Page 800 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

vehicle’s measurements. This raises an important question: to what extent can past user
experiences help predict the propagation environment? To explore this, we evaluate our ML
model for QoS prediction not only with data collected from a single leading vehicle but also
by incorporating insights from prior users. Our findings indicate that by leveraging data from
the leading vehicle, QoS prediction accuracy improves significantly for following vehicles on
the same route.

Furthermore, we employ Explainable AI (XAI) to demonstrate that ML models can learn wireless
network conditions without pre-programming [5]. Our dataset is collected from a real-world
network, covering platooning vehicular scenarios for multi-radio environments.

This work lays the foundation for reinforcing the following benefits:
1. Platooning promises higher fuel efficiency due to reduced aerodynamic drag and
increased overall traffic efficiency because of smoother velocity control.
2. Clustering more vehicles in a platoon optimizes costly data collection and reduces the
need for extensive data for pQoS evaluation
3. pQoS-based decision for full/partial task offloading of expensive computing tasks

System Architecture:

Our proposed solution incorporates two E2E pipelines, along with several key modules,
designed to optimize the decision-making process for safe and efficient vehicle mobility in a
V2X network.

Pipelines:

1. Master Pipeline: The master pipeline is responsible for data collection, cluster formation,
and QoS prediction. The Software-Defined Network (SDN) Controller at the base station
aggregates network data, including GPS information from the vehicles, and shares it with
the clustering engine and the pQoS engine as shown in Fig. 1. First, the clustering engine
performs a stationarity test on the data, forms a vehicle cluster, and elects a leader vehicle.
The cluster data, including the number of vehicles and the leader vehicle's identity, is
then sent to the respective vehicles. Later, the pQoS engine predicts the network’s QoS
and relays this information to the leader vehicle, which subsequently uses it to make task
offloading decisions.
2. Invoked Pipeline: The invoked pipeline focuses on task offloading. The leader vehicle
evaluates whether the predicted QoS is adequate to offload tasks to a task offloading
manager co-located with a nearby Multi-Access Edge Computing (MEC). If the QoS
meets the requirements, the leader offloads its computational tasks to the task offloading
manager. These computationally intensive tasks are processed, and the task offloading
manager generates a decision. This decision is sent back to the leader, who then
communicates the outcome to the other vehicles in the cluster.

Modules:

1. Data Collector: The SDN Controller at the base station functions as the Data Collector. It
aggregates information from the vehicles, including GPS data, speed, and other relevant
parameters, as well as network-related data.
2. Decision Engine: It contains three modules, such as clustering engine, pQoS engine, and
task offloading manager.

• Clustering Engine: The clustering engine is tasked with forming dynamic clusters,
electing a leader vehicle, and ensuring safe distances between vehicles through

764