Learning the on-demand adaptable matching range with a reinforcement learning

Abstract

Ride-sourcing services have reshaped urban transportation, providing greater convenience and efficiency for city commuters. At the core of these services is the matching process, which directly impacts service efficiency, passenger satisfaction, and overall platform profitability. Consequently, developing highly effective matching algorithms, especially under imbalanced supply–demand conditions, is of utmost importance. In existing matching algorithms, the matching range is a key factor. A larger matching range can result in longer pickup waiting times, potentially leading passengers to abandon their requests. Conversely, a smaller matching range may shorten waiting times but can also reduce the overall matching rate. Previous research on optimizing the matching range has often overlooked future information, leading to short-term improvements. In this paper, we propose a generalized, on-demand adaptable matching range technique based on reinforcement learning framework, designed to optimize decision-making from a long-term perspective while accounting for future information. Additionally, we develop a flexible framework adaptable to different kinds of matching modes. To evaluate the effectiveness of our approach, we implement our strategy with real-world supply and demand data and conduct a series of sensitivity analyses. The experimental results demonstrate that our method can achieve improvements in terms of the platform’s revenue and passengers’ satisfaction simultaneously compared with benchmark algorithms.