MPBA: Meta-Predictive Beam Alignment for mmWave Systems With Environmental Variability

Accurate and fast beam alignment is non-trivial in millimeter-wave (mmWave) communication systems. Deep learning models hold great promise to yield accurate beams within these systems. However, in practice, deep learning-based approaches suffer from generalization issues that the model trained under a given environment may not effectively adapt to new environments. This motivates us to propose a novel meta-learning-based algorithm to provide a general model for beam alignment, which aims to adapt to the unknown environments quickly. Concretely, this paper first proposes a Meta-Predictive Beam Alignment (MPBA) algorithm to assist model convergence and improve the quality of predicted beams. The MPBA simplifies the data partitioning process and training process when comparing traditional meta-algorithm, which is more conducive for actual deployment. This simplified and effective mechanism aids in identifying more general initial model parameters and transferring to new scattering environments without the need of heavy beam collection. Numerical simulation results reveal that the proposed MPBA has robust performance under various conditions and has about 21% accuracy gain and 2.2 bits/s/Hz spectral efficiency gain over direct transfer learning, as well as showcasing superior generalization capabilities when we transfer it to unseen base station (BS) environments.