Learned Preconditioned Conjugate Gradient Descent for Massive MIMO Detection

Abstract

In this paper, we investigate the use of model-based neural networks for Massive Multiple-Input Multiple-Output (MMIMO) detection. Recently, a new M-MIMO detection architecture called LcgNet [1] was obtained by unfolding an iterative conjugate gradient descent algorithm into a layer-wise network and introducing additional trainable parameters. Herein, we extend this approach by introducing a preconditioner aimed at improving the spectrum of the filter matrix used in the uplink MIMO detector. Specifically, the preconditioning scheme reduces the eigenvalue spread of the filter matrix, thus resulting in better convergence of the conjugate gradient algorithm. The proposed extension of LcgNet with preconditioning, referred to as PrLcgNet, is evaluated by means of simulations over M-MIMO uncorrelated Rayleigh fading channels and correlated fading channels. Compared to the original LcgNet, Pr-LcgNet exhibits faster convergence and lower residual error in the training phase, while achieving comparable bit error rate (BER) performance using fewer layers.