Dynamic Fleet Management With Rewriting Deep Reinforcement Learning

Inefficient supply-demand matching makes the fleet management a research hotpot in ride-sharing platforms. With the booming of mobile network services, it is promising to abate the supply-demand gap with effective vehicle dispatching. In this article, we propose a QRewriter - Dueling Deep Q-Network (QRewriter-DDQN) algorithm, to dispatch multiple available vehicles in ahead to the locations with high demand to serve more orders. The QRewriter-DDQN algorithm factorizes into a Dueling Deep Q-Network (DDQN) module and a QRewriter module, which are parameterized by neural networks and Q-table with Reinforcement Learning (RL) methods, respectively. Particularly, DDQN module utilizes the Kullback-Leibler (KL) distribution distance between supply (available vehicles) and demand (orders) as excitation to capture the complex dynamic variations of supply-demand. Afterwards, the QRewriter module learns to improve the DDQN dispatching policy with the streamlined and effective Q-table in RL. Importantly, the higher performance improvement space of the DDQN dispatching policy can be obtained by aggregating QRewriter state into low-dimension meta state. A simulator is designed to train and test the performance of QRewriter-DDQN, the experiment results show the significant improvement of QRewriter-DDQN in terms of order response rate.