Webinar: The Dutch Optimization Platform (DOP)
1st speaker (12:00-12:30): Ilker Birbil (University of Amsterdam)
Title: Optimization with Constraint Learning
Abstract: What if your solver could handle not only traditional domain constraints but also constraints derived directly from data? In this talk, I will introduce a novel approach we call Optimization with Constraint Learning (OptiCL), which transforms machine learning models into mixed-integer optimization formulations. I will also discuss two strategies to manage the inherent uncertainty of learned constraints: trust regions that ensure decisions stay within data-rich, reliable areas, and ensemble methods that enhance robustness. Finally, I will showcase the practical impact of OptiCL through real-world applications, including humanitarian logistics, personalized medicine, and explainable artificial intelligence.
Bio: Ilker Birbil is a professor at the University of Amsterdam, where he leads the Business Analytics Department. Previously, he was a professor of Data Science and Optimization at the Econometric Institute of Erasmus University. Before that, he spent over a decade as a professor of Optimization in the Industrial Engineering Department at Sabancı University. His research centers on developing optimization methods for artificial intelligence and decision-making. Recently, his work has expanded to focus on explainable and trustworthy optimization.
2nd speaker (12:30-13:00): Jean Pauphilet (London Business School)
Title: Constructing a diverse set of nearly optimal solutions
Abstract: When implementing solutions provided by optimization models, computing a diverse set of near-optimal solutions, instead a single 'optimal' one, can be of critical importance. Unlike the single-solution approach, such a diverse set enables decision-makers to select among high-quality alternatives while considering additional, possibly unmodeled, criteria. This talk provides a self-contained review and computational assessment of the state of the literature for generating diverse near-optimal solutions to an optimization problem. In particular, our analysis highlights the strong empirical performance of an easy-to-use heuristic, based on random search directions, which can be readily adopted by practitioners. We also illustrate the value of generating such diverse near-optimal solutions on two real-world use cases.
Bio: Jean Pauphilet is an Assistant Professor of Management Science and Operations at London Business School. His research develops data analytics methods for positive-impact applications. He has designed new algorithms for large-scale optimization, machine learning, and optimization under uncertainty, and has applied them to improve hospital operations or reduce ocean plastic pollution. He holds a PhD in Operations Research from the Massachusetts Institute of Technology (MIT).