Towards smarter on-demand ride-sharing: Leveraging spatiotemporal flexibility to improve efficiency via delayed matching and walking

Ride-sharing services have become a key component of urban mobility systems by improving transport efficiency and supporting more effective use of vehicle resources. However, conventional immediate matching strategies often fail to fully exploit temporal and spatial flexibility in user behavior. This study explores an on-demand management strategy that delays matching decisions for selected requests, while allowing passengers to walk to nearby pick-up points. Such flexibility offers the potential to reshape the temporal distribution of demand and better align it with vehicle availability. A two-stage on-demand delayed matching framework is proposed that integrates a multi-agent reinforcement learning-based admission mechanism with a trip-to-vehicle graph-based matching model incorporating walking flexibility, which aims to enhance matching efficiency while reducing travel time costs. Experimental results based on real-world ride-sourcing travel records from central Beijing show that the proposed strategy enhances overall system performance by increasing matching rates, reducing per-order emissions, and boosting platform profitability. For passengers, spatiotemporal flexibility in matching not only lowers pick-up, waiting, and walking times but also slightly reduces average matching time, showing that operational gains can be achieved without sacrificing user experience. The results suggest that modest behavioral adjustments, such as short-distance walking, can ease service imbalances without requiring major infrastructure changes. Ride-sharing platforms can replace uniform immediacy with response times that reflect passengers’ temporal flexibility, while adopting differentiated strategies that account for heterogeneous walking willingness can enhance system feasibility while maintaining user acceptance.