Exploring Neural Combinatorial Optimization: Current State and Future Perspectives

• Autores: Andoni Irazusta Garmendia
• Directores de la Tesis: Josu Ceberio Uribe (dir. tes.), Alexander Mendiburu Alberro (codir. tes.)
• Lectura: En la Universidad del País Vasco - Euskal Herriko Unibertsitatea ( España ) en 2024
• Idioma: inglés
• Programa de doctorado: Programa de Doctorado en Ingeniería Informática por la Universidad del País Vasco/Euskal Herriko Unibertsitatea
• Menciones:
  • Internacional
• Calificación: Sobresaliente Cum Laude
• Materias:
  • Ciencias tecnológicas
• Enlaces
  • Tesis en acceso abierto en: ADDI
• Resumen

  • Traditional problem-solving approaches have relied heavily on handcrafted heuristics, which are developed based on domain-specific knowledge and offer practical but non-optimal solutions under time constraints or incomplete information. While effective, these heuristics require significant expertise and are often limited in generalizability across different problem variations. Recent advances in Deep Learning (DL) and Neural Network (NN) architectures have revolutionized this landscape, particularly in fields like medical diagnosis, game strategies, spam detection, and credit scoring. These systems learn complex patterns from large datasets, automating the decision-making process and surpassing traditional rule-based methods.

    In the realm of Combinatorial Optimization (CO), this shift has given rise to Neural Combinatorial Optimization (NCO), where DL techniques are used to automate the generation of heuristics for complex tasks. While NCO has shown promise in problems like the Traveling Salesman Problem (TSP), research on its broader applicability remains limited, and there is a gap in understanding why and under what conditions these methods are most effective.

    This thesis aims to explore the potential of NCO across a broader set of CO problems, analyzing its performance compared to traditional methods and identifying its strengths and limitations. Key contributions include the development of neural constructive methods and neural improvement methods, the integration of memory modules to enhance search efficiency, and the proposal of the MARCO framework, which incorporates memory augmentation to improve combinatorial optimization.