Graph Neural Network Architecture for MIMO Channel Estimation

Graph Neural Networks (GNNs) are gaining popularity to solve wireless communication problems due to the inherent nature of viewing wireless networks as graphs. In this paper, we investigate the potential of using GNN for channel estimation in point-to-point multi-input and multi-output (MIMO) systems. The analysis includes an assessment of the channel estimation error performance, alongside the associated computational complexity. The performance is compared against benchmarks such as the conventional least squares, (linear) MMSE estimators, and shallow 1D convolutional neural network (1D-CNN) architectures from the literature. We evaluate the performance-complexity trade-off of GNN versus shallow 1DCNN which have been shown to outperform deep architectures for MIMO channel estimation.