Code, experiment results, and trained models accompanying the paper "Cloud-Fog-Edge Deployment of ML-Based Traffic Speed Prediction: An Empirical Analysis of Performance, Cost, and Scalability" (University of Manchester).
| Path | Contents |
|---|---|
paper_artifacts/ |
Start here for reviewer materials: trained models + scalers, the model evaluation document, the observed monitoring metrics document, and a SHA-256 manifest. |
experiments/ |
The cross-platform experiment system (runner, training/inference/feature-engineering phases, per-platform configuration). |
results/ |
Sample experiment result records (per-run JSON with wall-clock, accuracy, and resource telemetry). |
model/ |
Shared data-loading, feature-engineering, and evaluation modules used by the experiment. |
tools/ |
Configuration template (credentials blanked). |
One ~142k-parameter multilayer perceptron per sensor direction predicts five-minute average traffic speed from 15 calendar and event features, for 37 sensor directions in Greater Manchester (Dec 2020 – Feb 2025 data). The identical training and inference pipeline is evaluated across six platforms — HPC cloud GPU (A100), a fog workstation (RTX 4090), simulated Jetson / roadside-unit / Raspberry Pi edge tiers, and an AWS t2.medium instance — over four sensor-count tiers and multiple training scenarios (initial training, cold-start re-training, and warm-start re-training over recent and progressive windows).