Cascaded Binary Classifiers for Blind Beam Alignment in mmWave MIMO Using One-Bit Quantization

This paper proposes a new approach for partial and blind Machine Learning (ML)-based Beam Alignment (BA) for massive mmWave MIMO. It models an uplink scenario using one-bit quantization through a low-complexity fully-analog system architecture. The proposed BA is based on sub-sampled codebooks holding possible beam patterns at UE and BS. We propose to sound a small subset of beams based on instantaneous Received Signal Energies (RSE). These sounded RSE values are then quantized into binary integers. The proposed cascaded structure of Binary Logistic Regression (BLR) aims to iteratively filter the large dataset input-matrix (by deleting low-RSE beams) into a smaller one where our benchmark, the Exhaustive BA is feasible and the signaling overhead remains low. In addition to the theoretical monotonic-convergence guarantees, BLR has good classification quality and low computational complexity. Our extensive numerical simulations illustrate encountering the large signaling overhead problem with high prediction accuracy using one-bit quantization scheme and 14% of the total beam samples.