AI-Driven Optimization of Wave-Controlled Reconfigurable Intelligent Surfaces
Gal Ben Itzhak, Graduate Student Member, IEEE,, Miguel Saavedra-Melo, Graduate Student Member, IEEE,, Ender Ayanoglu, Fellow, IEEE,, Filippo Capolino, Fellow, IEEE,, and A. Lee Swindlehurst, Fellow, IEEE · 2025
Read AI-Driven Optimization of Wave-Controlled Reconfigurable Intelligent Surfaces by Gal Ben Itzhak, Graduate Student Member, IEEE,, Miguel Saavedra-Melo, Graduate Student Member, IEEE,, Ender Ayanoglu, Fellow, IEEE,, Filippo Capolino, Fellow, IEEE,, and A. Lee Swindlehurst, Fellow, IEEE (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-07126.txt.
Introduction
Gal Ben Itzhak and co-authors present a data driven way to set a smart radio surface that can steer signals, in a paper from 2025. The surface uses standing waves on wires below it to set each cell, but real parts behave in messy ways that models miss. Exact setup by theory alone is very hard.
The team trained a neural network on measured signal patterns from the real surface. A search method then tuned the network to match the measured patterns well. A second search used the network to pick wave settings that gave a strong desired signal and low leak. The best settings were saved in a table for instant use when the surface runs.
Catalog record
| CREATOR | Gal Ben Itzhak, Graduate Student Member, IEEE,, Miguel Saavedra-Melo, Graduate Student Member, IEEE,, Ender Ayanoglu, Fellow, IEEE,, Filippo Capolino, Fellow, IEEE,, and A. Lee Swindlehurst, Fellow, IEEE |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 61 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-07126
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
AI-Driven Optimization of Wave-Controlled Reconfigurable Intelligent Surfaces Thanks: The authors are with the Center for Pervasive Communications and Computing (CPCC), Department of Electrical Engineering and Computer Science, University of California, Irvine. Thanks: This work is partially supported by NSF grant 2030029. Gal Ben Itzhak, Graduate Student Member, IEEE,, Miguel Saavedra-Melo, Graduate Student Member, IEEE,, Ender Ayanoglu, Fellow, IEEE,, Filippo Capolino, Fellow, IEEE,, and A. Lee Swindlehurst, Fellow, IEEE Abstract A promising type of Reconfigurable Intelligent Surface (RIS) employs tunable control of its varactors using biasing transmission lines below the RIS reflecting elements. Biasing standing waves (BSWs) are excited by a time-periodic signal and sampled at each RIS element to create a desired biasing voltage and control the reflection coefficients of the elements. A simple rectifier can be used to sample the voltages and capture the peaks of the BSWs over time. Like other types of RIS, attempting to model and accurately configure a wave-controlled RIS is extremely challenging due to factors such as device non-linearities, frequency dependence, element coupling, etc., and thus significant differences will arise between the actual and assumed performance. An alternative approach to solving this problem is data-driven: Using training data obtained by sampling the reflected radiation pattern of the RIS for a set of BSWs, a neural network (NN) is designed to create an input-output map between the BSW amplitudes and the resulting sampled radiation pattern. This is the approach discussed in this paper. In the proposed approach, the NN is optimized using a genetic algorithm (GA) to minimize the error between the predicted and measured radiation patterns. The BSW amplitudes are then designed via Simulated Annealing (SA) to optimize a signal-to-leakage-plus-noise ratio measure by iteratively forward-propagating the BSW amplitudes through the NN and using its output as feedback to determine convergence. The resulting optimal solutions are stored in a lookup table to be used both as settings to instantly configure the RIS and as a basis for determining more complex radiation patterns. I INTRODUCTION A Reconfigurable Intelligent Surface (RIS) is a promising building block for next-generation wireless networks, potentially solving issues related to limited coverage, sensitivity to blockages, high path loss, etc., by providing alternative propagation paths and enabling beamforming for desired directions [ 1 ] . An RIS is a type of metasurface [ 2 , 3 ] consisting of periodic arrangements of subwavelength-sized passive reflecting elements whose reflection coefficients can be individually controlled. The RIS architecture studied in this paper is illustrated in Fig. 1 , and is based on the general wave-controlled approach described in [ 4 ] . In this approach, biasing standing waves (BSWs) are excited on a transmission line (TL) behind the RIS structure and sampled at discrete locations to provide the DC bias voltages for varactor diodes that modulate the reflection coefficients. In a previous paper [ 5 ] , various optimization algorithms were developed that demonstrate the capability of the wave-controlled RIS approach to create beams and nulls in the radiation pattern for multiple desired directions. The performance of the optimization algorithms were demonstrated for two different types of voltage sampling circuits: a rectifier that is easier to implement but more difficult to use for optimization, and a sample-and-hold circuit, which is more costly to implement but enables analytical solutions to the beam pattern optimization problem. In this work, we consider the rectifier implementation, where the beam pattern optimization is nonlinear and constrained, and we assume the RIS is deployed in an unknown environment without knowledge of channel state information (CSI). Thus, there is no exact mathematical model that can be used to predict the response created by the RIS to an excitation by the BSWs. Consequently, we will turn to the use of machine learning (ML) tools to solve the RIS configuration problem. Fig. 1 : Wave-controlled RIS composed of two physical layers [ 5 ] . Top layer: M elements in each row along the x direction, where each element is connected to a varactor diode. Bottom layer: N BSWs are excited on each TL to create the biasing voltages sampled at each RIS element. Each row is controlled via the connection on the left where a signal with N adjustable frequency components is injected by a waveform generator. Adjacent metallic patches and varactors on the top layer, and adjacent DC voltage outputs on the bottom layer are uniformly separated by distance d_{x} in the x direction and d_{y} in the y direction. There exist many algorithms for RIS optimization in the literature, such as model-based, ML-based, heuristic, and combinations of ML / model-based and heuristic algorithms [ 6 ] . Model-based optimization in the physical environment is difficult, in particular because of arbitrary control of the individual RIS reflection coefficients not being possible due to the coupling between the RIS elements. Moreover, typically available mathematical models for the RIS behavior do not take into account factors such as frequency dependence or device nonlinearities. Many of the supervised ML algorithms in [ 6 ] have been demonstrated to work well for energy efficiency and data rate maximization tasks, though they often make assumptions about existing knowledge of the channels, which is limited in practice given the complexity of CSI acquisition with the passive RIS. There is also difficulty in architecting the neural networks for system modeling while avoiding overfitting, with some solutions including adding dropout layers (randomly ignoring neuron contributions) [ 7 ] , using larger datasets, and decreasing the number of hidden layers [ 8 ] . Though, the model performances are very sensitive to hyperparameter selection and will vary on a case-to-case basis, requiring time and experience in fine-tuning. It is desirable to have a systematic way to generate and evaluate the performance of models without manually optimizing them. In [ 9 ] , a deep-learning convolutional neural network was demonstrated to capture the relationships between RIS configurations and achievable rates at arbitrary receiver (Rx) locations, and an optimal set of discrete RIS phase shifts was recommended to optimize the propagation environment. This is done by an exhaustive search on an existing codebook of RIS phase shifts during inference, which adds additional delays. Different learning algorithms, such as unsupervised learning, are also feasible, as shown for the case of signal-to-noise ratio (SNR) maximization in the setting of a single user [ 10 ] . However, the size of the training dataset required for optimal network convergence increases significantly with the number of RIS elements, posing a challenge for the scalability of the RIS. In the approach presented in this paper, we address these shortcomings. In particular, we focus on the use of wave-controlled RIS, which as described above is a novel architecture that accommodates the physical limitations of an RIS and simplifies the hardware required to operate it. The wave-controlled RIS achieves reduced-dimension control of the M RIS elements by configuring N\ll M BSW amplitudes on a TL. This approach also accounts for the coupling between RIS elements by maintaining a relatively smooth variation in the voltage profiles across adjacent elements. Further, we develop an ML model to not only approximate the behavior of the wave-controlled RIS in a given environment, but also to optimize its radiation pattern. Using the data-driven approach discussed