Generating large-scale real-world vehicle routing dataset with novel spatial data extraction tool.

Area of Science:

• Computer Science
• Artificial Intelligence
• Operations Research

Background:

• Deep reinforcement learning (DRL) shows promise for vehicle routing optimization.
• Practical DRL application is limited by a lack of realistic, real-world compatible datasets.
• Existing datasets often use simplistic metrics, failing to capture urban complexities.

Purpose of the Study:

• To address the scarcity of real-world data for DRL in vehicle routing.
• To develop a novel methodology for generating realistic vehicle routing datasets.
• To facilitate the practical implementation of DRL in complex routing scenarios.

Main Methods:

• Developed a spatial data extraction and curation tool for geocoded urban locations.
• Refined location data considering unique urban environmental characteristics.
• Integrated specialized distance metrics and demand information to create vehicle routing graphs.

Main Results:

• Successfully generated datasets closely aligned with DRL model requirements.
• Demonstrated efficacy on varied real-world testbeds.
• Dataset structure facilitates efficient access and manipulation (independent graphs with location, distance, demand).

Conclusions:

• The proposed methodology effectively generates real-world compatible data for DRL-based vehicle routing.
• This approach enhances the adaptability and reliability of DRL in complex routing challenges.
• Significant progress towards practical DRL application in real-world vehicle routing problems.