Content area
In today’s industrial landscape, organizations are increasingly challenged by the growing complexity of operational decisions. From operational logistics to global supply chains, minimizing total costs while maintaining efficient service levels requires decision policies that are not only accurate but also adaptive to real-world uncertainty.
The Inventory-Routing Problem integrates inventory management and vehicle routing into a unified framework. When uncertainty and sequential decision-making are introduced, the problem evolves into the Dynamic and Stochastic Inventory Routing Problem (DSIRP). In this scenario, customer demands are progressively revealed over time and are leveraged to support informed and proper decisions. Each decision includes which customers to visit, how much to deliver, and how to combine the orders into a vehicle route.
Recent challenges related to scalability, computational complexity, and the need to adjust to real-time information realizations have fostered growing interest in applying Machine Learning to dynamic combinatorial optimization problems, including the DSIRP. In this context, this dissertation aims to explore the application of Supervised Learning, specifically to support the decision of which customers to visit in each decision point. The goal is to evaluate whether learning from past high-quality solutions can approximate expert behavior under uncertainty while reducing computational costs.
The proposed solution is based on a three-step decision policy. First, a trained Supervised Learning model selects which customers should be visited. Second, the delivery quantity is computed using two classical inventory management rules. Third, the vehicle route is optimized by solving a Traveling Salesman Problem for the selected customers. Models were trained in an iterative development process, using different feature sets refined over time. Their performance was evaluated using a custom simulation environment, which enabled cost-based comparisons against reference solutions computed under full knowledge of future demand.
Finally, the obtained results suggest that while the models consistently improved across iterations and outperformed a baseline from the literature, they still fell short of more complex benchmark policies. Nonetheless, Supervised Learning is a promising method for learning decision policies using limited computational resources, though further refinements are needed to close the performance gap to the best-performing strategies.