vLOFT: Variable Learning Rate for Online Federated Traffic Prediction

Deep learning (DL) models have established remarkable success in traffic state estimation (TSE); however, the widely-studied approaches' centralized nature presents deployment challenges in Intelligent Transportation Systems (ITS), especially regarding data ownership, privacy, and response times. Online models trained using Federated Learning (FL) approaches can harvest the rich real-time data streams in ITS for continuously updating the TSE models. Though, there are limitations during the online learning process, such as model drift. In this paper, we propose a variable learning rate (VLR) scheduling approach that integrates online FL with cooperative edge computing (CEC). Continuously collected local traffic data by roadside units (RSUs) and connected sensors train localized models, and only the updated model parameters are shared instead of raw data. VLR adjustments within the Adaptive Gradient Algorithm (ADAM) optimizer framework adjust the learning rate based on traffic patterns. We conduct extensive experimental analysis and compare example VLR strategies with a control model. The results indicate that the best-performing VLR strategy reduced the average mean absolute error (MAE) across all detectors by about 10.29%.