Data-Driven Based Preconditioning for One-Bit MIMO Detection

Multiple-input multiple-output (MIMO) technology is a keystone for a wide range of wireless communications applications. Therefore, great efforts have been made to design efficient and high-performance MIMO detection algorithms. Current state-of-the-art methods are model-based algorithms such as the alternating direction method of multipliers (ADMM). However, these methods are limited to the ill-conditioned nature of the channel matrix producing subpar convergence and recovery performance. To address this issue, preconditioning methods employ a matrix (denoted as preconditioning matrix) in the algorithm iterations to mitigate the instabilities caused by the ill-conditioned channel matrix. Here, we present a method to design the preconditioning matrix in a data-driven manner. The method is inspired by the knowledge distillation theory of neural networks. Specifically, we use a well-conditioned (virtual) channel matrix, that has a high MIMO detection performance, as guidance to optimize the preconditioning matrix. Intuitively, the preconditioning matrix is designed such that the estimated signal via the preconditioned algorithm with the ill-conditioned channel matrix is similar to the one obtained with the well-conditioned channel matrix. We validate the method using an ADMM algorithm, showing a significant detection improvement using the proposed preconditioning method. Additionally, we show that the proposed guidance method for designing the preconditioning matrix allows better detection than using a supervised loss function.