On July 24th, Àngel Ruiz defended his master thesis for the Master In Intelligent Systems. Its title is “Last mile routing with Graph Neural Network and Pointer Network: A comparison between global and zone-based training”, and it has been supervised by Dr. Carlos Granell and Dr. Sergi Trilles. The work has been produced with the support of a 2024 Leonardo Grant for Scientific Research and Cultural Creation, BBVA Foundation within the CoIoT project, which aims to generate geographically distributed repositories of ML models for them to be used by edge devices.

The thesis tackles the last-mile routing use case, trying to give a solution to it with an encoder-decoder architecture using a Graph Neural Network and a Pointer Network respectively and trained with REINFORCE loss. By clustering H3 hexagons, different geographical zones are created and a different instance of the model is trained for each zone. Finally, the work studies if ordering the stops of a delivery route inside a specific zone using the particular model for that zone outperforms an inference performed with a single model which predicts the entire route. It is verified that the zone-based training performs better than the general training, drastically reducing the errors.