The solver says _optimal_. Now what? Transparency tools for large (mixed-integer) linear models.
Abstract
Large mathematical programming models, such as the ones arising in Energy System modelling, can be somewhat close to black boxes, as they typically possess thousands of parameters, millions of constraints and variables. This talk presents recent work on making large (mixed-integer) linear models more transparent & explainable. We will first focus on sensitivity analysis of multi-coefficient modifications of the constraint matrix, which arise when a single uncertain parameter, such as a technology’s efficiency, appears in thousands of coefficients. The talk presents how we can efficiently bound the impact of these modifications on the objective and the solution and present warm-starting methods, greatly reducing the computational burden. The second part will show how language models can detect silent modeling errors in models that solve without complaint.
Bio
Guillaume Derval is a postdoctoral researcher at the University of Liège, Belgium, and holder of the SCK CEN Chair on Small Modular Reactor Integration in the Belgian Energy Landscape. His research covers (mixed-integer) linear programming, constraint programming and modeling languages, with a focus on sensitivity analysis, parametric and robust optimization, and the explainability of large mathematical programming models such as energy-system models.