Machine Learning Models for Route Consolidation

This report was originally published in April 2020

Abstract

Finding common routes from a large set of individual trips is a difficult problem due to the natural complexity involved with nontrivial trips. Much of the current research for route consolidation has relied on clustering- or distance-based methods, along with ad-hoc rules for combining routes. We compare a more traditional physical-based method that uses clustering, graph theory and ad-hoc rules with a machine-learning method. In particular, the machine-learning method uses an autoencoder to reduce the number of trips in the dataset and find common or standard routes. The routes identified using the autoencoder are then post-processed using K-means clustering and Procrustes distance. We apply both methods to a mine haul truck trip dataset and show how the machine-learning method can largely replicate the results produced by the physical modeling method, thus providing a more generalizable alternative for route consolidation.