Deep Learning Based Adaptive Joint mmWave Beam Alignment

TL;DR

This work tackles two-sided mmWave beam alignment under analog/hybrid constraints by proposing an end-to-end trainable joint BA. The architecture combines UE-side adaptive, codebook-free sensing with a BS-side learnable beamcodebook and a final grid-free beam mapper, optimizing $\frac{|\mathbf{w}_{T}^H \mathbf{H} \mathbf{f}_{T}|^2}{\|\mathbf{H}\|_2}$ through backpropagation. Compared with exhaustive search and prior DL-based schemes, the method yields substantial beamforming gains and higher satisfaction probabilities, while maintaining compatibility with existing cellular signaling. The framework supports parallel UE BA via BS codebook sweeps and offers a practical path toward faster, higher-gain initial access in mmWave systems; future work includes adaptive termination with DRL and extension to diverse BS codebooks.

Abstract

The challenging propagation environment, combined with the hardware limitations of mmWave systems, gives rise to the need for accurate initial access beam alignment strategies with low latency and high achievable beamforming gain. Much of the recent work in this area either focuses on one-sided beam alignment, or, joint beam alignment methods where both sides of the link perform a sequence of fixed channel probing steps. Codebook-based non-adaptive beam alignment schemes have the potential to allow multiple user equipment (UE) to perform initial access beam alignment in parallel whereas adaptive schemes are favourable in achievable beamforming gain. This work introduces a novel deep learning based joint beam alignment scheme that aims to combine the benefits of adaptive, codebook-free beam alignment at the UE side with the advantages of a codebook-sweep based scheme at the base station. The proposed end-to-end trainable scheme is compatible with current cellular standard signaling and can be readily integrated into the standard without requiring significant changes to it. Extensive simulations demonstrate superior performance of the proposed approach over purely codebook-based ones.

Figures (7)

• Figure 1: System model of the proposed BA algorithm. Note how at each timestep $t$, the beampatterns of BS and UE (i.e., $\mathbf{f}_t$ and $\mathbf{w}_t$) are determined by a respective CU, i.e., joint BA is performed.
• Figure 2: Unrolled depiction of the proposed joint BA algorithm. For $0 \leq t \leq T-1$ the BS selects its beam based on its learnable codebook, but for $t = T$, the beam is determined codebook free, based on the received feedback $\mathbf{m}_{\text{FB}}$ from the UE. Note that green shaded boxes represent trainable components. The red arrow indicates the flow of the gradient used for updating the trainable parameters in the scheme.
• Figure 3: Unrolled visualization of the learned beampatterns for the proposed joint BA algorithm for a specific channel realization.
• Figure 4: Comparison of the performance impact of the various NN used in the proposed scheme, top: beamforming gain versus SNR, bottom: satisfaction probability versus SNR.
• Figure 5: Comparison of the performance impact of varying $T$ for fixed $N_{\text{CB}} = 8$ for the proposed scheme. Solid lines: beamforming gain versus $T$, dashed lines: satisfaction probability versus $T$.