Optimization algorithms for adaptative route sequencing on real-world last-mile deliveries

Abstract

This article explores the design and application of machine learning techniques to enhance traditional approaches for solving NP-hard optimization prob-lems. Specifically, it focuses on the Last-Mile Routing Research Challenge (LMRRC), supported by Amazon and MIT, which sought innovative solutions for cargo routing optimization. While the challenge provided travel times and zone identifiers, the dependency on these factors raises concerns about the algorithms’ generalizability to different contexts and regions with standard delivery services registries. To address these concerns, this study proposes personalized cost matrices that incorporate both distance and time models, along with the relationships between delivery stops. Additionally, it presents an improved approach to sequencing stops by combining exact and approxi-mate algorithms, utilizing a customized regression technique alongside fine-tuned metaheuristics and heuristics refinements. The resulting methodology achieves competitive scores on the LMRRC validation dataset, which comprises routes from the USA. By carefully delineating route characteristics, the study enables the selection of specific technique combinations for each route, considering its geometrical and geographical attributes. Furthermore, the proposed methodologies are successfully applied to real-case sce-narios of last-mile deliveries in Montevideo (Uruguay), demonstrating similar average scores and accuracy on new testing routes. This research contributes to the advancement of last-mile delivery optimization by leveraging personalized cost matrices and algorithmic refinements. The findings highlight the potential for improving existing approaches and their adaptability to diverse geographic contexts, paving the way for more efficient and effective delivery services in the future.