Communication Dans Un Congrès Année : 2026

Meta-learning with similarity-weighted transfer for manufacturing optimization

Résumé

The MBAE (Meta-Bayesian Active Explorer) framework addresses the critical challenge of sample efficiency in Industry 4.0 manufacturing optimization, where high-dimensional search spaces and expensive physical evaluations limit experimental budgets. By integrating hierarchical Bayesian meta-learning with similarity-weighted transfer, the model effectively overcomes the "cold-start" problem by leveraging historical task data while preventing negative transfer through a learned task-feature mapping. Unlike traditional evolutionary algorithms or standard Bayesian Optimization, MBAE utilizes a manufacturing-aware acquisition function that balances exploration, exploitation, and operational preferences with formal regret-bound guarantees. Experimental validation on 9D manufacturing problems demonstrates a 4.3x to 4.6x convergence speedup, reaching 95% optimality in approximately 6 iterations compared to over 26 for baseline methods, representing a significant advancement for rapid adaptation in reconfigurable production systems.

Fichier principal
Vignette du fichier
ROADEF2026 (1).pdf (292.86 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05548907 , version 1 (17-03-2026)

Licence

Identifiants

  • HAL Id : hal-05548907 , version 1

Citer

Madani Bezoui, Hajar Nouinou, Ahcène Bounceur. Meta-learning with similarity-weighted transfer for manufacturing optimization. Société Française de Recherche Opérationnelle et d'Aide à la Décision, Feb 2026, Tours, France. ⟨hal-05548907⟩
74 Consultations
27 Téléchargements

Partager

  • More