Improve Federated Learning Stability for Vehicle Trajectory Prediction

The crowdsourced information is useful to calibrate Advanced Driver Assistance Systems/Autonomous Driving (ADAS/AD) parameters for automated and autonomous vehicles. However, learning such information in vehicular networks is challenging. On the one hand, data collected by individual vehicle may be not sufficient to train a large scale machine learning model. On the other hand, uploading raw data to cloud server is likewise impractical due to enormous communication bandwidth requirement and data privacy threat. This paper seeks a solution by applying federated learning (FL). We aim to improve FL algorithm stability to increase prediction accuracy. Accordingly, we propose a variance-based and structure-aware FL (VSFL), in which a variance-based model aggregation method is introduced for FL server to make optimal model aggregation and a structure-aware model training scheme is provided for vehicle clients to tackle statistical heterogeneity without compromising performance. We first provide theoretical analysis for the proposed VSFL. We then validate the effectiveness of VSFL algorithms on vehicle trajectory prediction using both synthetic data and real data.