Dr. Dimitrios LETSIOS, currently postdoctoral fellow at the Department of Computing (DoC) at Imperial College London, will give a seminar on Tue, Jan 22 at 14:30-15:20 in Salle Philppe Flajolet (in front of Vanessa’s office on the second floor).

Abstract: Machine learning and data science methodologies leverage historical data to produce powerful models approximating unknown functions and predicting uncertain parameters. Mathematical optimization offers strong modeling paradigms, algorithms, and solvers for solving complex decision-making problems with analytical methods. This talk shall present approaches for combining machine learning and data science models with mathematical optimization methods, in a unified setting, towards effective resource allocation. The goal is the design of tailor-made, exact optimization methods exploiting the structure of predictive model outcomes. We shall discuss using gradient-boosted trees (GBTs) to approximate unknown functions and solving new optimization problems with trained GBTs embedded for reducing the energy consumption in industrial processes. We shall also discuss using robust combinatorial optimization with uncertainty sets obtained from data for effective production planning and scheduling. Finally, the talk shall present applications of the proposed approaches and numerical results with real data.