Optimizing Cluster-Based Vehicle Dispatching in Ride-Hailing Services: A Reinforcement Learning Approach to Demand-Supply Dynamics

This paper introduces a Reinforcement Learning (RL) framework that optimizes vehicle dispatching in urban ride-hailing systems. In this approach, we use a single-agent framework built upon Double Deep Q-Networks (DDQN) to maintain steady decision-making. Rather than running multiple agents in parallel, the agent sequentially handles each city-grid cluster, effectively mirroring Multi-Agent Reinforcement Learning (MARL) coordination without the accompanying computational overhead. By considering both local and neighboring demand supply conditions for each dispatch decision, the model achieves flexible and effective fleet management across clusters. Validation with New York City taxi data reveals enhanced service efficiency, low rejection rates, and fleet balance, demonstrating the scalability and effectiveness of this data-driven approach for high-demand urban environments. The results highlight the model’s strong performance in minimizing delays and ensuring optimal dispatching decisions.