CHEN Jing, ZONG Chang-ming, YUAN Chang-wei, MA Ning-yuan, ZHANG Jie, ZHANG Ming-hui, FANG Xin-jie
To address the limitations of existing taxi dispatching methods, which often focus on optimizing the interests of only one or two stakeholders among platforms, passengers, and drivers, and to overcome the low solution efficiency of traditional methods in dynamic environments, this study proposes a multi-objective taxi dispatching optimization method based on deep reinforcement learning. First, the taxi order dispatching problem is formulated as a Markov decision process. The dispatching platform is modeled as the agent, and prioritizing orders from a selected grid cell is defined as the action of the platform agent, which improves the efficiency of order dispatching. Second, a multi-objective reward function is designed by integrating platform efficiency, passenger waiting time, driver income, and drivers' preferred operating areas, so that the interests of multiple stakeholders can be considered simultaneously. Finally, a Deep Q Network is used to learn the order dispatching policy, enabling efficient decision making and long term reward optimization in dynamic environments. The proposed method improves order response capability and enhances passengers' waiting experience, while also better accounting for drivers' preferred operating areas and achieving coordinated optimization among multiple stakeholders.Experiments on a Xi'an taxi order dataset during the morning peak period over a complete week show that the proposed method outperforms the average performance of four baseline methods. Specifically, the order response rate and total revenue increase by approximately 13.42% and 10.37%, respectively, while the 95th percentile response time and average pickup time decrease by approximately 13.17% and 36.17%, respectively. In addition, the proposed method maintains a high order response rate with low decision time, indicating its practical value in large scale, dynamic, and real time order dispatching scenarios. Ablation experiments and multi-objective emphasis experiments further show that passenger waiting time and drivers' preferred operating areas are major factors affecting order dispatching performance. A balanced weight setting can provide a more reasonable trade off among platform efficiency, passenger experience, driver income, and drivers' preferred operating areas. The findings provide a reference for precise taxi fleet allocation and the optimization of urban transport supply and demand balance.
