End-to-end logistics in metropolitan areas: A stochastic dynamic order-assignment and dispatching problem
针对都市区端到端物流平台，提出随机动态订单分配与调度问题，用马尔可夫决策过程建模，结合参数化成本函数近似和自适应大邻域搜索求解，实验显示平均成本降低22%，案例研究可降30.5%。
The growth of e-commerce requires efficient integration of first-mile pickup, middle-mile consolidation, and last-mile delivery. These so-called integrated end-to-end logistics operations are particularly visible in metropolitan areas where fast delivery services are in high demand. Inspired by real-world practices at our industry partner, this paper introduces the Stochastic Dynamic Order-Assignment and Dispatching Problem (SDOA-DP). It concerns stochastic and dynamic pickup-and-delivery orders arising at an end-to-end logistics delivery platform, for which the company, as a decision maker , needs to determine in real-time how to assign orders to middle-mile linehaul schedules and when to dispatch first- and last-mile two-echelon vehicle routes. We model the SDOA-DP as a Markov Decision Process and propose a novel solution approach based on a parameterized Cost Function Approximation (CFA) for order assignment in the middle mile and a parameterized Adaptive Large Neighborhood Search (ALNS) for vehicle dispatch and two-echelon routing in the first and last-mile. The CFA balances the cost of using linehauls with the time slack available for first- and last-mile planning while ensuring time windows are met. The parameterization in the ALNS ensures that we balance routing cost and delivery speed by limiting the frequency and timing of dispatching vehicle routes. We learn the best value of the parameterization using Bayesian optimization . Computational experiments show that our approach yields a 22% on-average improvement compared to a baseline policy. If we learn a single best parameterization for various system settings, we observe almost as good cost savings, showing that our approach is robust and reliable for practitioners. Finally, we applied our method to a case study of our industry partner and showed that our method could potentially reduce daily costs by 30.5% across various operational contexts.