Adjacency-Aware Deep Reinforcement Learning for Centralized Vehicle Repositioning

Efficient vehicle repositioning remains a key challenge in large-scale mobility-on-demand systems due to highly dynamic and uncertain supply-demand patterns. Multi-agent reinforcement learning (MARL) has been widely adopted for such problems; however, it often faces significant challenges including environmental non-stationarity, the curse of dimensionality, and various coordination-related difficulties. In this work, we propose a centralized repositioning framework that models vehicle repositioning as continuous actions based on adjacency relations over spatiotemporal grids. The policy is trained via Soft Actor-Critic (SAC), enhanced with frame stacking and spatiotemporal attention mechanisms to jointly capture temporal dynamics and spatial dependencies. We demonstrate that a properly designed centralized approach can effectively handle large-scale repositioning tasks, achieving performance comparable to or better than multi-agent methods. Comprehensive experiments on both synthetic and real-world datasets show that the proposed approach achieves efficient and robust repositioning while maintaining scalability and adaptability under varying fleet sizes and supply fluctuations.