Can language models be used for real-world urban-delivery route optimization?

Area of Science:

• Artificial Intelligence
• Operations Research
• Logistics and Supply Chain Management

Background:

• Language models (LMs) have advanced interdisciplinary research, including protein design and molecular dynamics.
• Existing delivery route optimization methods often fail to capture the implicit knowledge of complex operational environments.
• Human behaviors, when mappable to sequences, can be learned by LMs, offering new application avenues.

Purpose of the Study:

• To propose a novel approach using language models to optimize delivery routes based on drivers' historical experiences.
• To integrate implicit knowledge of delivery environments into route optimization by learning from experienced drivers' behaviors.
• To demonstrate the fusion of language models and operations research for enhanced delivery logistics.

Main Methods:

• Analogized real-world delivery routes to natural language sentences to leverage LM capabilities.
• Employed unsupervised learning to derive vector representations of driving behaviors and infer delivery chains.
• Developed a hybrid framework combining LMs for inter-zone delivery inference and the Traveling Salesman Problem (TSP) for intra-zone optimization.

Main Results:

• The proposed LM-based approach significantly outperformed pure optimization methods in numerical experiments.
• Real-world data from Amazon's delivery service validated the model's effectiveness, efficiency, and extensibility.
• The model successfully learned and preserved drivers' implicit behavioral patterns in optimized delivery routes.

Conclusions:

• Language models can effectively learn and optimize complex real-world sequential tasks beyond traditional language processing.
• The hybrid framework offers a versatile solution for delivery route optimization, integrating implicit human knowledge.
• This approach serves as a foundation for applying LMs to diverse interdisciplinary problems with sequential data.