Impact of federated deep learning on vehicle-based speed control in mixed traffic flows

This study investigates the application of Federated Learning (FL) in an environment of mixed traffic flow with Connected and Automated Vehicles (CAVs). The focus of this study is set on the CAV cruise control speed adjustment which is enforced by the Intelligent Speed Adaptation (ISA) system. It is modelled as an Actor-Critic-based learning architecture with the goal of computing the vehicle speed based on the data of the nearby vehicles. Some CAVs achieve poor travel time performance in comparison to others. Thus, the proposed FL framework selects periodically best-performing CAVs which are local clients in the context of FL. The updated global model is computed based on the weighted averaging of the best-performing models of clients. The use case environment is modelled as a simulation of a synthetic motorway with an induced capacity drop in the form of a work-zone under moderate and high traffic demand. It is shown that the FL framework can stabilise the individual client learning convergence and reduce their travel time compared to the local learning approach. The overall motorway throughput is also increased by the proposed FL framework. This indicates that the global model of FL can utilise the optimal ratio between generalisation and bias towards the most effective control policy of individual speed control. Results show that those effects are more prominent under moderate traffic demand.