Cross-Environment Transfer Learning for Location-Aided Beam Prediction in 5G and Beyond Millimeter-Wave Networks

TL;DR

This work tackles the overhead of maintaining mmWave beam alignment by predicting the best beam pair from UE location information. It introduces a cross-environment transfer-learning approach that trains a location-to-beam model on a reference gNB and then fine-tunes it on a target gNB with limited data, evaluated via ray-tracing in Frankfurt and Seoul. Results show that with as little as 5% of the target data, the model can achieve up to 80% top-1 accuracy, with substantial gains over naive transfer and broad generalization improvements across both intra-city and inter-city transfers. The findings imply that transfer learning can substantially reduce training data requirements and computational burden for practical location-aided beam prediction in 5G and beyond networks.

Abstract

Millimeter-wave (mm-wave) communications requirebeamforming and consequent precise beam alignmentbetween the gNodeB (gNB) and the user equipment (UE) toovercome high propagation losses. This beam alignment needs tobe constantly updated for different UE locations based on beamsweepingradio frequency measurements, leading to significantbeam management overhead. One potential solution involvesusing machine learning (ML) beam prediction algorithms thatleverage UE position information to select the serving beamwithout the overhead of beam sweeping. However, the highlysite-specific nature of mm-wave propagation means that MLmodels require training from scratch for each scenario, whichis inefficient in practice. In this paper, we propose a robustcross-environment transfer learning solution for location-aidedbeam prediction, whereby the ML model trained on a referencegNB is transferred to a target gNB by fine-tuning with a limiteddataset. Extensive simulation results based on ray-tracing in twourban environments show the effectiveness of our solution forboth inter- and intra-city model transfer. Our results show thatby training the model on a reference gNB and transferring themodel by fine-tuning with only 5% of the target gNB dataset,we can achieve 80% accuracy in predicting the best beamfor the target gNB. Importantly, our approach improves thepoor generalization accuracy of transferring the model to newenvironments without fine-tuning by around 75 percentage points.This demonstrates that transfer learning enables high predictionaccuracy while reducing the computational and training datasetcollection burden of ML-based beam prediction, making itpractical for 5G-and-beyond deployments.

Figures (6)

• Figure 1: Building layout (grey) and gNB locations (green) in our two urban network study areas.
• Figure 2: Top-1 BPL prediction: generalization and transfer learning accuracy using 5% fine-tuning dataset size across all gNB combinations of reference and target gNBs in (a)-(b) Frankfurt and (c)-(d) Seoul. The diagonal entries correspond to the accuracy of training and testing on the same gNB.
• Figure 3: Percentage coverage with respect to the total $500$$\text{m}$$\times$$500$$\text{m}$ study area and corresponding dataset size of the gNBs in (a) Frankfurt and (b) Seoul, differentiating between LoS and NLoS coverage.
• Figure 4: Top-1 transfer learning accuracy across the cities using 5% fine-tuning for gNB combinations of (a) reference gNBs in Seoul and target gNBs in Frankfurt and (b) vice-versa.
• Figure 5: Test accuracy vs. percentage of target gNB dataset used for fine-tuning in transferring between selected gNB combinations in Frankfurt. The plots also include the baseline accuracy of training and testing on target gNB.