Distributed Fleet Control with Maximum Entropy Deep Reinforcement Learning

Takuma Oda, Yulia Tachibana

Abstract: In the context of modern vehicle fleets, such as ride-hailing platforms and taxi companies, the ability to proactively dispatch vehicles is instrumental in reducing passenger waiting time and unoccupied cruising time, improving driver profit and decreasing environmental and traffic impact. Among the complex dynamics of fluctuating demand, supply, and traffic conditions inherent in vehicle dispatch, which are almost impossible to fully model explicitly, model-free approaches have shown marked strengths. We present a framework which extends the model-free DQN formulation with soft-Q learning entropy maximization and graph-based diffusion convolution. Comparison against a non-diffusive ’hard’ formulation shows significant improvement across several metrics, such as passenger waiting times.