Every day, countless drivers traverse the nation, delivering packages and parcels to customers and businesses. Many of these deliveries occur within only a few days, creating a substantial challenge for logistical coordination. Optimizing the last leg of delivery routes has long been a focus of operations research. This final phase is often the most expensive part of the supply chain, exacerbated by inefficiencies like lengthy distances between stops, weather issues, traffic congestion, limited parking, customer delivery preferences, and partially loaded trucks—inefficiencies that became particularly pronounced during the pandemic.
Advancements in technology and the availability of detailed data have empowered researchers to create models with enhanced routing options, all while balancing the computational demands involved. Matthias Winkenbach, MIT’s principal research scientist and director at the MIT Center for Transportation and Logistics (CTL), explains how artificial intelligence could yield more efficient solutions to intricate combinatorial optimization challenges like vehicle routing.
Q: What is the vehicle routing problem, and how do traditional operations research (OR) methods tackle it?
A: The vehicle routing problem is a common challenge faced by logistics and delivery services such as USPS, Amazon, UPS, FedEx, and DHL. At its core, it involves determining the most efficient route to connect multiple customers requiring deliveries or pickups. It’s about deciding which customers each vehicle should visit on a given day and in what sequence. The goal is often to find the shortest, quickest, or cheapest routes, but various constraints, such as specific customer needs or delivery time windows, complicate the process. For example, a delivery to a customer residing on the 15th floor of a high-rise presents unique challenges compared to one on the ground floor.
To effectively address the vehicle routing problem, implementing accurate demand information and customer characteristics is essential; knowing package sizes and weights and the quantity to be shipped significantly impacts service time at each stop. Additionally, it’s critical to identify where drivers can safely park. Traditionally, route planners relied on estimates, leading to blanket assumptions due to a lack of stop-specific data.
Machine learning plays a transformative role in improving this process. With most drivers equipped with smartphones or GPS trackers, a wealth of information is now available regarding delivery times. This data can be systematically extracted and used to accurately model each delivery stop.
Traditional OR methods involve formulating an optimization model, defining an objective function—typically a cost function—along with several equations that encompass routing problem dynamics. This includes ensuring that if a vehicle visits a customer, it must also leave, a concept known as flow conservation. Every customer must be visited exactly once during the route. While it appears straightforward, these constraints must be meticulously defined for a viable route.
Once the optimization problem is laid out, solvers help identify the best solutions. Over time, they refine their outputs to find efficient routes, aiming for a solution deemed satisfactory or, ideally, optimal. Given that an average U.S. delivery vehicle stops around 120 times, the computational intensity of solving these routes can be prohibitive. Consequently, companies often resort to heuristics—algorithms that quickly generate reasonable solutions without guaranteeing the optimal outcome.
Q: How are you integrating machine learning into the vehicle routing problem to outperform traditional OR methods?
A: We are actively exploring this with colleagues from the MIT-IBM Watson AI Lab. Our approach involves training a machine learning model on vast datasets of existing routing solutions—either observed from real-world operations or generated using efficient heuristics. Unlike traditional models that rely on an explicit objective function, these machine learning models require training to recognize the problem type and identify effective solutions. Much like training a language model on vocabulary, our route-learning model must comprehend the intricacies of various delivery stops and their demand characteristics. By understanding how to connect delivery stops efficiently, the model can create cost-effective or rapid routes even when faced with completely new customer demands.
In this approach, we utilize model architectures common in language processing. While it might seem unusual to apply language processing techniques to routing, the ability of transformer models to uncover structure—akin to how words combine to form sentences—makes them suitable for optimizing routes. In a city like Cambridge, there are about 40,000 potential delivery addresses; the challenge lies in effectively combining a smaller subset of these addresses into a well-organized route.
This innovative method aims to demonstrate that a purely machine learning-based approach can predict routes on par with, or even better than, those derived from traditional optimization techniques born out of decades in operations research.
Q: What advantages does your method hold over other cutting-edge OR techniques?
A: Currently, leading methods demand significant computational resources for model training, but this effort can be front-loaded. Once established, the trained model quickly generates solutions as required. Moreover, the dynamic nature of urban environments means factors like road conditions, traffic laws, and available parking constantly change. An OR-based approach could falter, necessitating a complete resolution of the routing problem with each new data point. In contrast, a well-trained model able to recall similar situations can suggest optimal routes swiftly, helping companies navigate an increasingly unpredictable landscape.
Additionally, optimization algorithms often rely on custom development to address specific company problems. The solutions’ quality is tied directly to the intricacy of the algorithm’s design. Meanwhile, a learning-based model continually refines its routing policy based on vast data exposure, making improvements without the need for continuous manual intervention.
Finally, traditional optimization methods generally focus on narrowly defined objectives—often purely cost-centric—while the realities companies face are multifaceted and sometimes conflicting. For example, a company might want efficient routing while minimizing emissions, and drivers prioritize safety and convenience. A robust route-learning model can adapt and account for these complex, high-dimensional objectives more flexibly than traditional methods.
This exciting evolution in machine learning holds tangible potential for real-world impacts, reshaping industries and benefiting society and the environment. While optimizing entire supply chains—like managing product flow from manufacturers in China through the global port network to retailers in North America—presents even more intricate challenges, our foundational work aims to spark future research and private sector innovations to enhance end-to-end supply chain optimization.
Photo credit & article inspired by: Massachusetts Institute of Technology
