Solving optimization problems with machine learning
Faculty project looks at using machine learning to improve how difficult optimization problems are solved
Curiosity drives everything at Dickinson. In labs, studios, archives and communities around the world, students and faculty are asking questions big and small—and often finding surprising answers. Here is a look at a query.
Researchers: Dick Forrester (and Lucas Waddell, Bucknell University)
Project
The quadratic multidimensional knapsack problem models decisions in areas such as finance, logistics and resource allocation, where several resource limits must be satisfied and some combinations of choices are worth more together than separately. The research compares traditional optimization methods with newer approaches that incorporate machine learning and genetic algorithms.
As Forrester explains: “Machine learning is usually designed to make predictions, not to solve optimization problems. What interested us was whether those predictions could be used to make optimization algorithms better. We trained machine-learning models to recognize which choices were most likely to appear in the best solutions, then used that information within several optimization methods. The strongest approach combined those predictions with a genetic algorithm and produced solutions within 0.75% of the best possible solution, on average.”
This content is sourced from
Dickinson College
. It reflects the author's views and has not been edited by our newsroom. It may have been generated using AI assistance.
