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

Millimeter-wave (mm-wave) communications require beamforming and consequent precise beam alignment between the gNodeB (gNB) and the user equipment (UE) to overcome high propagation losses. This beam alignment needs to be constantly updated for different UE locations based on beamsweeping radio frequency measurements, leading to significant beam management overhead. One potential solution involves using machine learning (ML) beam prediction algorithms that leverage UE position information to select the serving beam without the overhead of beam sweeping. However, the highly site-specific nature of mm-wave propagation means that ML models require training from scratch for each scenario, which is inefficient in practice. In this paper, we propose a robust cross-environment transfer learning solution for location-aided beam prediction, whereby the ML model trained on a reference gNB is transferred to a target gNB by fine-tuning with a limited dataset. Extensive simulation results based on ray-tracing in two urban environments show the effectiveness of our solution for both inter- and intra-city model transfer. Our results show that by training the model on a reference gNB and transferring the model by fine-tuning with only 5 % of the target gNB dataset, we can achieve 80 % accuracy in predicting the best beam for the target gNB. Importantly, our approach improves the poor generalization accuracy of transferring the model to new environments without fine-tuning by around 75 percentage points. This demonstrates that transfer learning enables high prediction accuracy while reducing the computational and training dataset collection burden of ML-based beam prediction, making it practical for 5G-and-beyond deployments.