Machine Learning-Assisted Codebook Design for MMSE Channel Estimation

In order to realize the high spectral efficiency promised by the new radio, the channel estimation algorithm shall be carefully designed to achieve high accuracy. Conventional optimal estimation techniques require the knowledge of second-order statistics, which is difficult to compute in real-time systems. In this work, we propose a machine learning-assisted codebook design for minimum mean square error (MMSE) channel estimation, where ML techniques are applied to cluster the channels that minimize the sum normalized mean square error (MSE). Specifically, we use k-medoids as the clustering method and we propose the normalized mismatched weight error performance as the clustering dissimilarity measure. Simulation results demonstrate the close-to-optimal channel estimation performance with reasonable codebook size, and robustness to signal-to-noise-ratio (SNR) changes and truncated codeword length.