KPI-oriented Probing Beam Selection and Beam Prediction for mmWave MISO communications

Millimeter wave (mmWave) communications present significant challenges in beam alignment and tracking (BA/T) due to the substantial overhead and delays associated with beam training in dynamic environments. Deep learning (DL) has emerged as a promising solution by exploiting spatial channel correlations. This work addresses the problem of data-driven probing beam selection and beam prediction, where a trained prediction model utilizes selected probing beams to optimize BA/T. Given the NP-hard combinatorial optimization problem, we propose an iterative strategy to reduce computational complexity, with each iteration focusing on selecting a probing beam. The training of the prediction models and the beam selection process are guided by a targeted key performance indicators (KPI). Extensive simulations and tests using both measured and simulated data from line of sight (LoS) and non-line of sight (NLoS) scenarios validate the effectiveness of the proposed strategies. The results show marked improvements in targeted KPIs, in NLoS environments such as DeepMIMO and 3GPP scenarios.