Learning Statistically Robust MIMO Detection with Imperfect CSI

For multi-input multi-output (MIMO) systems, the detection performance can be severely deteriorated by the channel state information (CSI) uncertainties. In this paper, we propose a learnable robust MIMO detector by taking the statistics of CSI imperfection into account. Specifically, we first formulate a robust maximum likelihood (ML) detection problem and then develop an alternating direction method of multipli-ers (ADMM) based solution, which involves the calculations of closed-form expressions in each iteration. Furthermore, a model-driven neural network is established by unfolding the derived ADMM algorithm whose penalty parameters are learned via offline training. Simulation results demonstrate that the proposed network can considerably outperform the conventional mismatched ML detector and even approach the optimal robust ML detector with only 5 layers.