[Paper Review] Predicting the Path Loss of Wireless Channel Models Using Machine Learning Techniques in MmWave Urban Communications

Saud Alhajaj Aldossari, Kwang‐Cheng Chen|arXiv (Cornell University)|May 2, 2020

Advanced MIMO Systems Optimization13 references4 citations

TL;DR

This paper proposes a machine learning-based approach to predict mmWave path loss in urban environments using regression models trained on wireless channel measurements, significantly reducing the need for extensive new measurements. By leveraging features like distance, delay, and angle of departure/arrival, the method achieves up to 72% R-squared accuracy in path loss prediction, demonstrating that models trained on one environment can generalize to others with high efficiency and reduced complexity.

ABSTRACT

The classic wireless communication channel modeling is performed using Deterministic and Stochastic channel methodologies. Machine learning (ML) emerges to revolutionize system design for 5G and beyond. ML techniques such as supervise leaning methods will be used to predict the wireless channel path loss of a variate of environments base on a certain dataset. The propagation signal of communication systems fundamentals is focusing on channel modeling particularly for new frequency bands such as MmWave. Machine learning can facilitate rapid channel modeling for 5G and beyond wireless communication systems due to the availability of partially relevant channel measurement data and model. When irregularity of the wireless channels lead to a complex methodology to achieve accurate models, appropriate machine learning methodology explores to reduce the complexity and increase the accuracy. In this paper, we demonstrate alternative procedures beyond traditional channel modeling to enhance the path loss models using machine learning techniques, to alleviate the dilemma of channel complexity and time-consuming process that the measurements were taken. This demonstrated regression uses the measurement data of a certain scenario to successfully assist the prediction of path loss model of a different operating environment.

Motivation & Objective

• To address the high complexity and measurement burden of traditional mmWave channel modeling in urban environments.
• To reduce the number of required channel measurements by leveraging machine learning to generalize models across different urban scenarios.
• To improve path loss prediction accuracy beyond classical deterministic and stochastic models using supervised learning techniques.
• To validate the effectiveness of regression models using real-world channel measurement data from micro- and macro-urban environments.

Proposed method

• The study employs supervised machine learning regression techniques—specifically linear and multiple linear regression—on measured mmWave channel data.
• Features used include transmitter-receiver separation distance (m), time delay (ns), received power (dBm), RMS delay spread (ns), and angles of departure/arrival (AoD/AoA) in degrees.
• The path loss model is formulated as a linear function of these features: $\hat{PL} = \alpha + L_0[\text{dB}] + \dots + X_\sigma[\text{dB}]$, where $L_0[\text{dB}]$ represents the reference path loss.
• The model is trained on data from a micro-urban (UMi) environment and tested for generalization to macro-urban (UMa) environments.
• Model performance is evaluated using standard metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R-squared ($R^2$).
• The approach enables transfer learning from one environment (e.g., UMi) to predict path loss in another (e.g., UMa), minimizing new measurement campaigns.

Experimental results

Research questions

• RQ1 Can a machine learning model trained on path loss data from one urban environment accurately predict path loss in a different urban environment?
• RQ2 How does the inclusion of multiple channel features (e.g., delay, angles, power) affect the accuracy of path loss prediction compared to single-feature models?
• RQ3 To what extent can regression models reduce the number of required channel measurements in mmWave urban communications?
• RQ4 What is the performance gain of multiple linear regression over simple linear regression in predicting mmWave path loss?

Key findings

• The multiple linear regression model using eight features (distance, delay, power, RMS delay spread, and AoD/AoA angles) achieved the best performance with an R-squared value of 0.72.
• The model trained on micro-urban (UMi) data successfully predicted path loss in macro-urban (UMa) environments, achieving a Mean Absolute Error (MAE) of 6.66 dB.
• The Root Mean Squared Error (RMSE) decreased from 11.25 dB (single-feature model) to 6.67 dB (eight-feature model), indicating improved prediction accuracy.
• The R-squared value increased from 0.21 (single feature) to 0.72 (eight features), showing that adding relevant channel features significantly enhances model fit and predictive power.
• The model demonstrated generalization capability, enabling accurate path loss prediction in new environments with minimal additional measurements.
• The study confirms that machine learning can effectively reduce the complexity and measurement burden of mmWave channel modeling while improving accuracy.