Stochastic Black-Box Optimization using Multi-Fidelity Score Function Estimator - Archive ouverte HAL
Pré-Publication, Document De Travail (Preprint/Prepublication) Année : 2024

Stochastic Black-Box Optimization using Multi-Fidelity Score Function Estimator

Résumé

Optimizing the parameters of physics-based simulators is crucial in the design process of engineering and scientific systems. This becomes challenging when the simulator is stochastic, computationally expensive, black-box, multi-modal, and has a high-dimensional parameter space, such as when simulating complex climate models that involve numerous interacting variables and uncertain parame- ters. Many traditional optimization methods rely on gradient information, which is frequently unavailable in legacy black-box codes. To address these challenges, we present SCOUT-Nd (Stochastic Constrained Optimization for N dimensions), an algorithm designed to efficiently estimate gradients, reduce noise in the gradient estimator, and enhance convergence properties through the use of natural gradients. SCOUT-Nd also incorporates multi-fidelity schemes and an adaptive number of samples to minimize computational effort. We validate our approach using stan- dard benchmark analytical problems, demonstrating its superior performance in parameter optimization compared to existing methods. Additionally, we showcase the algorithm’s efficacy in a complex real-world application: optimizing wind farm layout.
Fichier principal
Vignette du fichier
scout_nd_preprint_neurips_formal_ml_scitech.pdf (7.19 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04659802 , version 1 (23-07-2024)

Identifiants

  • HAL Id : hal-04659802 , version 1

Citer

Atul Agrawal, Phaedon-Stelios Koutsourelakis, Kislaya Ravi, Hans-Joachim Bungartz. Stochastic Black-Box Optimization using Multi-Fidelity Score Function Estimator. 2024. ⟨hal-04659802⟩

Collections

TDS-MACS
76 Consultations
81 Téléchargements

Partager

More