ML-Driven Seeding for Constrained Optimization Problems

• Autores: Bermúdez, Carlos; Alfonso, Hugo Alfredo; Carnero, Mercedes; Hernández, José Luis; Minetti, Gabriela F.; Salto, Carolina
• Año de publicación: 2025
• Idioma: inglés
• Tipo de recurso: documento de conferencia
• Materia: Ciencias Informáticas; Machine Learning; Constrained Optimization Problem; Seeding; Metaheuristic
• Institución: Universidad Nacional de La Plata

The generation of feasible solutions is a crucial step in metaheuristic approaches to constrained optimization problems. The development of seeding strategies relies on random or problem-specific heuristic methods, which can be challenging to design and adapt effectively for diverse problem domains. This work proposes the Machine Learning Seeding Tool (MLST), a data-driven approach that leverages supervised learning to improve solution generation. MLST is proven through its application to the Water Distribution Network Design (WDND) problem, where feasibility constraints pose significant computational challenges. MLST leverages five supervised classification algorithms to generate high-quality WDND solutions, reducing the Total Investment Cost compared to the traditional method. While results demonstrate MLST’s effectiveness, class imbalance in pipe diameters highlights the need for future refinements using advanced resampling techniques.
Red de Universidades con Carreras en Informática