Are you waiting for a package to be delivered? Complex math problems need to be solved before the van arrives at your door, and MIT researchers have a strategy that can expedite their resolution.

This approach applies to vehicle route planning problems such as last mile delivery where the goal is to deliver goods from the Central Depot to multiple cities while reducing travel costs. While there are mechanisms that are designed to solve this problem for several hundred cities, this solution is much slower in larger cities.

Gilbert W. Snyder from the Institute for Data, Systems and Society for Civil and Environmental Engineering will speak about this. Kathy Woo, Winslow’s assistant professor of career development, and her students developed machine learning techniques that accelerate some of the most powerful techniques. Algorithm solution 10 to 100 times.

To solve another 200 problems with running vehicles in 2,000 cities, the solution mechanism works by breaking the delivery problem into smaller incremental problems. Wu and colleagues have refined this process with a new machine learning algorithm that identifies the most effective sub-problems to be solved rather than solving all sub-problems, thereby improving the quality of the solution when a lower order computation is used.

The researchers say their approach, which they call “learn-to-represent,” can be applied to a variety of solutions and similar problems, including planning and routing for warehouse robots.

Founder and CEO Mark Guo said the mission is to push boundaries to quickly resolve major vehicle routing problems. Route plan, Smart Logistics Site to Improve Shipping Routes. Some of Rudyf’s latest algorithm improvements are inspired by Woo’s work, he notes.

“Most academics focus on specific mechanisms for small problems, trying to find the best solution at the expense of turnaround time. But in the real world, companies aren’t interested in finding the best solution, especially if it takes too long to calculate, ”explains Guo Money, and your entire warehouse operation can’t wait for the slow transfer process. An algorithm must be able to be implemented very quickly. “

Wu, Sirui Li, PhD student in Social Organization and Engineering, and Zhongxia Yan, PhD student in Electrical Engineering and Computer Science. Your research This week at the NeuroIPS conference 2021.

Pick a good problem

Vehicle navigation problems are a built-in class of problems that use a horizontal algorithm to find a “good enough solution” to a problem. “Better” answers to these problems cannot usually be found, as the number of possible solutions is overwhelming.

“The name of the game for these types of problems is to develop efficient algorithms … they are optimized within certain factors,” explains Wu. “But the goal is not to find the optimal solution. It is very difficult. Instead, we want to find the best possible solution. Even a 0.5% increase in solutions can result in a huge increase in sales for a company.

Over the past few decades, researchers have developed various heuristics to provide quick solutions to built-in problems. They usually do this by starting with a bad but correct initial solution and then gradually increasing the solution – for example by making small changes to improve the routes between nearby cities. However, for larger problems such as city routing challenges of over 2,000, this approach will take longer.

Machine learning methods have been developed recently to solve problems, but if they are fast they can be very accurate even at the level of a few dozen cities. Wu and his colleagues decided whether there was an effective way to combine the two methods to find a faster, but higher quality, solution.

“This is where machine learning comes in for us,” said Wu. “Which of these sub-problems can we predict if we want to solve them, which leads to further improvements to the solution and saves computing time and costs?”

Conventionally, large vehicle routing problem heuristics can be selected in roughly any order or with sub-problems using other carefully designed horistics. In this case, the MIT researchers automatically recognize an additional series of problems via the neural network they have created, the solution of which leads to an enormous increase in the quality of the solution. According to Wu and colleagues, this process speeds up the sub-problem selection process by 1.5 to 2 times.

“We don’t know why this sub-problem is better than the other sub-problems,” noted Wu. “This is really an exciting future area of ​​responsibility. If we have some insight, it could lead to even better algorithm designs. “

Amazing speed

Woo and his colleagues were amazed at how well this approach worked. In machine learning, the idea of ​​trash in, garbage out applies – that is, the quality of the machine learning approach depends heavily on the quality of the data. Collective problems are so difficult that even the additional problems cannot be optimally solved. Input data Neural networks that have been trained in the “medium quality” sub-problem solution “usually give results of moderate quality,” says Woo. In this case, however, the researchers were able to use a medium-quality solution, which is much faster than advanced methods, to produce high-quality results.

For vehicle navigation and similar problems, users need to develop highly specialized algorithms to solve their specific problems. Some of these horizontals have been developed over the decades.

The Learning-to-Delicate-System offers an automated way to speed up this horoscope for big problems.

Because this method can work with a variety of solutions, it can be useful for a variety of resource allocation problems. “Since the cost of troubleshooting is 10 to 100 times lower, we can now unlock potential new apps.”

This research was supported by the MIT Indonesia Seed Fund, the US Department of Transportation’s Dwight David Eisenhower Transportation Scholarship Program, and the MIT-IBM Watson AI Lab.