Machine Learning-Aided 3D User Localization with STAR-RIS using SNR-Based Refinement

Abstract

This paper presents a reconfigurable intelligent surface (RIS)-assisted three-dimensional (3D) localization system that combines machine learning (ML) with optimization tech-niques for enhanced positioning accuracy in indoor environments. The proposed framework employs a multi-layer perceptron (MLP) regressor trained on comprehensive spatial data to pro-vide initial position estimates, followed by a sophisticated two-stage refinement algorithm that incorporates time-of-flight, angle-of-arrival, and signal-to-noise ratio constraints. To validate its effectiveness, we compare the proposed MLP+Refinement method against a few key state-of-the-art alternatives, such as, geometric localization, optimization-only, random forest (RF), XGBoost (XGB), and their refined versions. Experimental results demon-strate that the proposed MLP+Refinement method consistently outperforms all baselines, achieving the lowest mean error of only 0.0285 m, which represents a 65% reduction over the raw MLP prediction and a 99.7% improvement over the optimization-only baseline. The method also achieves an 84.5% refinement success rate while maintaining computational efficiency suitable for real-time applications.