[Paper Review] MmWave Beam Prediction with Situational Awareness: A Machine Learning Approach

Yuyang Wang, Murali Narasimha+1 more|arXiv (Cornell University)|May 23, 2018

Millimeter-Wave Propagation and Modeling14 references3 citations

TL;DR

This paper proposes a machine learning framework that leverages situational awareness—specifically vehicle locations and surrounding traffic conditions—to predict mmWave beam power with minimal feedback overhead. By using regression models trained on ray-traced channel data, the approach achieves near-zero overhead beam selection with throughput loss of less than 2% compared to ideal beam training.

ABSTRACT

Millimeter-wave communication is a challenge in the highly mobile vehicular context. Traditional beam training is inadequate in satisfying low overheads and latency. In this paper, we propose to combine machine learning tools and situational awareness to learn the beam information (power, optimal beam index, etc) from past observations. We consider forms of situational awareness that are specific to the vehicular setting including the locations of the receiver and the surrounding vehicles. We leverage regression models to predict the received power with different beam power quantizations. The result shows that situational awareness can largely improve the prediction accuracy and the model can achieve throughput with little performance loss with almost zero overhead.

Motivation & Objective

• Address the challenge of high-latency and high-overhead beam training in mmWave vehicular networks due to rapid blockages and mobility.
• Leverage situational awareness from vehicle sensors (GPS, radar, V2X) to predict optimal beam configurations without relying on real-time beam training.
• Reduce feedback overhead in mmWave beam selection by replacing traditional training with learned beam power predictions based on vehicle locations.
• Evaluate the impact of CQI quantization granularity and situational awareness levels on prediction accuracy and beam selection performance.
• Demonstrate that multi-beam power regression outperforms classification-based beam selection in terms of achieved throughput.

Proposed method

• Use ray-tracing simulations (Wireless Insite) to generate mmWave channel data in a two-lane urban canyon with fixed buildings and moving vehicles.
• Model the mmWave channel using geometric channel modeling based on ray-tracing outputs, extracting path parameters such as angles of arrival/departure and path gains.
• Train regression models (e.g., random forest, gradient boosting) to predict the received power of each beam pair in the codebook using vehicle locations as input features.
• Incorporate situational awareness features including receiver location, vehicle type (truck vs. car), and surrounding vehicle density to improve prediction accuracy.
• Apply CQI quantization to the beam power predictions to reduce feedback overhead, with quantization parameters tuned based on dataset statistics.
• Compare beam selection performance using regression-based power prediction versus classification-based beam index prediction in terms of alignment probability and throughput.

Experimental results

Research questions

• RQ1 Can situational awareness—specifically vehicle locations and surrounding traffic conditions—significantly improve the accuracy of mmWave beam power prediction in vehicular networks?
• RQ2 How does CQI quantization granularity affect the prediction accuracy and beam selection performance in mmWave beam training with machine learning?
• RQ3 Does multi-beam power regression provide better beam selection performance than beam index classification in terms of throughput and alignment probability?
• RQ4 To what extent can feedback overhead be reduced while maintaining high beam alignment accuracy using learned beam prediction models?
• RQ5 What is the optimal configuration of CQI quantization bounds (P_u, P_l) and resolution (r_CQI) for maintaining high prediction accuracy in real-world mmWave V2X scenarios?

Key findings

• Full situational awareness, including vehicle locations and surrounding traffic, significantly improves beam power prediction accuracy compared to using only receiver location.
• With CQI quantization granularity of 1 dBm or less, the prediction error remains below 1 dBm for over 50% of predictions, showing minimal performance degradation compared to full-resolution prediction.
• The upper bound P_u has a greater impact on prediction accuracy than the lower bound P_l, as beam power is typically high and underestimation due to small P_u leads to large errors.
• Regression over all beam pair powers achieves higher throughput (98.8%) than classification (98.4%) despite slightly lower alignment probability, due to better beam order discrimination.
• Even with 5 dBm quantization granularity, the achieved throughput remains high at 97.8%, indicating that coarse CQI quantization has minimal impact when resolution is sufficient.
• The model achieves near-zero feedback overhead by replacing traditional beam training with learned predictions, maintaining throughput within 2% of ideal beam training performance.