Deep Active Learning Approach to Adaptive Beamforming for mmWave Initial Alignment

This paper proposes a deep learning approach to the adaptive and sequential beamforming design problem for initial access phase in mmWave environment with single-path channel. For single-user scenario where is equivalent designing sequence of sensing beamformers learn angle arrival (AoA) dominant path, we propose novel neural network (DNN) that designs vectors sequentially based on available information so far at base station (BS). By recognizing AoA posterior distribution sufficient statistic solving problem, use as input proposed DNN strategy. However, computing can be computationally challenging when channel fading coefficient unknown. To address this issue, an estimate compute approximation distribution. Further, shows deal practical constraints such constant modulus constraint. Numerical results demonstrate compared existing non-adaptive schemes, DNN-based strategy achieves significantly better acquisition performance.