Source-linked AI summary
Indoor Signal Focusing with Deep Learning Designed Reconfigurable Intelligent Surfaces
Chongwen Huang, George C. Alexandropoulos, Chau Yuen, Mérouane Debbah
TL;DR
Indoor RIS deployments face complex configuration operations, motivating an efficient wireless alternative. The paper trains a DNN on coordinate fingerprints to predict RIS phases online, and simulations show improved focusing, throughput, and phase-matrix MSE performance.
Problem
Indoor signal focusing with deep-learning-configured RISs had not been reported, while existing RIS configuration approaches require complex operations.
Method
A fingerprint database trains a five-layer DNN to map measured user coordinates to RIS phase configurations maximizing received signal strength.
Results
The simulations show increased achievable rate at target locations and good MSE performance for predicted RIS phase matrices.
Takeaways & Limitations
The method provides efficient online wireless RIS configuration for indoor signal focusing at intended user locations.
Abstract
from arXiv · showhide
Reconfigurable Intelligent Surfaces (RISs) comprised of tunable unit elements have been recently considered in indoor communication environments for focusing signal reflections to intended user locations. However, the current proofs of concept require complex operations for the RIS configuration, which are mainly realized via wired control connections. In this paper, we present a deep learning method for efficient online wireless configuration of RISs when deployed in indoor communication environments. According to the proposed method, a database of coordinate fingerprints is implemented during an offline training phase. This fingerprinting database is used to train the weights and bias of a properly designed Deep Neural Network (DNN), whose role is to unveil the mapping between the measured coordinate information at a user location and the configuration of the RIS's unit cells that maximizes this user's received signal strength. During the online phase of the presented method, the trained DNN is fed with the measured position information at the target user to output the optimal phase configurations of the RIS for signal power focusing on this intended location. Our realistic simulation results using ray tracing on a three dimensional indoor environment demonstrate that the proposed DNN-based configuration method exhibits its merits for all considered cases, and effectively increases the achievable throughput at the target user location.
I. INTRODUCTION
Indoor coverage can be weak because building walls attenuate outdoor base-station signals, while RISs offer programmable reflections for focusing energy at intended locations. The paper addresses the unreported use of deep learning to configure RISs for indoor signal focusing and proposes a fingerprint-trained DNN.
- Indoor wall penetration loss can reduce outdoor base-station signal strength below acceptable performance.
- Deploying additional base stations or relays improves coverage but incurs location-finding, backhaul, power-consumption, and implementation costs.
- RISs use electronically controlled reflecting elements to alter incident electromagnetic fields and focus transmitted energy at intended 3D user locations.
- Deep learning has improved several applications, but indoor signal focusing with deep-learning-configured RISs had not been reported.
- The proposed DNN maps measured user positions to RIS unit-cell configurations that maximize received signal strength.
II. SYSTEM MODEL AND PROPOSED METHODOLOGY
The system consists of an indoor AP, a reflecting RIS, and users whose channels and positions determine the desired RIS phase configuration. Throughput is evaluated from the received-user SNR.
- The modeled 3D indoor environment contains an AP with M antennas, an N-element RIS, obstacles, reference points, and a target user.
- Active RIS elements support estimation of channels involving the RIS, while the paper leaves the challenging estimation task for future work.
- The model assumes 3D propagation from the AP and RIS reflections, with reference points and user locations on the room floor.
- The RIS phase matrix Φ is diagonal, with entries representing phase shifts applied by its reflecting elements.
- Achievable throughput is defined as R = log2(1+γ), where γ is the intended user's SNR.
B. Proposed Deep Learning Method
The proposed method replaces repeated complex RIS optimization with a two-phase position-to-configuration workflow. Offline fingerprints train the DNN, which predicts the RIS phases online for a measured target position.
- Existing channel-optimization approaches require complex, time- and power-consuming RIS reconfiguration during each channel coherence time.
- The method exploits the dependence of an intended user's optimal RIS phase configuration on that user's position.
- The proposed workflow has distinct offline training and online phases for RIS-based indoor signal-focusing improvement.
- Offline (Training) Phase: The offline phase records measured coordinates at reference positions together with their corresponding optimal phase configurations.
A. Deep Learning Basics
A feedforward DNN composes parameterized layer mappings to transform inputs into outputs. Dense layers use weight matrices, biases, and activation functions to implement these mappings.
- A feedforward DNN with L layers maps an input vector through L iterative processing steps.
- Each layer mapping depends on the preceding layer's output and that layer's parameter set.
- A dense layer connects every neuron in the layer to all neurons in the following layer.
- For a dense layer, parameters comprise a weight matrix and bias vector, while σ denotes the activation function.
B. Fingerprinting Database Design
The fingerprinting database pairs repeated coordinate estimates at reference points with their corresponding optimal RIS phase configurations for DNN training.
- Each database set gathers the j-th coordinate estimate from all K reference positions.The coordinate vector is denoted p_k,j = [x_k,j y_k,j].
- The database is formed from J distinct estimates for each of K reference-point positions.
- Every coordinate estimate is paired with its corresponding optimal RIS phase configuration Φ_k,j.
C. Proposed DNN Design
The proposed five-layer DNN maps estimated user coordinates to an RIS phase matrix that approximates the optimal configuration for indoor signal focusing.
- The five-layer architecture contains an input layer, three hidden layers, and an output layer.
- Estimated user position coordinates enter the DNN, whose hidden layers learn their relationship with the optimal RIS phase configuration.
- The output layer provides the estimated phase matrix Φ̂ by minimizing its difference from the desired optimal phase matrix Φ.
- The network uses tanh activation functions because they support nonlinear approximation and provide full mapping in [−1, 1].The passage also states that tanh leads to stronger gradients than sigmoid.
D. Proposed DNN Training Process
The DNN is trained offline on coordinate–phase pairs using MMSE loss and stochastic-gradient optimization, with Adagrad used for gradient-based parameter updates.
- Training uses KJ measured coordinate inputs and corresponding optimal phase-matrix outputs gathered from the fingerprinting database.
- The DNN parameters are optimized with a Minimum Mean Squared Error criterion between predicted and desired phase configurations.
- The predicted phase configuration Φ̂_k,j depends on the five layers’ weights and bias parameters Θ.
- Stochastic-gradient descent is used to minimize the loss with small training batches and reasonable computational complexity.
- Adagrad with backpropagation computes efficient gradient updates for the DNN design.
IV. NUMERICAL RESULTS
The numerical evaluation uses ray-traced simulations in a three-dimensional office environment with reference and test points, obstacles, and a RIS deployment.
- The simulated office measures 6 m in width, 10 m in length, and 3 m in height.
- The evaluation layout includes RIS and obstacle deployments together with reference and test points on the room floor.
- The fingerprinting database uses estimated coordinates from 240 different reference-point positions.
- The setup assumes timely online user-position estimation, independent Rayleigh fading, and 200 Monte Carlo runs for average results.
- To isolate RIS contributions, only the RIS provides signal reflections in the simulated room.
B. Performance Evaluation
The evaluation examines throughput-based signal focusing and DNN phase-matrix estimation across user offsets, SNR values, and RIS sizes. RIS assistance improves target-user throughput, while the DNN achieves good MSE performance after relatively few training epochs.
- Figure 5 compares throughput versus transmit SNR at the target position, three user distances, and the non-RIS target-user case.
- RIS deployment substantially improves achievable throughput at the target user compared with non-RIS transmission.
- The DNN achieves good RIS phase-configuration MSE performance after around 20 training epochs for all tested reflecting-element counts.
- N = 8 yields the minimum MSE but requires a few more epochs for convergence than the other tested RIS sizes.
- Smaller N reduces estimation error because fewer RIS phase values must be estimated.
V. CONCLUSION
The paper presents a DNN-based method for online wireless RIS configuration in indoor communication environments. Simulations in a 3D office environment show effective signal focusing and good phase-matrix MSE performance.
- The proposed method uses offline fingerprinting to train a DNN that maps user position to the RIS unit-cell configuration maximizing received signal strength.
- During online operation, estimated user coordinates are fed to the DNN to produce the desired RIS configuration.
- Simulations in a 3D office environment show effective signal focusing at intended user locations and good RIS phase-matrix MSE performance.