Location-Aided Iterative Beam Search via Learning-based Generative Models

Beamforming is a fundamental technique that utilizes multiple antennas to enhance coverage, capacity, and efficiency in modern communication systems. Nevertheless, traditional beam management involves excessive measurement overhead to search and track the optimal communication beam. The availability of user location and historical data can be leveraged to train machine learning models, enabling the prediction of optimal beams and minimizing measurement overhead. However, existing solutions based on neural networks (NNs) face challenges in adapting to dynamic environments and varying wireless propagation conditions, as they cannot incorporate new data without undergoing retraining. This work introduces a probabilistic beam prediction method that relies on the generative model trained on measurements and location information. An iterative refinement approach further improves beam predictions by progressively incorporating data measured in the online phase. Results obtained from a real-world dataset show how the proposed method significantly reduces the measurement overhead while maintaining high prediction accuracy, outperforming existing NN-based solutions in challenging scenarios.