Graph Prompt Learning Method for the Demand-Responsive Transport Routing Problem

Demand Responsive Transport (DRT) plays a crucial role in mitigating the inefficiencies of current public transit systems. Efficient routing is paramount for enhancing the flexibility and applicability of this transportation mode. Machine learning techniques, such as the attention-based encoder-decoder methodology, have the capability to produce solutions within seconds after offline training. However, these algorithms encounter convergence issues during training process, and demonstrate limited generalization ability, particularly across different scales. Thus, this paper proposes a graph prompt learning-based method comprising an information encoder, token generation, and token mapping to effectively train models that can adapt to diverse vehicles and demand variations. Particularly, token generation considers the characteristics of the problem by integrating vehicle and customer urgency information each time step. Token mapping obtains vehicle decoding sequences through attention mechanisms and mask function. The proposed model's performance is comprehensively evaluated against commonly baselines across various request contexts. Results show that our method can significantly reduce the computational time, and improve the quality of routing solution compared with baselines. Overall, the proposed model can enhance the routing efficiency of DRT systems through token mapping and prompts design.