Efficiently routing holiday packages is an intricate computational problem for delivery companies such as FedEx. So complex is the problem that companies often implement specialized software, termed a mixed-integer linear programming (MILP) solver. Yet, the solver may take prolonged times to offer a solution, leading companies to conclude midway, settling for suboptimal solutions bounded by time constraints.
Scientists from Massachusetts Institute of Technology (MIT) and the Swiss Federal Institute of Technology in Zurich (ETH Zurich) have utilized machine learning processes to quicken the system. They observed a crucial intermediate phase in MILP solvers, requiring such a considerable amount of time to solve that it decelerates the entire operation. The researchers utilized a filtering method to streamline this phase and then used machine learning to find the optimal solution for a specific type of problem. This enabled the company to customize a general-purpose MILP solver to suit its unique problem based on its data.
This novel technique has quickened MILP solvers by between 30 and 70 percent with no compromise in precision. It may provide an optimal solution more rapidly, and for notably complex problems, a feasible solution within a practical time. This technique can be applied where MILP solvers are employed, such as ride-hailing services, electric grid operators, vaccination distributors, or entities dealing with intricate resource-allocation problems.
Researchers noted that MILP problems lead to an exponential number of potential solutions and are considered NP-hard, meaning there’s likely no efficient algorithm to solve them. The researchers devised a filtering mechanism that substantially reduced the search space for the optimal combination of algorithms. They then used a machine-learning model to select the best combination from the remaining options.
This model uses a dataset specific to the user’s optimization problem, making it more User-specific and efficient. This data-driven approach accelerated MILP solvers without any drop in accuracy. The researchers hope to apply this technique to more complex MILP problems, and the project has the support of Mathworks, the National Science Foundation (NSF), the MIT Amazon Science Hub, and MIT’s Research Support Committee.
