Impact of RIS size on machine learning-enabled beam sweeping for user localization

This paper investigates machine learning (ML)-assisted user localization in reconfigurable intelligent surface (RIS)-aided millimeter-wave (mmWave) systems using minimal beam probing. A system model with three RIS apertures (10×10, 20×20, and 30×30) is considered, where the surface sequentially applies a limited set of probing beams, and the user equipment (UE) records the corresponding received power. Using ML regression, we train the regressor models to predict the UE positions from these measurements and update the RIS phase distribution
for efficient beam forming towards the intended UE. Simulations show that small RIS arrays achieve accurate predictions with limited probing, whereas larger apertures need deeper sweeps to curb outliers. The best approach attains mean errors below 1.5 dB for a 10 × 10 aperture and improves further on larger apertures with six probing beams. This underscores a trade-off among RIS size, probing overhead, and ML method choice for efficient RIS-aided localization in future sixth-generation (6G) wireless networks