Orateur
Description
The integration of inventory management and vehicle routing decisions creates a complex combinatorial optimization problem, known as the Inventory Routing Problem (IRP), which is a fundamental challenge in supply chain optimization and has been widely studied over the past decades. However, in the Stochastic IRP (SIRP), where retailer demand varies over time, the problem becomes more challenging. This paper focuses on the SIRP where the demand distribution is unknown in advance, stockouts are allowed during the planning horizon without back-ordering, and vehicle capacity is limited. To address this challenge, we propose an Enhanced Learning to Optimize (E-L2O) pipeline, jointly optimizing replenishment and routing decisions. The multi-stage SIRP is decomposed into a per-period solving process, leveraging the predictive and statistical capabilities of neural networks to incorporate future stochastic variables and related factors into current decisions. To enable more flexible replenishment strategies, we introduce an enhanced Fenchel-Young loss function to guide both replenishment and routing decisions. Through extensive computational experiments, we demonstrate that the resulting algorithm achieves solutions of unmatched quality to date, especially on large-scale benchmark instances.
