Integrating reinforcement learning-based vehicle dispatch algorithm into agent-based autonomous taxi fleet system simulation

On-demand mobility has drastically changed the way transportation systems are operated and greatly improved people’s access to transportation services. Meanwhile, autonomous driving technology has matured over time, and driverless vehicles have already been operated in real life. Replacing traditional taxi fleet with reliable autonomous taxi fleet would improve the service quality of transportation systems even further. However, with a large fleet of fully controllable objects, the operational optimization of such system becomes challenging as well. Existing studies fail to address both the realisticness of system simulation and the advantage of optimization-based algorithms at the same time. To precisely measure the benefits of operating an AV taxi fleet, this thesis integrates a reinforcement learning algorithm into to an agent-based simulation model of a ride hailing system. A real-world scale simulation of New York City (NYC) taxi fleet is conducted, and the system performance with the algorithm is compared with the common rule-based and heuristic dispatch algorithms in relevant literatures. It was observed that (1) DQN dispatched vehicles conservatively but achieved similar rider service level with proactive dispatch methods; and (2) DQN outperformed all other dispatch methods evaluated in this study with significantly higher dispatch efficiency.