Time-dependent and Privacy-Preserving Decentralized Routing using Federated Learning
With rapid urbanization, route planning is gaining more importance. As transportation networks become more complex and mobility in our society more important, the demand for efficient methods in route planning increases even further. State of the art solutions for route planning in a time-dependent network assumes a centralized approach, where parallelization of the search algorithm uses a shared memory model. Hence, its deployment is limited to multiprocessing environment such as in a data center, where it is assumed that a shared memory allows a constant time direct communication between each pair of processors. It is not well suited for a distributed setting, which is prone to communication failures and can incur higher response times due to network latency. Furthermore, storing and using location data of users raises privacy concerns. This thesis describes a resilient, decentralized approach for route planning in a time-dependent network, where the computing devices have limited resources and operate in an environment with intermittent network connectivity. We leverage recent advances in federated learning to collaboratively learn shared prediction models online while keeping all the training data on the device, thus preserving privacy. As everyday devices are becoming more powerful, our approach can effectively tackle the urban routing problem by harnessing the device resources. Our approach will particularly help cities with a limited budget and network coverage, provide self-sustaining mobility services for its residents while still preserving their privacy. We show the effectiveness of our approach and provide analysis using a case study from the Metropolitan Nashville area.