Abstract:This paper introduces a novel learning-based methodology for Direction of Arrival (DoA) estimation in environments featuring an Intelligent Reflecting Surface (IRS) or Reconfigurable Intelligent Surface (RIS), particularly under challenging Non-Line-of-Sight (NLoS) conditions. The core innovation lies in the integration of a specialized Neural Network (NN) layer, termed the NN-based RIS layer, within a broader network architecture. The NN-based RIS layer is meticulously designed to autonomously learn the optimal phase shifts for the RIS, specifically tailored to enhance the accuracy of DoA estimation. In this framework, pre-processed real-valued input observations are directed into the RIS layer, which possesses a distinct and specialized structure. Unlike conventional neural network layers, the NN-based RIS layer imposes constraints on its weights, requiring them to be sinusoidal functions. The phase arguments of these sinusoidal functions serve as the tunable parameters during the training process, effectively enabling the layer to simulate the physical behavior of an RIS. To accommodate the unique properties of the NN-based RIS layer, the standard feed-forward and back-propagation algorithms are adapted, ensuring that the learning process aligns with the specialized requirements of the RIS-based architecture. This innovative approach allows the network to seamlessly integrate the RIS functionality into the DoA estimation process. Extensive numerical simulations validate the effectiveness of the proposed machine learning-based technique, demonstrating its superior performance compared to traditional non-learning-based methods across a range of RIS-assisted scenarios. The results underscore the potential of this approach to significantly enhance DoA estimation accuracy in complex wireless communication environments, paving the way for future advancements in RIS-enabled systems.

Enhanced‐Resolution Learning‐Based Direction of Arrival Estimation by Programmable Metasurface

Machine–learning-enabled metasurface for direction of arrival estimation

Programmable Metasurface RCS Prediction under Obstacles Based on DNN

A comprehensive review of metasurface-assisted direction-of-arrival estimation

Spatial Sectorized Neural Network for 2-D DOA Estimation in the Full Azimuth

Programmable Meta‐Reflector for Multiple Tasks in Intelligent Connected Environments

Space-Time-Coding Metasurface for Direction-of-Arrival Estimation

Two-Dimensional Direction-of-Arrival Estimation Using Stacked Intelligent Metasurfaces

Joint Detections of Frequency and Direction of Arrival in Wideband Based on Programmable Metasurface

Deep learning estimation of complex reverberant wave fields by a programmable metasurface

Stacked Intelligent Metasurface Performs a 2D DFT in the Wave Domain for DOA Estimation

Metasurface-Enabled High-Resolution Localization of Non-Cooperative Wireless Source in an Indoor Environment

DeepAoANet: Learning Angle of Arrival From Software Defined Radios With Deep Neural Networks

Super-resolution diffractive neural network for all-optical direction of arrival estimation beyond diffraction limits

Neural Network-Based IRS Assisted NLoS DoA Estimation

On the Application of Deep Learning in Modeling Metasurface

Single-pixel polarimetric direction of arrival estimation using programmable coding metasurface aperture

SDOA-Net: An Efficient Deep-Learning-Based DOA Estimation Network for Imperfect Array

Polarized Intelligent Reflecting Surface Aided 2D-DOA Estimation for NLoS Sources

A Physics Based Approach for Neural Networks Enabled Design of All-Dielectric Metasurfaces

On-demand Doppler-offset Beamforming with Intelligent Spatiotemporal Metasurfaces