To meet personalized travel demand, ride-sourcing platforms provide differentiated service modes for travelers. These service modes are usually operated independently, and because of the heterogeneity of travel demand, the fragmented ride-sourcing services struggle with imbalances between supply and demand within each sub-market. For that, cross-mode dispatching can be adopted to reduce the vehicle idle time and improve passengers’ experience. However, it may undermine the efficiency of the premium mode, such as increased matching failures. To consider the long-term impact of cross-mode dispatching, the dispatching problem of multi-service modes is modeled by the Markov decision process. A multi-service mode dispatching framework is proposed, comprised of a simulator and a two-stage reinforcement learning algorithm. The simulator provides environment support while avoiding matching failures of premium mode caused by cross-mode dispatching. The two-stage reinforcement learning algorithm refines requests based on the k-nearest neighbor algorithm to solve the problem of the dynamic change of the candidate request set. Experiments show that the framework fulfills 3.1% additional basic requests without the decreasing of fulfilled premium requests, by cross-mode dispatching, demonstrating the feasibility of the cross-mode dispatching and the performance of the framework. Sensitivity analyses are designed to test the robustness of the framework under different scenarios.
