DROP: Deep relocating option policy for optimal ride-hailing vehicle repositioning

In a ride-hailing system, an optimal relocation of vacant vehicles can significantly reduce fleet idling time and balance the supply–demand distribution, enhancing system efficiency promoting driver satisfaction retention. Model-free deep reinforcement learning (DRL) has been shown to dynamically learn relocating policy by actively interacting with intrinsic dynamics in large-scale systems. However, issues sparse reward signals unbalanced demand supply distribution place critical barriers developing effective DRL models. Conventional exploration strategy (e.g., ϵ-greedy) may barely work under such environment because dithering low-demand regions distant from high-revenue regions. This study proposes option (DROP) that supervises vehicle agents escape oversupply areas effectively relocate potentially underserved areas. We propose Laplacian embedding time-expanded graph, as approximation representation policy. The generates task-agnostic signals, which combination task-dependent constitute pseudo-reward function for generating DROPs. present hierarchical framework trains high-level set low-level note DROP is general method be incorporated into existing model-free RL advances further improvements applications. effectiveness our approach demonstrated using custom-built high-fidelity simulator real-world trip record data. report improves baseline value iteration algorithms resolve issue