Large Neighborhood Search and Structured Prediction for the Inventory Routing Problem

International audienceWe consider the packaging return logistics of a large scale car manufacturer, representing dozens of millions of euros and hundreds of thousands of tons of CO2 per year. This car manufacturer has a network of depots (factories) sending commodities (packaging) to customers (suppliers) by truck. This problem can be modelled as a multi-depot multi-commodity Inventory Routing Problem (IRP) [4]. The main challenges we face are the size of the car manufacturer’s network (with about 600 customers here compared with 50 maximum in the instance library of the literature [1]), the multi-depot aspect (leading to routing complexity), the multi-commodity aspect (with 30 concurrent commodities leading to binding constraints for the truck loading whereas most of the literature studies focus on the single-commodity framework with isolated cases of up to 5 commodities) and the large horizon (20 days compared with the 3 − 6 days often considered in the previous studies). Our instances are thus one order of magnitude larger than the ones addressed in the IRP literature. We emphasize that no algorithm is known to properly scale to our context