Enhancing Online Food Delivery with Transfer Points: A Data-Driven Decompose-Then-Optimize Approach

Online food delivery platforms can improve efficiency by consolidating orders with similar origins, destinations, and time windows at intermediate transfer locations. This research investigates the online food delivery problem with transfer (OFDP-T) and assesses how transfer-based routes and courier assignments enhance delivery performance. We propose a novel data-driven decompose-then-optimize framework that tames the exponential growth in route space from transfer and synchronization decisions and supports near-real-time decision making. The decomposition policy couples a first-step transfer-related routing and assignment subproblem with a tractable second-step generalized linear assignment problem. We develop a tailored hierarchical reinforcement learning (HRL) model to learn this policy and explicitly address learning barriers inherent to HRL and to the transfer-involved food delivery problem. Numerical experiments show that the model improves system net revenue by 13.3% over the best-known heuristics and by 26.9% over the nontransfer baseline. We further analyze the operational characteristics of transferred orders and the impact of courier heterogeneity on system performance. Applied to real-world Meituan data, the model yields a 16.1% increase in system revenue and an 11.48% reduction in the required courier fleet relative to optimized nontransfer operations. The framework thus provides a real-time solution by leveraging transfers and improving courier utilization, ultimately supporting a more sustainable and scalable food delivery system. Funding: This research was partially supported by research grants from the USDOT SMART project for Regional Mobility Engine, the U.S. Department of Energy [Grant DE-EE0011187], and the U.S. Department of Defense [Grant A9550-25-1-0278]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/trsc.2025.0147 .