Beam Alignment for Millimeter Wave Wireless Communications : A Multiscale Approach

Millimeter-wave communications use narrow beams to overcome the enormous signal attenuation. Such narrow-beam communication demands precise beam-alignment between transmitter and receiver and may entail huge overhead, especially in high mobility scenarios. Moreover, detection of the optimal beam is challenging in the presence of beam imperfections and system noise. This thesis addresses the challenges in the design of beam-training and data-communication by proposing various schemes that exploit different timescales. On a short timescale, we leverage the feedback from the receiver to efficiently perform beam-training and data-communication. To this end, we have worked in three different areas. In the first research direction, we design an optimal interactive beam-training and data-communication protocol, with the goal of minimizing power consumption under a minimum rate constraint. The optimality of a fixed-length beam-training phase followed by a data-communication phase is proved under the assumption of perfect binary feedback. In the second research direction, we propose a coded energy-efficient beam-training scheme, robust against the feedback/detection errors. In the third research direction, we investigate the design of the beam-training in the presence of uncertainty due to noise and beam imperfections. Based on the bounding of value-function, the second-best preference policy is proposed, which achieves a promising exploration-exploitation tradeoff. On the other hand, on longer timescales, we exploit the mobility and blockage dynamics and beam-training feedback to design throughput-efficient beam-training and data-communication. We propose a point-based value iteration (PBVI) algorithm to determine an approximately optimal policy. However, the design relies on the a-priori knowledge of the state dynamics, which may not be available in practice. To address this, we propose a dual timescale approach, where on the long timescale, a recurrent deep variational autoencoder (R-VAE) uses noisy beam-training observations to learna probabilistic model of system dynamics; on the short timescale, an adaptive beam-training procedure is optimized using PBVI based on beam-training feedback and a probabilistic knowledge of the UE's position provided by the R-VAE. In turn, the observations collected during the beam-training procedure are used to refine the R-VAE via stochastic gradient descent in a continuous process of learning and adaptation.