An Order Dispatch System Based on Reinforcement Learning for Ride Sharing Services

Ride-sharing has been widely used in many cities, such as Didi and Uber. Ride-sharing is regarded as an effective way to solve urban traffic congestion and pollution. However, most of the existing dispatch methods take the minimization of the travel distance as the optimization goal, without considering other factors. In this paper, we consider not only detour distance, but also consider seat utilization, future profit, and hidden profit. Firstly, we propose a deep evaluation network to evaluate factors that affect vehicle dispatch, and we exploit the reinforcement learning strategy to train the deep evaluation network. Then, we propose the dynamic external factor calculation (DEFC) algorithm to calculate those factors. Secondly, to calculate the driver's future profit, we propose a prediction model based on the K-Nearest Neighbor(KNN) to predict the number of passengers and vehicles. Thirdly, we use a vehicle search method to search vehicles that satisfy passenger spatial-temporal constraint. Based on a real-world dataset in Rome, we evaluate our algorithm to confirm the effectiveness of our method.