Adaptive federated reinforcement learning for critical realtime communications in UAV assisted vehicular networks
J. Hao, Rola Naja, Djamal Zeghlache
Abstract
This paper sheds the light on road active safety measurements implemented in unmanned aerial vehicles assisted vehicular networks. Despite the great potential of deploying high computing drones, the drone battery life is the major concern, on one hand. On the other hand, road active safety is a critical real-time process that should be tackled in a tight time window in vehicular networks. To meet the mentioned concerns, we adopt federated machine learning on the local vehicles, sending local updates to drone servers. Moreover, a dynamic frequency adaptation framework is proposed to achieve the optimal trade-off between the road active safety performance and drone’s energy consumption. The thresholds for the local update frequency are calibrated according to road safety measurements (i.e., collision rate, risky and impolite driving time on the road) and drone energy consumption. Additionally, an accurate mathematical modeling based on M/G/1 multi-class was conducted in order to access the queuing time at the drone.