Machine Learning-based Path Loss Prediction in Suburban Environment in the Sub-6 GHz Band

Ferdaous Tarhouni, Muneer M. Al-Zu'bi, Mohamed‐Slim AlouiniarXiv 2025 · DOI · arXiv

TL;DR

This paper demonstrates that Machine Learning (ML) models, trained on synthetic data from high-fidelity ray-tracing simulations, can predict path loss in suburban environments with significantly higher accuracy than traditional empirical models. By using Random Forest and K-Nearest Neighbors algorithms, the researchers achieved precise signal strength predictions that generalize well even when tested against real-world measurement data from entirely different geographical locations.

What Problem Does This Paper Address?

In wireless communications, Path Loss (PL) is the reduction in power density of an electromagnetic wave as it propagates through space. Predicting PL accurately is the "bread and butter" of network planning—it determines where to place base stations and how to design antennas for 5G and 6G networks.

Currently, engineers face a trade-off:

  1. Empirical Models (e.g., Hata, COST-231): These are simple mathematical formulas. They are fast but often inaccurate because they don't account for specific local terrain or building layouts.
  2. Deterministic Models (e.g., Ray-Tracing): These simulate physical wave interactions (reflection, diffraction). They are highly accurate but computationally "expensive" and slow, making them difficult to use for real-time optimization.

The authors aim to bridge this gap by using ML to capture the accuracy of ray-tracing while maintaining the inference speed of a simple formula.

Methods

The researchers developed a pipeline that uses a "teacher" (Ray-Tracing) to train a "student" (ML models).

  • Synthetic Data Generation: They used the MATLAB RF propagation tool and the Shooting and Bouncing Rays (SBR) method to simulate signal behavior across the KAUST campus.
  • The ML Suite: They evaluated five different architectures:
  • Artificial Neural Networks (ANN)
    • Recurrent Neural Networks with LSTM (RNN-LSTM)
    • Decision Tree Regression (DTR)
    • Random Forest Regression (RFR)
    • K-Nearest Neighbors (KNN)
  • Feature Engineering: Instead of just distance, they fed the models eight features, including transmitter height, power, frequency, elevation angle, Line-of-Sight (LoS) status, and geographical coordinate differences.
  • Baselines: They compared these against industry standards like COST-231 Hata, Longley-Rice, and the Close-in (CI) model.

Datasets

The study utilizes two primary data sources:

  1. KAUST Synthetic Dataset (New): Generated specifically for this study, it consists of ~95,000 samples per site across three distinct locations on the KAUST campus (Suburban environment). It covers frequencies from 1.5 to 6 GHz.
  2. Rio de Janeiro Measurement Dataset: To prove their models weren't just "memorizing" the simulation, they validated them against real-world outdoor measurements taken in Brazil at 750 MHz, 2.5 GHz, and 3.5 GHz.

Key Insights and Ideas

  • The Power of Ensembles: Random Forest Regression (RFR) emerged as a top performer. Because RFR is an ensemble of decision trees, it naturally handles the complex, non-linear relationships of signal propagation (like a signal suddenly dropping behind a building) without requiring complex feature transformations.
  • Generalization is Possible: One of the most "aha" moments is that models trained on synthetic data from Saudi Arabia performed remarkably well on real-world data from Brazil. This suggests that ML can learn the underlying physics of suburban propagation rather than just site-specific quirks.
  • Empirical Models are Failing Modern Needs: The traditional COST-231 Hata model showed massive errors (RMSE > 35 dB) in these scenarios. This highlights that as we move toward higher frequencies and more complex suburban layouts, our old "rule-of-thumb" formulas are no longer sufficient.

Key Results

  • Synthetic Accuracy: On the simulated test sets, the Decision Tree (DTR) and Random Forest (RFR) models achieved a Root Mean Squared Error (RMSE) of approximately 1.35 dB and 1.44 dB, respectively. For context, the benchmark threshold for "good" prediction in this field is usually 6-7 dB.
  • Real-World Validation: When tested against the Brazil measurement data, the RFR model achieved an RMSE of 10.11 dB. While higher than the synthetic results, this outperformed existing hybrid models in literature (which ranged from 11.50 to 13.46 dB) and far surpassed traditional models like Okumura-Hata (17.25 dB).
  • Speed: The authors note that while training deep learning models (ANN/RNN) is time-consuming, once trained, they can generate predictions in seconds on standard consumer hardware.

Limitations and Future Work

The authors identify three main areas for improvement:

  • Vegetation: The current model uses a simplified representation of buildings and does not account for signal attenuation caused by trees and foliage, which is significant in suburban areas.
  • Dynamic Environments: The study assumes a static world. Real-world path loss fluctuates due to moving cars, weather changes, and human traffic.
  • Computational Complexity of Training: Generating the synthetic "ground truth" via ray-tracing still requires significant GPU power and time before the ML model can even begin learning.

Future research will focus on incorporating "vegetation loss" and more realistic terrain profiles into the training pipeline to further close the gap between simulation and reality.

Summary updated

Abstract

Accurate path loss (PL) prediction is crucial for successful network planning, antenna design, and performance optimization in wireless communication systems. Several conventional approaches for PL prediction have been adopted, but they have been demonstrated to lack flexibility and accuracy. In this work, we investigate the effectiveness of Machine Learning (ML) models in predicting PL, particularly for the sub-6 GHz band in a suburban campus of King Abdullah University of Science and Technology (KAUST). For training purposes, we generate synthetic datasets using the ray-tracing simulation technique. The feasibility and accuracy of the ML-based PL models are verified and validated using both synthetic and measurement datasets. The random forest regression (RFR) and the K-nearest neighbors (KNN) algorithms provide the best PL prediction accuracy compared to other ML models. In addition, we compare the performance of the developed ML-based PL models with the traditional propagation models, including COST-231 Hata, Longley-Rice, and Close-in models. The results show the superiority of the ML-based PL models compared to conventional models. Therefore, the ML approach using the ray-tracing technique can provide a promising and cost-effective solution for predicting and modeling radio wave propagation in various scenarios in a flexible manner.
BibTeX
@misc{tarhouni2025machine,
  author = {Ferdaous Tarhouni and Muneer M. Al-Zu'bi and Mohamed‐Slim Alouini},
  title = {{Machine Learning-based Path Loss Prediction in Suburban Environment in the Sub-6 GHz Band}},
  year = {2025},
  eprint = {2510.00696},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.00696},
  howpublished = {arXiv},
}
Figure