Adaptive Beam Alignment using Noisy Twenty Questions Estimation with Trained Questioner

TL;DR

This work tackles the latency and interpretability bottlenecks of beam alignment (BA) in 6G mmWave MIMO systems by recasting BA as a noisy twenty questions estimation problem. It introduces two feasible, interpretable questioner-training strategies to map queries into beamforming vectors: (i) Linear Weighted Sum (LWS), which forms a normalized sum of steering vectors within the current query region, and (ii) a Deep Neural Network (DNN) mapper that learns a direct mapping from steering vectors to a beamformer while keeping the learning localized to the mapping step. Through extensive simulations, the proposed BA algorithms consistently outperform benchmark schemes (naive and hierarchical beam sweeping, ideal hiePM-based, and end-to-end DNN BA) under both 1-bit and full measurement rules, with additional robustness to unknown fading $\alpha$ when using Kalman filtering. The results demonstrate that feasible, interpretable, and low-latency BA is attainable for practical mmWave/MIMO deployments, offering gains in accuracy and latency that are critical for 6G connectivity.

Abstract

The 6G communication systems use mmWave and MIMO technologies to achieve wide bandwidth and high throughout, leading to indispensable need for beam alignment to overcome severe signal attenuation. Traditional sector-search-based beam alignment algorithms rely on sequential sampling to identify the best sector, resulting in a significant latency burden on 6G communication systems. Recently proposed adaptive beam alignment algorithms based on the active learning framework address the problem, aiming to identify the optimal sector with the fewest possible samples under an identical sector partition. Nevertheless, these algorithms either lack feasibility (Chiu, Ronquillo and Javidi, JSAC 2019) due to ideal assumptions or lack interpretability (Sohrabi, Chen and Yu, JSAC 2021) due to the use of end-to-end black-box neural networks. To avoid ideal assumptions and maintain interpretability, we address all above problems by proposing an adaptive beam alignment algorithm using the framework of noisy twenty questions estimation with a trained questioner. Specifically, we use two methods for training the questioner to eliminate reliance on ideal assumptions. The first method maps queries of twenty questions estimation to beamforming vectors via weighted summation of steering vectors, as an initial attempt to address the feasibility problem encountered in prior pioneering study by Chiu, Ronquillo and Javidi (JSAC 2019). The second method uses multi-layer fully connected neural networks to achieve improved performance while only employing them to train the questioner, which can effectively mitigate the interpretability issues in prior study by Sohrabi, Chen and Yu (JSAC 2021). Furthermore, we provide numerical simulations to illustrate the effectiveness of our proposed adaptive beam alignment algorithms and demonstrate that our algorithms outperform all benchmark algorithms.

Figures (17)

• Figure 1: Illustration of an adaptive beam alignment procedure in an air-ground communication scenario with a BS and a UAV.
• Figure 2: Illustration of sub-intervals of search interval $[\theta_{\mathrm{min}},\theta_{\mathrm{max}}]$.
• Figure 3: Diagram of adaptive BA algorithm.
• Figure 4: Diagram of noisy twenty questions estimation framework
• Figure 5: Diagram of conversion relationship between adaptive BA algorithm and noisy adaptive twenty questions estimation framework

Theorems & Definitions (4)

• Definition 1
• Definition 2
• Definition 3
• Definition 4