FedCruise: Collaborative Cruise Guidance With Federated Policy Distillation in Multiple Ride-Hailing Platforms

Recent technological advancements have led to the emergence of intelligent cruise guidance systems tailored for ride-hailing platforms (RHPs), such as Uber and Didi Chuxing, offering potential solutions to issues like traffic congestion and vehicle emissions. However, they face challenges, such as passenger-driver matching, route, and price optimization, and ensuring safety and fairness. These challenges are exacerbated by heterogeneity in data across multiple RHPs and privacy concerns related to data sharing. In this article, we propose a novel cruise guidance framework, FedCruise, which tackles these issues by using customized federated policy distillation with deep reinforcement learning (DRL). FedCruise enables collaborative model training across different RHPs without exchanging raw data, preserving privacy while addressing nonidentically distributed (non-IID) data challenges. FedCruise employs two models in each DRL agent: 1) a local teacher model and 2) a global student model, enabling bidirectional learning and achieving a global optimum. Our framework optimizes ride-sharing services and addresses data heterogeneity and privacy challenges. The results of our extensive simulations demonstrate the effectiveness and efficiency of FedCruise, proving its superiority in convergence rate, pickup orders, and driver income over other benchmarks.