Integrated Probing-Beam Pattern Learning and Beam Prediction for mmWave Massive MIMO

With the widespread adoption of massive multiple-input multiple-output (MIMO) and millimeter wave (mmWave) communication techniques, the overhead of beam measurement and the complexity of beam management become even more severe issues due to the dramatic increase of the number of beams. Traditional methods often overlook the selection of optimal probing beams, thus limiting beam prediction performance. Additionally, existing solutions based on deep learning exhibit high complexity, which often hinders their practical deployment. In this work, we propose a lightweight, integrated neural network approach tailored for joint probing-beam pattern selection and beam prediction. Specifically, our solution includes two main components. First, formulating the selection of probing beams as a sampling operation, we envision a sampling network where, to enable gradient back-propagation of network parameters in spite the non-differentiable nature of sampling, the standard sampling function is approximated with a fitting function. Then, drawing inspiration from the physical structure of antenna arrays and 3D beam formation, we develop a beam-prediction network based on convolutional neural networks and self-attention mechanisms. Experimental results demonstrate that, thanks to the learned pattern, our proposed scheme achieves very good prediction performance (exceeding the state of the art by over 15% in top-1 accuracy), using a neural network with almost 90% less parameters than existing machine learning-based solutions.