When some commuting trains arrive at the end of the line, they must travel to the switching platform to surround themselves so that the station can later leave the platform than the one they arrived at.
Engineers use software programs called Algorithmic Shelvers for planning these movements, but at a station with thousands of weekly arrivals and departures, the problem becomes too complicated for the traditional solver to disintegrate at once.
Using machine learning, scientists have developed an improved planning system that shortens the resolution time by up to 50 pierieties and creates solutions that better meet the goal, such as early training. The new method could also be used to effectively solve other complex logistics problems, such as planning the hospital staff, striving airlines, or assigning tasks to factory machines.
Engineers often divide these types of problems into a sequence of overlapping sub -problems that can be solved in a feasible love of time. However, overlapping causes Mayry’s decisions to be unnecessarily renewed, so the solver takes a long time to achieve an optimal solution.
The new, artificial intelligence with increased approach teaches which parts of each partial problem should remain unchanged and freeze these variables to avoid red -based calculations. Then the traditional algorithmic solver deals with the remaining variables.
“A frequently devoted team can often spend months or even years of designing an algorithm to solve only one of these combinatorial problems. Modern deep learning gives us the opportunity to take advantage of new progress to make the design of these algorithms. and institute) and institutes and institutions) and institute) and institute) and institutions) and institute) and institute) and institute) and institutions (MIT institutions and a member of the Laboratory for Information and Decision Systems (Lids).
The paper was joined by the chief author Sirui Li, a postgraduate student IDSS; Wenbin Ouyang, in a postgraduate student; And Yining Ma, postdoc lids. The research will be presented at an international conference on learning representations.
Elimination
One of the motivation for this research is a practical problem identified by Master’s student Devin Camille Wilkins at the Wu at the basic level. The student wanted to apply learning to strengthen the real problem with the train-dispatch to the North Station in Boston. Transit organizations must assign many trains to a limited number of platforms where they can be rotated well in advance before arrival at the station.
It turned out that this is a very complicated problem with combinatorial planning – the exact type of problem laboratory WU has been in the last few years.
Facting a problem with a long country that includes accommodation of a limited set of resources, such as factory tasks, for a group of machines, planners often frame this problem as flexible work planning.
In flexible shop planning, every task needs to complete a different time, but the tasks can be assigned to each machine. At the same time, each task is composed of operations that must be performed in the correct order.
For traditional solvers, these problems quickly become too large and cumbersome, so users can employ the optimization of the horizon (RHO) to divide the problem into managing pieces that can be solved faster.
With RHO, the user assigns the initial tasks to machines in a fixed horizon of planning, a possibly four -hour time window. They then perform the first role in this sequence and move the four-hour planning horizon forward to add another task, repetition of the process until the problem is resolved and the final task schedule is created-Machine assignment.
The planning horizon should be longer than any task duration, because the solution will be better if the algorithm also considers the tasks that appear.
However, when the planning horizon proceeds, it creates some overlap with operations in the previous planning horizon. The algorithm has already come up with preliminary solutions to these overlapping operations.
“Maybe these preliminary solutions are good and Don needs to be calculated again, but maybe they are good. Machine learning comes here,” Wu explains.
For their technique, which they call optimization of learning tension (L-RHO), scientists teach a model of machine learning that predict that operations or variables should recover when the planning horizon moves forward.
L-RHO requires the data to train the model, so scientists solve a set of partial problems using a classic algorithmic solver. They took the best solution – the one with the most operations you don’t have to be suppressed – and used it as a training data.
Once the machine learning model has been trained, it will receive a new sub -project that has been previously and predicts which operations should not be recurred to. The remaining operations are brought back to the algorithmic solver, which performs the task, these operations recalculate and move the forward planning horizon. Then the loop begins again.
“If we did not have to re -optimize them when we look back, then we can remove these variables from the problem. Because these problems grow exponentially in size, this may be this advantage if we can abandon any of these variables,” he adds.
Adaptable, scalable approach
To test their approval, scientists have compared L-RHO with several basic algorithmic researchers, specialized researchers and approaches that only use machine learning. All of them overcame and reduce their time by 54 pers and improved the quality by up to 21 Pierge.
In addition, their method continued to overcome all the basic lines when they tested it on more complicated variants of the problem, such as when factory machines disintegrate or when the train is overloaded. It is about overcoming other basic lines that scientists have created to question their solver.
“Our approval can be used without adjustment to all these different variants, which is what we have determined with this line of research,” he says.
The L-RHO can also adapt if the goals change, automatically generates a new algorithm to solve the problem-about what it needs is a new data set of training.
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 consent into other types of comprehensive optimization problems such as inventory management or vehicle direction.
This work was partially supported by the National Science Foundation, the MIT research committee, Amazon Robotics PhD Fellowship and Mathworks.
(Tagstotranslate) Cathy Wu