From the shipping of online shopping goods to the delivery of postal mail, optimised transport routes save time and resources for companies and customers alike. Today, artificial intelligence (AI) models can help human coordinators plan delivery routes that cover multiple locations as efficiently as possible. However, such vehicle routing problems (VRPs)—which fall under a bigger umbrella of combinatorial optimisation problems (COPs)—are often time-consuming and computationally taxing to solve.
“COPs involve finding the best solution from a finite but extremely large set of possibilities,” explained Zhuoyi Lin and J. Senthilnath, respectively a Scientist and Senior Scientist from the A*STAR Institute for Infocomm Research (A*STAR I2R). “These problems are crucial as they also occur in manufacturing and even biology; think of production schedules and protein folding.”
In collaboration with a global team from Singapore Management University; Eindhoven University of Technology, the Netherlands; and Shandong University, China, Lin and Senthilnath investigated whether VRPs could be more speedily solved through a cross-problem learning approach based on neural heuristics.
“Heuristics are ‘rules of thumb’ that help to quickly find good solutions without searching every possibility,” said Lin and Senthilnath. “Neural heuristics are specifically those learned by neural networks—a type of AI model—instead of being manually designed.”
Inspired by how the human brain extracts patterns from data, neural networks can use training datasets to adapt and improve on their own rather than rely on hard and fixed rules encoded by developers. Much like how one might learn to solve a puzzle, the team wanted their model to not just identify the best VRP solutions, but—more importantly—to pick up on the logic behind efficient route planning, then apply these rules to solving other types of COPs.
The team first pre-trained a neural network on a basic routing problem, then explored different fine-tuning strategies for more complicated routing problems. “We found that adapter-based fine-tuning achieved nearly the same performance as full training, but with far fewer parameters, making the model lighter and more efficient,” said the researchers. “One unexpected but exciting result was achieving over 90 percent parameter efficiency, where fine-tuning the model on just a single task significantly improved performance across other complex routing problems. This dramatically reduces computational cost and training time, making the approach highly scalable and efficient.”
Based on the proposed cross-problem learning approach, the team’s model learnt how to navigate a theoretical delivery network and decide on the best routes, effectively generalising its heuristics across different VRPs.
“Knowledge learnt from solving one type of routing challenge can be reused for other problems,” said Lin and Senthilnath. “Our approach reduces the need to manually build a new heuristic algorithm from scratch for each specific task. This can lead to faster, more adaptive logistics and transport solutions.”
Looking ahead, the researchers aim to extend their approach to a broader spectrum of real-world optimisation challenges, such as job scheduling and bin packing. They also envision integrating the reasoning capabilities of large language models, which could enable real-time, adaptive decision-making in complex logistical environments.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Institute for Infocomm Research (A*STAR I2R).
