Completed Project

Independent Research Project

Interpretable Learning-Augmented Last-Mile Delivery Routing

An interpretable, learning-augmented approach to last-mile delivery routing that combines operations research with interpretable machine learning and hierarchical zone-preference modelling, evaluated on real Amazon last-mile delivery routes using the official challenge scoring.

Dates
November 2025 – December 2025

Research Problem

Last-mile delivery routes learned purely from historical data can be accurate but opaque. This project develops an interpretable, learning-augmented routing approach that combines operations research with interpretable machine learning and hierarchical zone-preference modelling, evaluated against real delivery behaviour using the official challenge scoring.

Methodology

Operations ResearchInterpretable Machine LearningRoute OptimisationHierarchical Zone-Preference ModellingChronological Train–Validation–Test SplittingOfficial Challenge ScoringSensitivity AnalysisRoute-Level Robustness Analysis

Tools

Python

Verified Data

  • Analysed 6,112 Amazon last-mile delivery routes
  • Evaluated on 926 unseen test routes

Verified Results

  • Reduced aggregate median official route score by 39.9% compared with nearest-neighbour routing
  • Improved 88.1% of test routes
  • Limited median travel-time increase to 4.83%
  • Reduced median zone re-entries from 21 to 0

Current Stage

Completed independent research project