C. D'Ambrosio, M. Cuesta Santa Teresa, M. Durbán, V. Guerrero, R. Spencer Trindade
This work introduces an innovative integration of data science into mathematical optimization to solve complex MINLP problems. Smooth additive models are used to approximate non-convex objective functions and/or constraints. A surrogate MINLP problem is then achieved by replacing these complex components with their tractable approximations. In addition, the shape-constrained estimation capabilities of smooth additive models are leveraged to incorporate expert knowledge, ensuring that properties such as non-negativity and monotonicity are preserved in the approximations performed. Thus, this approach builds accurate and tractable surrogate MINLPs that are both data-driven and knowledge-driven. Our approach is shown to be competitive in benchmark MINLP instances and real case studies, such as the Hydro Unit commitment problem.
Keywords: statistical modeling, data science, mathematical optimization, MINLP, surrogate MINLP
Scheduled
GT03. AMC1 Machine Learning
June 10, 2025 11:30 AM
Auditorio 1. Ricard Vinyes