MIMO Channel Estimation with Non-Ideal ADCS: Deep Learning Versus GAMP

Abstract

Channel estimation for massive MIMO using coarse quantizers is nontrivial due to severe nonlinear distortions caused by quantization and the large dimensional MIMO channel. The best solutions to this problem nowadays are based on the generalized approximate message passing (GAMP) and its variations. However, there are practical issues such as nonideal quantizers that may violate the assumptions in which GAMP algorithms rely. This motivates research on methods based on deep learning (DL), which provides a framework for designing new communication systems that embrace practical impairments. We explore DL applied to channel estimation of MIMO systems with low resolution analog-to-digital converters (ADCs). Assuming millimeter wave MIMO, the channel estimation results indicate that a single neural network trained in a range of practical conditions is more robust to ADC impairments than a GAMP variant. For example, for a fixed wireless scenario with channels obtained from raytracing, DL achieved a normalized mean-square error lower than GAMP’s by more than 5 dB.