The task of optimizing the delivery of holiday packages is a complex issue for logistics companies like FedEx, which often leverages specialized software known as a mixed-integer linear programming (MILP) solver. This software breaks down complex optimization problems into smaller parts and employs generic algorithms to find the best solutions. However, this process can take hours or even days, and often companies have to stop the software halfway, thus accepting suboptimal results due to time constraints.
To solve this challenge, researchers from the Massachusetts Institute of Technology (MIT) and ETH Zurich used machine learning (ML) to increase the pace of the process. The team identified an intermediate step in MILP solvers that involves searching through myriad potential solutions, thus slowing down the entire process. The researchers then developed a machine learning-based filtering technique to simplify this step, and utilized the resulting model to find the optimal solution for a specific type of problem.
By speeding up MILP solvers between 30 and 70 percent, without compromising accuracy, the newly developed technique effectively allows logistics companies to generate optimal solutions more quickly. Moreover, for particularly complex problems, the method enables finding superior solutions within a manageable timeframe.
The novel approach can be used across a range of industries where MILP solvers are employed, such as ride-hailing services, electric grid operators, vaccine distributors, and any organization encountering complicated resource allocation dilemmas.
The success of this approach underscores the potential of combining traditional solutions with machine learning to tackle complex optimization problems.
To streamline the process, the research team created a filtering mechanism, which reduces potential solutions from a myriad of combinations to around 20 options. Subsequently, they employed a machine learning model to select the ideal combination of algorithms from among these 20 options. The ML model is trained using a dataset that is specific to the optimization problem at hand, thus allowing it to select algorithms that best suit the task. This permits companies like FedEx to employ past experience and pre-existing data to resolve new problems more efficiently.
Next, the researchers plan to apply this model to even more complicated MILP problems. The team is also eager to delve deeper into the model to better understand the effectiveness of various algorithms.
This research was partially funded by Mathworks, the National Science Foundation (NSF), the MIT Amazon Science Hub, and MIT’s Research Support Committee. The research findings will be presented at the upcoming Conference on Neural Information Processing Systems.
