Site-Specific Beam Alignment without Explicit Channel Knowledge via Deep Learning

Bib

@ARTICLE{kwak2025site,
  author={Kwak, Jong Woo and Yoo, Hanju and Choi, Jaeyoung and Ko, Soomin and Roberts, Ian P. and Chae, Chan-Byoung},
  journal={Submitted},
  title={Site-Specific Beam Alignment without Explicit Channel Knowledge via Deep Learning},
  year={2025},
  note={equal contribution}
}

Jong Woo Kwak^*†, Hanju Yoo^*†, Jaeyoung Choi^‡, Soomin Ko^‡, Ian P. Roberts^§, and Chan-Byoung Chae^†

^* Equal contribution
^† School of Integrated Technology, Yonsei University, Seoul 03722, Korea
^‡ Samsung Electronics, Gyeonggi 16677, Korea
^§ Department of Electrical and Computer Engineering, University of California, Los Angeles, CA 90095, USA

The framework replaces explicit channel labels with practically obtainable beam-sweeping measurements.

Abstract

In this paper, we address the challenge of efficient beam alignment in wireless communications. Although recent studies demonstrate that deep learning-based methods can significantly accelerate site-specific beam alignment by leveraging propagation channel data between the base station (BS) and potential user locations, such datasets are impractical to acquire via exhaustive channel estimation and excessively time-consuming when using ray-tracing simulations. To overcome these limitations, we propose an innovative beam alignment framework that utilizes only noisy beam sweeping measurements, avoiding explicit channel estimation altogether. Our method involves training a model directly with these practically obtainable, noisy measurements, thus bypassing costly and impractical data acquisition processes. Specifically, the proposed approach constructs a dataset from beam sweeping measurements, enabling effective site-specific optimization without explicit knowledge of channel information. Simulation results confirm that our framework achieves beam alignment accuracy and speed comparable to state-of-the-art methods trained on perfect explicit channel datasets, demonstrating remarkable robustness even under significantly noisy measurements.

Why This Matters

Beam alignment gains often come from exploiting local propagation structure, but the price is environment-specific data collection. In many papers, that dataset is created with tools or labels that are far cleaner than what a real system can actually observe. This work closes that gap by building the learning pipeline around measurements that a beam-management system can realistically obtain.

That makes the contribution mainly about supervision design. The model is important, but the more important point is that practical supervision can be enough if the training target and dataset construction are chosen correctly.

What This Paper Does

Instead of building a supervised pipeline around ideal channel state information, the paper constructs the dataset from noisy sweeping responses. The method therefore reduces the cost of site adaptation while preserving the advantages of site-aware learning. That is a more realistic path to deployment than relying on exhaustive channel acquisition or ray-tracing labels.

The result is a beam alignment framework that remains grounded in the actual measurement pipeline rather than in an idealized offline labeling workflow.

Learned Probing Beam Patterns

The learned probing codebook is not just a smaller subset of conventional DFT beams. It adapts its angular coverage to the site. In the mostly LOS Rosslyn environment, the learned probing beams are distributed more broadly, while in the DeepMIMO O1_28B environment they concentrate energy toward useful reflected paths and UE regions.

Learned probing beam pattern in the Rosslyn environment — Rosslyn: the proposed probing beams cover the main user directions with broad, site-aware angular support.

Learned probing beam pattern in the DeepMIMO O1_28B environment — DeepMIMO O1_28B: the learned beams form multi-lobe patterns that exploit the NLOS propagation geometry.

Achieved SNR With Low Overhead

The practical metric is achieved downlink SNR after beam selection, not only whether the exact noiseless best-beam index is predicted. With top-$k$ refinement, the proposed method can recover most of the exhaustive-search SNR while probing far fewer beams.

Average downlink SNR versus probing codebook size in Rosslyn — Rosslyn: increasing the probing codebook size steadily closes the SNR gap to exhaustive search.

Average downlink SNR versus probing codebook size in DeepMIMO O1_28B — DeepMIMO O1_28B: the proposed method reaches exhaustive-search-level SNR with a small learned probing codebook.

Robustness to Measurement Noise

The method is trained and deployed with noisy power measurements. The SNR curves show that learned probing remains competitive across measurement-noise levels, especially in the NLOS DeepMIMO setting where naive exhaustive power selection can be corrupted by measurement noise.

Average downlink SNR versus measurement noise power in Rosslyn — Rosslyn: learned probing stays close to the CSI-aided trend over the tested noise range.

Average downlink SNR versus measurement noise power in DeepMIMO O1_28B — DeepMIMO O1_28B: learned probing remains robust as measurement noise increases.

Key Results

The learned probing beams adapt to the deployment site, forming broad coverage in LOS-heavy Rosslyn and multi-lobe NLOS-aware patterns in DeepMIMO O1_28B.
The proposed method approaches exhaustive-search SNR using far fewer probing beams, especially when a small top-$k$ refinement is allowed.
In DeepMIMO O1_28B, the method can match noisy exhaustive search with 8 probing beams for $k=3$ or 14 probing beams for $k=1$.
The remaining SNR gap to CSI-aided joint beam learning is modest, about 2 dB in Rosslyn and about 1 dB in DeepMIMO O1_28B.
The framework stays practical because training uses noisy narrow-beam power measurements instead of explicit CSI or ray-tracing labels.

Measurement-driven workflow showing how the proposed method turns beam-sweeping data into a practical site-specific learning pipeline.