Once some commuter trains arrive at the end of the line, they will need to move from a different platform than the one they arrived on to the switching platform so they can depart the station later.
Engineers plan these movements using software programs called algorithm solvers, but at stations with thousands of arrivals and departures each week, problems become complicated for traditional solvers to unravel at once.
Using machine learning, MIT researchers develop improved planning systems that reduce resolution times by up to 50%, generating solutions that better meet user goals, such as on-time train departures. This new method can also be used to efficiently solve other complex logistical issues, such as scheduling hospital staff, assigning airline crews, and assigning tasks to factory machines.
Engineers often divide these types of problems into a series of overlapping sub-problems that each can solve in a feasible time. However, overlapping causes many decisions to be recalculated unnecessarily, which takes the solver much longer to reach the optimal solution.
A new AI-enhanced approach will learn which parts of each sub-problem need to be kept unchanged and freeze these variables to avoid redundant calculations. The traditional algorithm solver then addresses the remaining variables.
“In many cases, dedicated teams can design algorithms for months or years to solve only one of these combination problems. Modern deep learning allows us to use new advances to streamline the design of these algorithms. Members of MIT’s Data, Systems and Society (IDSS), and Information and Decision Systems (LID) labs.
She was joined the paper by lead author Sirui Li, a graduate student at IDSS. Wenbin Ouyang, a graduate student at CEE. And then yining ma, postdoctoral cover. This research will be presented at the International Conference on Learning Expression.
Eliminate redundancy
One of the motivations for this study is the practical issues identified by Master’s Devin Camille Wilkins in WU’s entry-level transport course. Students wanted to apply reinforcement learning to actual train dispatch issues at Boston’s North Station. Transportation must allocate many trains to a limited number of platforms that can rotate before arriving at the station.
This turned out to be a very complicated combination scheduling issue. This is the exact type of issue that WU Labs have been working on over the past few years.
When faced with long-term problems, including assigning limited resources to groups of machines, such as factory tasks, planners often assemble the problems as flexible job shop scheduling.
Flexible Job Shop scheduling requires a different time for each task to complete, but you can assign tasks to any machine. At the same time, each task consists of operations that need to be performed in the correct order.
These problems become too big and difficult to handle for traditional solvers, so users can use Rolling Horizon Optimization (RHO) to split the problems into manageable chunks that can be solved faster.
With Rho, users assign the first few tasks to the machine on a fixed plan horizon, perhaps on a four-hour time frame. Then, run the first task in that sequence, shift the 4-hour planning horizon forward and add the next task, and repeat the process until the entire problem is resolved and the final schedule for the task machine assignment is created.
If you also consider tasks where algorithms appear, the planning period should be longer than the duration of one task, as the solution is better.
However, as the planning horizon progresses, this creates an overlap with previous operationalities on the planned horizon. The algorithm has already created a preliminary solution for these duplicate operations.
“Maybe these preliminary solutions are good and don’t have to do any calculations again, but they may not be good. This is where machine learning comes in,” explains Wu.
For a technique called learning guided rolling horizon optimization (L-RHO), researchers teach machine learning models and predict which operations or variables to be recalculated as the planning period advances.
Because L-RHO requires data to train the model, researchers use classical algorithm solvers to solve a set of sub-problems. They took the best solutions – the most solutions of operations that don’t need to be reimported – and used these as training data.
Once trained, the machine learning model receives new sub-problems that we have never seen before and predicts which operations should not be recalculated. The rest of the operations are fed to an algorithm solver that performs tasks, reconstructs these operations, and moves forward the planning horizon. The loop then begins again.
“In hindsight, if you didn’t need to re-express them, you can remove these variables from the problem. These problems get exponentially large, so dropping some of these variables could be very advantageous,” she adds.
An adaptable, scalable approach
To test their approach, researchers compared the L-RHO with some basic algorithm solvers, specialized solvers, and approaches that use only machine learning. It surpassed all of them, reducing time by 54% and improving the quality of the solution by up to 21%.
Furthermore, the method continued to surpass all baselines when tested with more complex variants of the issue, such as when factory machines broke down or when there was additional train crowds. It even surpassed the additional baselines researchers created to challenge solvers.
“Our approach can be applied to all these different variations without changing, which is what we’re trying to do with this research line,” she says.
L-RHO can also be adapted if the target changes. Automatically generate new algorithms to solve the problem. All you need is a new training data set.
In the future, researchers want to better understand the logic behind the model’s decision to freeze some variables, but others want to better understand them. They also want to integrate their approaches into other types of complex optimization problems, such as inventory management and vehicle routing.
This work was supported in part by the National Science Foundation, the Research Assistance Committee at MIT, the Amazon Robotics PhD Fellowship, and Mathworks.
Disclaimer: Includes third-party opinions. No financial advice.
