[Paper Review] Reconfigurable Intelligent Surfaces and Machine Learning for Wireless Fingerprinting Localization

Cam Ly Nguyen, Orestis Georgiou+1 more|arXiv (Cornell University)|Oct 7, 2020

Indoor and Outdoor Localization Technologies19 references21 citations

TL;DR

This paper proposes a machine learning-enhanced wireless fingerprinting localization system using Reconfigurable Intelligent Surfaces (RIS) to generate diverse, easily differentiable radio maps. By applying feature selection to reduce RIS state space complexity, the method achieves sub-meter localization accuracy with reduced training time and grid resolution, outperforming heuristic and random configuration selection.

ABSTRACT

Reconfigurable Intelligent Surfaces (RISs) promise improved, secure and more efficient wireless communications. We propose and demonstrate how to exploit the diversity offered by RISs to generate and select easily differentiable radio maps for use in wireless fingerprinting localization applications. Further, we apply machine learning feature selection methods to prune the large state space of the RIS, thus reducing complexity and enhancing localization accuracy and position acquisition time. We evaluate our proposed approach by generation of radio maps with a novel radio propagation modelling and simulations.

Motivation & Objective

• To enable high-accuracy indoor wireless localization in RIS-empowered environments where traditional methods fail due to ambiguous signal origins.
• To address the challenge of high complexity and large state space in RIS-based radio map generation for fingerprinting.
• To reduce localization error and position acquisition time by leveraging machine learning for optimal RIS configuration selection.
• To validate the effectiveness of machine learning-based feature selection (ML-FS) against heuristic and random RIS configuration selection methods.

Proposed method

• The system uses a novel radio propagation model based on impedance matrices of thin wire antennas to simulate composite EM fields from AP and RIS.
• Radio maps are generated by simulating 50 distinct RIS configurations: 10 with uniform impedance (planar reflection), 10 with linearly increasing impedance (beamforming), and 30 with random impedance (diffuse scattering).
• A machine learning feature selection (ML-FS) approach is applied to identify the most informative RIS configurations from the 50 candidates, reducing the effective state space.
• The method employs k-NN, Neural Networks, and Random Forests for localization, with k-NN showing the best performance under the tested conditions.
• Heuristic State Selection (HSS) and random configuration sampling are used as baselines for comparison.
• The system evaluates performance across two grid resolutions: 100 points (2×2 m) and 400 points (1×1 m), measuring localization error via cumulative distribution functions and mean error.

Experimental results

Research questions

• RQ1 Can RIS reconfiguration be used to generate diverse, differentiable radio maps that enhance fingerprinting-based localization accuracy?
• RQ2 Does machine learning-based feature selection outperform heuristic or random selection of RIS configurations in reducing localization error and complexity?
• RQ3 To what extent can the resolution of the radio map be reduced (e.g., from 400 to 100 grid points) while maintaining sub-meter localization accuracy through optimal RIS configuration selection?
• RQ4 How do different machine learning algorithms (k-NN, NN, RF) perform in RIS-aided fingerprinting localization under varying RIS configuration sets?
• RQ5 What is the trade-off between radio map resolution and the number of RIS configurations (M) in achieving a target localization error?

Key findings

• The k-NN algorithm with machine learning feature selection (ML-FS) achieved the best localization accuracy among the tested models, with a mean error below 1 meter when using M=12 configurations and L=100 grid points.
• ML-FS significantly outperformed both the heuristic state selection (HSS) and random configuration selection, with HSS performing worse than random sampling due to outlier bias in the distance metric.
• A mean localization error of 2 meters was achievable with either 400 grid points and M=22 random configurations, or with only 100 grid points and M=12 using ML-FS, demonstrating a key trade-off between resolution and configuration complexity.
• The cumulative distribution function (CDF) results showed consistent improvement in localization error across all algorithms when using feature selection, with k-NN showing the most pronounced gain.
• The proposed method enables sub-meter localization accuracy without requiring dense radio map sampling or multiple access points, by exploiting RIS reconfigurability to generate diverse, informative radio maps.
• The results confirm that RIS reconfiguration can be leveraged as a dynamic tool to enhance fingerprinting localization, reducing both time and computational complexity while maintaining high accuracy.