QoS prediction in radio vehicular environments via prior user information

TL;DR

This work tackles the challenge of predicting quality of service (QoS) in high-mobility vehicular networks over look-ahead horizons of minutes. It proposes an ML tree-ensemble approach (XGBoost) to predict the downlink QoS $y(t+\tau)$, and innovates by augmenting the self-vehicle feature set with prior information from leading vehicles through PHY and cell measurements, using a delay-based look-ahead $\\tau_{sn}(t)= d_{sn}(t)/v_s(t)$. The study demonstrates that incorporating next-vehicle PHY/cell data substantially improves prediction accuracy, achieving roughly a 45% reduction in MRPE over a baseline that uses only self-vehicle data, and shows the ability to extend horizons up to around 8 minutes. Key findings indicate that leading-vehicle information and PHY feature correlations are valuable for proactive QoS management, even with a small number of vehicles, suggesting potential for reduced data collection in larger-scale deployments. The work lays a foundation for exploiting radio-environment priors in commercial networks to enable safer, more reliable autonomous and connected mobility.

Abstract

Reliable wireless communications play an important role in the automotive industry as it helps to enhance current use cases and enable new ones such as connected autonomous driving, platooning, cooperative maneuvering, teleoperated driving, and smart navigation. These and other use cases often rely on specific quality of service (QoS) levels for communication. Recently, the area of predictive quality of service (QoS) has received a great deal of attention as a key enabler to forecast communication quality well enough in advance. However, predicting QoS in a reliable manner is a notoriously difficult task. In this paper, we evaluate ML tree-ensemble methods to predict QoS in the range of minutes with data collected from a cellular test network. We discuss radio environment characteristics and we showcase how these can be used to improve ML performance and further support the uptake of ML in commercial networks. Specifically, we use the correlations of the measurements coming from the radio environment by including information of prior vehicles to enhance the prediction of the target vehicles. Moreover, we are extending prior art by showing how longer prediction horizons can be supported.

Figures (5)

• Figure 1: "Motorway A9 5G-ConnectedMobility" measurement campaign.
• Figure 2: Driving scenario with four vehicles following the same route on highway.
• Figure 3: Cross-correlation of self-vehicle with either vehicle 3 or vehicle 1 traveling at $\sim 3$ and $\sim8$ minutes gap respectively.
• Figure 4: mrpe in qos prediction for self-vehicle at look-ahead time 3 and 8 minutes with vehicle 3 and vehicle 1 as next vehicles respectively.
• Figure 5: Real($y$) vs. predicted ($\hat{y}$) qos comparison between models using phy layer features of self-vehicle and next vehicle.