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

2025-04-01

AI Summary

To address the joint rate optimization challenge for terrestrial and aerial users in large-scale cellular networks, this paper proposes a data-driven black-box optimization framework integrated with meta-learning-based transfer learning for joint antenna parameter optimization—specifically, base station downtilt and half-power beamwidth. Leveraging real-world London network measurements and Sionna-based ray-tracing modeling, our approach achieves, for the first time, rapid cross-scenario generalization without requiring initial training data. Under the constraint of preserving terrestrial user performance, the 10% worst-performing terrestrial users experience over 2× rate improvement, while the median aerial user rate increases by 5×. Post-transfer convergence speed matches that of the source scenario, with negligible performance degradation. The core contribution is a lightweight, transferable antenna configuration optimization framework tailored for integrated air–ground communications.

Technology Category

• Search and Optimization: Learning to Search
• Machine Learning: Optimization
• Computer Vision: Learning & Optimization for CV

Application Category

• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Search and Retrieval-Augmented AI: Vertical and domain-specific search

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.

Problem

• Optimize cellular network antenna tilts and beamwidths
• Improve rates for worst and aerial users significantly
• Enable transfer learning for efficient model generalization

Innovation

• Data-driven optimization for cellular networks
• Transfer learning for model generalization
• Ray tracing for channel propagation modeling