When some commuter trains arrive at the end of the line, they should travel on a switching platform so that they can depart the station later, often from a separate platform on which they arrive.
Engineers use a software program called Elgorithm Solver to plan these movements, but at a station with thousands of weekly arrival and departure, the problem becomes very complicated to highlight everyone at a bar at a time for a traditional solver.
Using machine learning, MIT researchers have developed a better plan system that reduces the solution time by 50 percent and produces a solution that fulfills the purpose of a better user, such as train departure on time. The new method can also be used efficiently to solve other complex logical problems, such as schedule of hospital staff, assign airline crew, or allocate work to factory machines.
Engineers often break the overproflam in the sequence of overlapping, which can be resolved in possible time. But many decisions due to overlaps start unnecessarily, so the solver takes longer to reach an optimal solution.
The new, artificial intelligence-communicated approach learns which parts of each sub-availability should remain unchanged, those variables to avoid fruitless computation. Then a traditional algorithm solver deal with the remaining variables.
“Often, a dedicated team can spend for months or years, even in designing an algorithm to solve one of these combinatorial problems. Modern deep learning gives us an opportunity to use new advances to help us streamline the design of these algorithms. We can take whatever we know well, and use it to speed up, and use it, (using it) In MIT, and a member of the laboratory for information and decision systems (LIDS).
He is included on paper by Sirui Lee, an IDSS graduate student Sirui Lee; Venbin Oying, a CEE graduate student; And MA, a lids postdock. Research will be presented at the International Conference on the representation of learning.
Excess
An inspiration for this research is a practical problem identified by Wu’s Entry-Level Transportation Course by a master student of a master. The student wanted to implement learning reinforcement for a real train-dispatch problem at Boston’s North Station. The transit organization needs to assign multiple trains on a limited number of platforms, where they can be rotated well before their arrival at the station.
This is a very complex combinatorial scheduling problem – the exact type of problem Wu’s lab has spent working in the last few years.
When a long -term problem faces a group of machines involving assigning a limited set of resources like factory tasks, planners often frame the problem as flexible job shop time determination.
In flexible job shops scheduling, a different time is required to complete each task, but the tasks can be assigned to any machine. At the same time, each function is made up of operations that should be done in the right order.
Such problems are quickly large and unknown to traditional solvers, so the user can appoint rolling horizon adaptation (RHO) to break the problem in managing fragmentation that can be rapidly resolved.
With RHO, a user offers machines in a certain planning horizon, perhaps a four -hour time window. Then, they execute the first task in that sequence and move the four-hour plan to further connect the next task, repeat the process until the entire problem is solved and the final schedule of the task-masine assignment is formed.
A planning horizon should be longer than the duration of any task, as the solution would be better if the algorithm also considers the tasks that are coming.
But when the planning horizon moves forward, it creates some overlap with operations in the previous planning horizon. The algorithm had already come up with an initial solution to these overlapping works.
“Perhaps these initial solutions are good and do not need to be re -calculated, but perhaps they are not good. This is where the machine learning comes,” Wu explains.
For their technology, which they call the learning-guided rolling horizon optimization (L-RHO), researchers taught a machine-learning model to estimate which operations, or variables, should be resumed when the plan proceeds when the plan proceeds.
The L-RHO requires data to train the model, so researchers solved a set of subtypes using classical algorithm solver. They took the best solution – the most operating who do not need to re -order – and used them as training data.
Once trained, the machine-learning model receives a new subtle that she has not seen before and predicts which operations should not be rearranged. The remaining operations are fed back to the algorithm solver, which executes the task, resumes these operations, and the planning carries forward the horizon. Then the loop starts again.
“If, if, we did not need to re -prepare them, we can remove those variables from the problem. Because these problems grow rapidly, it can be quite beneficial if we can leave some of those variables,” she says.
A adaptable, scalable approach
To test their approach, researchers compared the L-RHO to several base algorithm solvers, special solvers, and approaches, which only use machine learning. This improved all of them, reduced the time to solve time by 54 percent and improved the quality of the solution by 21 percent.
In addition, their method continued to improve all baseline when they tested it on more complex variants of the problem, such as when the factory machines break or when there is a congestion of additional train. It even improved additional base lines that researchers made their solver to challenge.
She says, “Our approach can be implemented without amending all these different variants, which really we want to do with this line of research,” she says.
The L-RHO can also adapt. If the objectives change, automatically generate a new algorithm to solve the problem-its need is a new training dataset.
In the future, researchers want to better understand the argument behind their model’s decision to freeze some variables, but not others. They want to integrate their approach in other types of complex adaptation problems such as inventory management or vehicle routing.
The work was supported, in part by the National Science Foundation, MIT’s Research Support Committee, an Amazon Robotics PhD Fellowship and Mathwords.
