A Distributed Spatio-Temporal Deep Q-Network for On-Demand Ridesharing

Ridesharing holds significant potential to transform urban mobility by offering passengers affordable, shared transportation while increasing profits for transit operators. However, existing approaches often struggle to accurately identify and respond to zones of high demand, leading to suboptimal dispatch decisions and reduced efficiency. To address this challenge, this paper proposes a novel spatio-temporal Deep Q-Network (DQN) architecture designed to optimize real-time vehicle dispatching in ridesharing systems. The proposed model utilizes a diffusion convolutional layer to capture complex spatial dependencies in demand patterns, alongside a recurrent layer to learn temporal demand dynamics. These components are jointly integrated within a deep reinforcement learning framework to determine optimal dispatch actions. We evaluate our approach through extensive experiments using the New York City taxicab public dataset. Results demonstrate that our method significantly improves key operational metrics including profit, idle time, and vehicle utilization, highlighting its effectiveness in enhancing ridesharing system performance.