When some suburban trains arrive at the end of the line, they have to go to the switching platform to turn around so that they can leave the station later, often from a different platform than the one they arrived.
Engineers operate programs called algorithmic Solvers to plan these moves, but at a station with thousands of weekly newcomers and departures the problem becomes too convoluted for classic Solver can discover everything at the same time.
Using machine learning, MIT scientists have developed an improved planning system that reduces the time of solving by up to 50 percent and gives a solution that better meets the user’s goal, such as the departure of trains on time. The modern method can also be used to effectively solve other convoluted logistics problems, such as hospital staff planning, assignment of airline crews or assignment of tasks to factory machines.
Engineers often distribute this type of problems into the sequence of overpracks that can be resolved at feasible time. But overlapping means that many decisions are unnecessarily calculated, so Solver occupies achieving the optimal solution.
The modern, from artificial intelligence, enhanced by intelligence, learns which parts of each sub -compulsory should remain unchanged, freezing those variables to avoid excess calculations. Then classic algorithmic solver solves the remaining variables.
“A often dedicated team can spend months and even years in designing the algorithm to solve only one of these combinatorial problems. Modern deep learning gives us the opportunity to use new progress to improve the design of these algorithms. We can accept what we know, works well, and use AI to accelerate it,” says Cathy Wu, Thomas D. and Virginia W. Cabot Association Association in civil and civil engineer (CEE) and CEE) and CEE) and CEE) and CEE) and CEE) Institute of Data, Systems and Society (IDSS) in MIT and a member of the Laboratory of Information and Decision Systems (LIDS).
He joins her paper by the main author of Sirui Li, an IDSS graduate; Wenbin Ouyang, CEE graduate; and yining has, postdoc lids. The research will be presented at an international conference on the representation of learning.
Eliminating redundancy
One motivation for these research is a practical problem identified by a master’s student Devin Camille Wilkins at Wu’s basic transport course. The student wanted to operate learning to strengthen in a real problem with the Boston North Station train discount. The transit organization must assign many trains to a circumscribed number of platforms on which they can be turned long before arriving at the station.
It turns out that this is a very convoluted combinatorial problem – the exact type of problem of the Wu laboratory has spent the last few years.
In the face of a long -term problem, which includes assigning a circumscribed set of resources, such as factory tasks, a group of machines, planners are often a problem as pliant workshop planning.
In pliant workshop planning, each task needs a different time to perform, but the tasks can be assigned to any machine. At the same time, each task consists of operations that should be performed in the right order.
Such problems quickly become too huge and bulky for classic solvers, so that users can operate the optimization of the horizon (RHO) to divide the problem into fragments that can be managed that can be solved faster.
Thanks to RHO, the user assigns some initial tasks to the machines in a set planning horizon, perhaps a four -hour time window. Then they perform the first task in this sequence and change the four -hour front planning horizon to add the next task, repeating the process until the whole problem is solved and creating the final schedule of assigning tasks.
The planning horizon should be longer than any duration of one task, because the solution will be better if the algorithm also considers the tasks that appear.
But when the planning horizon develops, it causes some overlapping to activities in the previous planning horizon. The algorithm has already invented preliminary solutions to these overlapping operations.
“Perhaps these preliminary solutions are good and do not have to be calculated again, but maybe they are not good. Machine learning here,” explains Wu.
In the case of their technique, which they call the optimization of the science horizon (L-RHO), scientists teach the machine learning model to predict which operations or variables should be re-calculated when the planning horizon will develop.
L-RHO requires data to train the model, so scientists solve a set of sub-composition using a classic algorithmic solver. They took the best solutions – those of the most operations that do not have to be re -calculated – and used them as training data.
After training, the machine learning model receives a modern sub -department, which he has not seen before, and predicts which operations should not be calculated again. Other operations are transferred back to the algorithmic Solver, which performs the task, again calculates these operations and moves the horizon of forward planning. Then the loop begins again.
“If, in retrospect, we did not have to optimize them again, we can remove these variables from the problem. Because these problems increase exponently in terms of size, it can be quite beneficial, if we can drop some of these variables,” he adds.
Adaptable, scalable approach
To test their approach, scientists compared L-RHO with several basic algorithmic solvers, specialized solutions and approaches that only operate machine learning. All this exceeded, shortening the solution time by 54 percent and improving the quality of the solution by up to 21 percent.
In addition, their method still exceeded all base lines when they tested it on more convoluted variants of the problem, for example, when factory machines are distributed or when there is additional congestion of the train. Even surpassed the additional foundations of researchers, which scientists created to challenge them.
“Our approach can be used without modifying all these different variants, which we really decided to do with this test line,” he says.
L-RHO can also adapt if the goals change, automatically generating a modern algorithm to solve the problem-everything he needs is a modern set of training data.
In the future, scientists want to better understand the logic of the decision of their model to freeze some variables, but not others. They also want to integrate their approach with other types of convoluted optimization problems, such as inventory management or vehicle routing.
These works were partly supported by the National Science Foundation, Mit’s Research Support Committee, Dr. Robotics Dr. Scholarship and Mathworks.