Data-driven Optimization and Transfer Learning for Cellular Network Antenna Configurations

Mohamed Benzaghta ⋆ , Giovanni Geraci † ⁣ ⋆ {}^{\dagger\,\star} , David López-Pérez ♯ , and Alvaro Valcarce ♭ · 2025

Read Data-driven Optimization and Transfer Learning for Cellular Network Antenna Configurations by Mohamed Benzaghta ⋆ , Giovanni Geraci † ⁣ ⋆ {}^{\dagger\,\star} , David López-Pérez ♯ , and Alvaro Valcarce ♭ (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-2504-00825.txt.

Introduction

Mohamed Benzaghta and co-authors tune cell tower antennas with data and ray tracing, in work from 2025. Antenna tilt and beam width decide how well a network serves its weakest users. Hand tuning cannot search the vast setting space a real city presents.

A live London network supplied the test case, with a ray tracer modeling signal paths site by site. Optimized settings more than doubled the rates of the worst-served tenth of users. For drone users aloft, median rates rose fivefold with no loss for ground users. Transfer learning then carried a tuned solution to a new scenario cheaply. It matched fresh tuning speed with negligible loss. The flow suits operators facing many similar districts.

Catalog record

CREATORMohamed Benzaghta ⋆ , Giovanni Geraci † ⁣ ⋆ {}^{\dagger\,\star} , David López-Pérez ♯ , and Alvaro Valcarce ♭
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-00825

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Data-driven Optimization and Transfer Learning for Cellular Network Antenna Configurations Thanks: This work was supported by a) the Spanish State Research Agency through grants PID2021-123999OB-I00 and CEX2021-001195-M, b) the UPF-Fractus Chair on Tech Transfer and 6G, c) the Spanish Ministry of Economic Affairs and Digital Transformation and the European Union NextGenerationEU through actions CNS2023-145384, CNS2023-144333, and the UNICO 5G I+D SORUS project, d) the Generalitat Valenciana, Spain, through the CIDEGENT PlaGenT, Grant CIDEXG/2022/17, Project iTENTE, and e) HORIZON-SESAR-2023-DES-ER-02 project ANTENNAE (101167288). Mohamed Benzaghta ⋆ , Giovanni Geraci † ⁣ ⋆ {}^{\dagger\,\star} , David López-Pérez ♯ , and Alvaro Valcarce ♭ Abstract We propose a data-driven approach for large-scale cellular network optimization, using a production cellular network in London as a case study and employing Sionna ray tracing for site-specific channel propagation modeling. We optimize base station antenna tilts and half-power beamwidths, resulting in more than double the 10%-worst user rates compared to a 3GPP baseline. In scenarios involving aerial users, we identify configurations that increase their median rates fivefold without compromising ground user performance. We further demonstrate the efficacy of model generalization through transfer learning, leveraging available data from a scenario source to predict the optimal solution for a scenario target within a similar number of iterations, without requiring a new initial dataset, and with a negligible performance loss. I Introduction The large-scale optimization of cellular networks remains a significant challenge due to the complex interdependencies among settings across multiple cells. Coverage and capacity are heavily influenced by the configuration of base station (BS) antennas, with adjustments in parameters such as the tilt angle and half-power beamwidth (HPBW) being critical for optimizing signal strength and minimizing interference. This process, known as cell shaping, becomes increasingly difficult in large-scale networks, as interactions between cells lead to a non-convex and NP-hard optimization problem [ 1 ] . The conflicting goals of maximizing both coverage and capacity further complicate cellular network design. Coverage optimization involves directing energy toward cell edges. Capacity optimization prioritizes the signal-to-interference-plus-noise ratio (SINR) for users closer to the cell center. These challenges are exacerbated in scenarios involving nonhomogeneous user distributions, where uniform down-tilt configurations become ineffective [ 2 , 3 , 4 ] . Traditional methods for optimizing cellular networks, such as those employed in 3GPP frameworks, rely heavily on stochastic simulations and are typically limited to regular hexagonal deployments [ 5 ] . In real networks, site-specific radio frequency planning tools rely on time-consuming trial-and-error methods. A model-based framework for cellular network optimization, based on quantization theory and recently proposed in [ 6 ] , enables the fine-tuning of antenna parameters for each BS within a given deployment to achieve optimal coverage, capacity, or any trade-off thereof. In this paper, we propose an alternative approach that leverages available data to maximize real-world key performance indicators (KPIs), which are often mathematically intractable. Our main contributions can be summarized as follows: Data-driven optimization We employ high-dimensional Bayesian Optimization (HD-BO) to address a practical large-scale mobile network optimization problem, using a production cellular network in London as a case study. We employ a 3D representation of the area and use Sionna RT [ 7 ] to model site-specific channel propagation, considering the actual cell locations and configuration. To evaluate the effectiveness of our data-driven approach, we jointly optimize antenna tilts and HPBWs and identify configurations that achieve more than double the 10 %-worst rates with respect to a 3GPP baseline. Aerial connectivity In a second case study, we maximize user rates on the ground as well as along 3D aerial corridors for uncrewed aerial vehicle (UAV) users. Ensuring reliable air-to-ground connectivity demands configurations that cannot be easily determined through heuristic methods [ 8 , 9 ] . Our approach identifies configurations that improve UAV median rates by fivefold, without degrading ground performance. Transfer learning In alignment with the 3GPP vision on data-driven model generalization [ 10 ] , we use transfer learning to leverage knowledge from a previously optimized scenario ( scenario source ), to predict the optimal solution for a new scenario ( scenario target ). We demonstrate that, when the aerial corridor height changes from 140 m–160 m to 40 m–60 m, convergence occurs within a similar number of iterations, without the need for a new initial dataset, and with only a marginal 1% performance decline. II System Model In this section, we present the network deployment, channel model, and performance metrics utilized in our study. Cellular network topology We consider a site-specific scenario corresponding to a real-world production radio network owned by a leading commercial mobile operator in the UK. The network segment under consideration includes 16 deployment sites, with heights ranging from 22 to 56 meters. Each site consist of three sector antennas, resulting in a total of 48 cells across the network. The geographical area selected for our study covers approximately 1400 m by 1275 m and is situated in London, between latitudes [51.5087,51.5215] and longitudes [-0.1483,-0.1296] . Fig. 1 illustrates a 3D model of the selected area, highlighting several cell site locations. Fig. 1: A section of the area considered, with cell deployment sites indicated by black markers and 3D aerial corridors shown in blue. Antenna model We characterize the antenna configuration of each BS b by four main parameters: tilt \theta^{b} , bearing \phi^{b} , vertical HPBW \theta^{b}_{\text{3dB}} , and horizontal HPBW \phi^{b}_{\text{3dB}} . The tilt is defined as the angle between the antenna boresight and the horizon and can be electrically adjusted. The bearing indicates the horizontal orientation of each sector. The vertical HPBW is the angular range over which the antenna gain is above half of the maximum gain in the vertical plane, while the horizontal HPBW is the corresponding range in the horizontal plane. In this study, the bearings \phi^{b} and the horizontal HPBWs \phi^{b}_{\text{3dB}} are assumed fixed: the former as per the real cellular deployment, the latter set to \phi^{b}_{\text{3dB}}=65^{\circ} for all BSs. The tilts \theta^{b} and HPBWs \theta^{b}_{\text{3dB}} are the object of optimization. The normalized antenna gain for a specific pair of azimuth and elevation, \phi and \theta , between a BS b and a UE k , is [ 5 ] , and where the maximum antenna gain depends on the HPBW (e.g., \theta^{b}_{\text{3dB}}=10^{\circ} and \phi^{b}_{\text{3dB}}=65^{\circ} yield a maximum gain of 14 dBi [ 5 ] ). A(\phi,\theta)_{\textrm{dB}}=-\min\left\{-\left[A_{\text{H}}(\phi)+A_{\text{V}}(\theta)\right],25\right\}\,, (1) where A_{\text{H}}(\phi)_{\textrm{dB}}=-\min\left\{12\left[\left(\phi-\phi^{b}\right)/\phi_{\text{3dB}}\right]^{2},25\right\}\,, (2) A_{\text{V}}(\theta)_{\textrm{dB}}=-\min\left\{12\left[\left(\theta-\theta^{b}\right)/\theta^{b}_{\text{3dB}}\right]^{2},20\right\}\,. (3) Site-specific propagation channel A 3D representation of the selected area is constructed using OpenStreetMap, incorporating terrain and building information. The BSs are positioned and configured according to the actual cellular network topology. The omnidirectional large-scale channel gain (excluding the antenna gain) between BS b and UE k is calculated using Sionna RT [ 7 ] , a widely-used 3D ray-tracing tool for analyzing site-specific radio wave propagation. Simulations are conducted at a carrier frequency of 2 GHz. The material itu_concrete is used to model the permittivity and conductivity of all buildings. The maximum number of reflections and diffractions are set to 5 and 1, respectively. The total large-scale channel gain G_{b,k} is then obtained from the omnidirectional ray tracing channel gain by adding the antenna gain as per ( 1 ). User distribution We denote by {\mathcal{U}} the set of users under consideration. User

Read the complete text