Entropy-based Probing Beam Selection and Beam Prediction via Deep Learning

TL;DR

This work tackles the overhead of mmWave beam management by formulating a probabilistic beamspace model and solving joint probing and beam prediction via entropy minimization. It introduces a greedy iterative framework (Iter-BP&PBS) with a diagonal covariance approximation and a two-stage variant (2S-BP&PBS) that uses a location-aware codebook to reduce interactions and compute load. A transformer-based mean/variance beam predictor is developed, featuring specialized embedding and attention-based processing, enabling location-informed predictions of the beampower distribution. Simulations in urban mmWave scenarios show substantial improvements over hierarchical search and uniform probing, with 2S-BP&PBS offering a favorable balance between performance, complexity, and online latency. The approach provides a scalable, data-driven strategy for efficient BA/T under tight training and interaction constraints.

Abstract

Hierarchical beam search in mmWave communications incurs substantial training overhead, necessitating deep learning-enabled beam predictions to effectively leverage channel priors and mitigate this overhead. In this study, we introduce a comprehensive probabilistic model of power distribution in beamspace, and formulate the joint optimization problem of probing beam selection and probabilistic beam prediction as an entropy minimization problem. Then, we propose a greedy scheme to iteratively and alternately solve this problem, where a transformer-based beam predictor is trained to estimate the conditional power distribution based on the probing beams and user location within each iteration, and the trained predictor selects an unmeasured beam that minimizes the entropy of remaining beams. To further reduce the number of interactions and the computational complexity of the iterative scheme, we propose a two-stage probing beam selection scheme. Firstly, probing beams are selected from a location-specific codebook designed by an entropy-based criterion, and predictions are made with corresponding feedback. Secondly, the optimal beam is identified using additional probing beams with the highest predicted power values. Simulation results demonstrate the superiority of the proposed schemes compared to hierarchical beam search and beam prediction with uniform probing beams.

Figures (13)

• Figure 1: Plots of information interactions between the BS and the MU.
• Figure 2: Plots of selected probing beams and prediction results, where the x and y axes respectively are the numbers of antennas in the horizontal and vertical dimensions. The optimal beams are marked with red circles. The first sub-figure is the ground-truth of RSRP. The second and third sub-figures respectively are the mean and the corresponding variance estimates of the RSRP. The fourth sub-figure is the selected probing beams.
• Figure 3: The 2D location-specific probing codebook $\mathcal{C}$ for probing is composed of $N_{\textup{x}} \times N_{\textup{y}}$ grids. Each grid stores a codeword including a probing beam set and a counterpart relative MU location. Given the location, the probing beams are selected by binary searching.
• Figure 4: An illustration of the mean and variance networks (left) and a transformer (right).
• Figure 5: Layout of a mmWave communication scenario. The BSs marked in circles are located on the buildings or along the streets, and the MUs are distributed outdoors.