The rise of e-commerce has transformed last-mile delivery, with companies’ prioritising faster, more flexible options and implementing innovations such as route optimisation. Efficient last-mile delivery is now critical to customer satisfaction and business success. This work aims to bridge the gap between planned and actual delivery routes, a challenge highlighted by the 2021 Amazon Last-Mile Routing Research Challenge. The solution uses a sophisticated hybrid approach, combining machine learning algorithms with automated hyperparameter optimisation. Instead of focusing on individual stops, it predicts sequences of zones. The process involves data pre-processing, a Prediction by a Partial Matching algorithm to identify optimal zone combinations, a Rollout Algorithm to compute zone sequences for unexplored routes, and a Lin-Kernighan-Helsgaun solver for zone-to-zone routing. These steps are seamlessly integrated into a repeatable pipeline that automates hyperparameter fine-tuning. The results obtained indicate a robust solution capable of producing high-quality predictions.

The authors acknowledge Fundação para a Ciência e a Tecnologia (FCT) for its financial support via the grant CEECINST/00096/2021 and the project UIDB/50022/2020 (LAETA Base Funding), and under project UIDB/00006/2020 project. DOI: 10.54499/UIDB/00006/2020

Last-mile delivery,VRP,PPM,Rollout Algorithm,Machine learning,Hybrid solutions

• Matemáticas - Ciências Naturais
• Ciências da Computação e da Informação - Ciências Naturais
• Engenharia Civil - Engenharia e Tecnologia
• Economia e Gestão - Ciências Sociais
• Outras Ciências Sociais - Ciências Sociais