Efficient angular localization in RIS-assisted mmWave network via learning-based estimations

Among the candidate spectrum bands for sixth- generation (6G) networks, millimeter-wave (mmWave) frequencies are widely considered for high-capacity links, yet exhaustive beam sweeping introduces significant overhead in reconfigurableintelligent surface (RIS)-assisted systems. This paper proposes a lightweight machine learning (ML) regression framework that directly maps received power measurements from a small set of sector-aligned probing beams to the user equipment (UE)elevation angle, enabling RIS phase optimization without explicit channel state information (CSI). Four classical regressors—K- nearest neighbors (KNN), decision tree (DT), random forest (RF), and support vector regressor (SVR) are evaluated underan angle-stratified protocol across different beam configurations. KNN consistently achieves the best performance, attaining a mean absolute error (MAE) of 0.79 degrees and R2 of 0.99 with six probing beams, while reducing beam-probing overhead by over 93% compared to exhaustive search. The predicted angles are used to construct 1-bit RIS phase matrices, yielding radiation patterns within 1–2 degrees of the ground-truth direction, confirming that accurate angular localization translates directlyinto effective beamforming.