Randomized Local Search for Real-Life Inventory Routing

In this paper, a new practical solution approach based on randomized local search is presented for tackling a real-life inventory routing problem. Inventory routing refers to the optimization of transportation costs for the replenishment of customers' inventories: based on consumption forecasts, the vendor organizes delivery routes. Our model takes into account pickups, time windows, drivers' safety regulations, orders, and many other real-life constraints. This generalization of the vehicle-routing problem was often handled in two stages in the past: inventory first, routing second. On the contrary, a characteristic of our local search approach is the absence of decomposition, made possible by a fast volume assignment algorithm. Moreover, thanks to a large variety of randomized neighborhoods, a simple first-improvement descent is used instead of tuned, complex metaheuristics. The problem being solved every day with a rolling horizon, the short-term objective needs to be carefully designed to ensure long-term savings. To achieve this goal, we propose a new surrogate objective function for the short-term model, based on long-term lower bounds. An extensive computational study shows that our solution is effective, efficient, and robust, providing long-term savings exceeding 20% on average, compared to solutions built by expert planners or even a classical urgency-based constructive algorithm. Confirming the promised gains in operations, the resulting decision support system is progressively deployed worldwide.