Companies like FedEx utilize intricate software to efficiently deliver holiday parcels, but these complex processes can often take hours or even days to complete. The software, known as a mixed-integer linear programming (MILP) solver, is often halted partway through by firms, accepting the best solution that can be gleaned in a particular timeframe, even if it isn’t ideal. However, a collaborative team of researchers from MIT and ETH Zurich are looking to speed up this process through machine learning techniques.
Specifically, the researchers identified a key intermediate phase in MILP solvers that has an enormous number of potential solutions, which makes the process lengthy and slow. To combat this, they used AI to simplify this stage, permitting machine learning to uncover the optimal solution to a given type of issue. This enables firms to make use of their own data to optimize generic MILP solvers for the issues they’re facing.
The new speedy machine learning method can speed the MILP solvers up by anywhere between 30-70%, without compromising on accuracy. This technique could save time when attempting to find an optimal solution or could even create a superior solution in a timeframe that is more manageable for complex issues.
The researchers’ method could be of benefit wherever MILP solvers are used. This could be in the use of ride-hailing services, electric grid operators, vaccine distribution, or for an entity dealing with a difficult resource-distribution issue.
Cathy Wu, Senior Author of the paper discussing the research, believes these hybrid solutions which combine machine learning and classical approaches can provide the best solutions. Co-lead authors Sirui Li and Wenbin Ouyang; and Max Paulus, a graduate student at ETH Zurich, also contributed to the paper. The research will be presented at the Conference on Neural Information Processing Systems.
MILP problems contain an exponential number of potential solutions, making them incredibly difficult to solve efficiently. A total MILP solver uses an array of practical techniques and methods to uncover reasonable solutions in a realistic timeframe.
Researchers found the process of identifying the ideal combination of separator algorithms to use was a problem with a vast number of solutions. The researchers developed a filtering mechanism that reduced this separator search space from over 130,000 potential combinations to around 20 options. Finally, a machine-learning model was used to determine the best combination of algorithms from the remaining 20 options.
This model learns from specific user data to select algorithms that best serve each user’s unique task, improving the final solutions. The machine-learning model, known as contextual bandits, follows an iterative learning process whereby a solution is chosen, feedback on its effectiveness is obtained, and then the process repeats in order to find a more efficient solution.
The new data-centric approach decreased the time taken by MILP solvers by 30-70% without reducing accuracy. The researchers plan to apply the technique to even more complex MILP problems in future.
This project has been part-funded by Mathworks, the National Science Foundation (NSF), the MIT Amazon Science Hub, and MIT’s Research Support Committee.
