Using machine learning to identify hidden constraints in vehicle routing problems

Abstract

Last-mile delivery involves a series of complex tasks in an unpredictable environment. Decision support tools based on optimization algorithms construct efficient routes for drivers, optimizing the cost of making deliveries. However, drivers often deviate from these routes due to factors not considered in the decision-making process. This discrepancy raises the question of how to identify routes that are useable in real-world scenarios. Our research proposes using modern machine learning techniques to classify routes based on their practical usability. In a controlled environment, we demonstrate that machine learning can learn hidden factors influencing route viability by focusing on variants of the vehicle routing problem with additional constraints like time window, capacity and precedence. For each underlying constraint, we show that a machine learning model can be trained to classify routes based on whether or not they violate the constraint. Using datasets generated from well-known benchmark instances, we present computational experiments to evaluate model performance. We discuss which types of constraints are more challenging to recognize and how large a dataset must be to allow for accurate classification. This research has the potential to improve existing decision tools, enabling them to generate routes that better account for real-world complexities.