistic per-antenna power constraints. We map the fixed number of iterations in the developed mean-square-error uplink-downlink duality iterative algorithm into a layer-wise trainable network to solve it. In contrast to black-box approximation neural networks, this proposed unfolding network has higher explanatory power due to fusing domain knowledge from existing iterative optimization approaches into deep learning architecture. Moreover, it could provide faster robust beamforming decisions by using several trainable key parameters. We optimize the determination of the channel error's spectral norm constraint to improve the sum rate performance. The experimental results verify that the proposed deep unfolding "RMSED-Net" could combat channel errors in comparison with the non-robust baseline. It is also confirmed by the simulations that the proposed RMSED-Net in a fixed network depth could substantially reduce the computing time of the conventional iterative optimization method at the cost of a mild sum rate performance loss.