Across the U.S., hundreds of thousands of drivers deliver innumerable parcels daily, with most deliveries taking a few days. Coordinating such a enormous supply chain in a predictable and timely manner is a challenging problem in operations research, particularly optimizing the last leg of delivery routes, which is often the costliest due to factors such as long distances between stops, weather delays, lack of parking availability, and customer delivery preferences. These inefficiencies became more apparent during the COVID-19 pandemic.
Improvements in technology and the availability of more detailed data has allowed researchers to develop models with better routing options, but balancing the computational cost is crucial. Matthias Winkenbach, research scientist at MIT, discusses how artificial intelligence could provide more efficient solutions to this complex optimization problem.
The vehicle routing problem is faced by logistics and delivery companies like USPS, Amazon, UPS, FedEx, and DHL daily. The task is finding an efficient route connecting a set of customers either for delivery or pick-up. To solve this problem, accurate demand information and customer-related characteristics are needed such as the size or quantity of packages. Traditionally, route planners would make estimates for these parameters, resulting in blanket assumptions due to lack of specific data.
Machine learning can offer solutions as drivers have smartphones or GPS trackers that provide information on how long it takes to deliver a package. With machine learning, this information can be extracted and used to model every single stop realistically.
Traditional operations research approach involves writing up an optimization model and defining the objective function. Then, algorithms help find the best possible solutions that meet all the constraints. This is a computationally expensive process, so companies often use efficient heuristics to find reasonably good solutions.
Researchers at the MIT-IBM Watson AI Lab are applying machine learning to the vehicle routing problem. The machine-learning models are trained on a large set of existing routing solutions collected from real-world operations. These models understand the nature of the problem and what a good solution looks like, and can connect delivery stops to find efficient routes, drawing on language processing techniques.
The method still has its limitations, being computationally expensive, however, its benefits far outweigh them. It adapts better to a constantly changing operational environment compared to traditional OR methods, produces instant solutions, and learns from the data over time. It can also handle complex, multi-dimensional objectives more effectively than traditional optimization approaches.
A learning-based routing tool can have a tangible real-world impact in the logistics industry, on society, and the environment. It could lay the foundation for more sophisticated research in optimizing and improving end-to-end supply chain processes.
