Sampling Criteria for Constrained Bayesian Optimization under Uncertainty - Mines Saint-Étienne Access content directly
Conference Papers Year : 2022

Sampling Criteria for Constrained Bayesian Optimization under Uncertainty

Abstract

We consider the problem of chance constrained optimization where the objective and the constraint functions are affected by uncertainties and are computationally costly. Bayesian optimization is an appropriate family of methods to address such problems. We first propose a two-step acquisition criterion defined in the joint space of optimization variables and uncertain parameters. The objective and the constraints are aggregated through a feasible improvement measure and the two steps consist in the optimization of the expectation and the one-step-ahead variance of this criterion. To ease the computational burden, an analytical approximation to the one-step-ahead variance is proposed. Additionally, we also account for the possible correlation between the constraints. This is done by considering a vector-valued "input as output" joined Gaussian process which improves the constraints modeling accuracy and consequently the optimization procedure. The correlations between the constraints are further exploited by allowing each constraint to be evaluated for different uncertain parameters and by optimally selecting a subset of constraints to be evaluated at each iteration, thus avoiding unnecessary computations. Numerical tests confirm the applicability and potential gains brought by these methods, such a faster convergence speed and better scaling with respect to the number of constraints if compared to alternative optimization methods.
Pres_SIAM2022_JPI.pdf (1015.12 Ko) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

emse-03657435 , version 1 (18-05-2022)

Identifiers

  • HAL Id : emse-03657435 , version 1

Cite

Julien Pelamatti, Mohamed Reda El Amri, Rodolphe Le Riche, Christophette Blanchet-Scalliet, Céline Helbert. Sampling Criteria for Constrained Bayesian Optimization under Uncertainty. Hybrid Conference: 2022 SIAM Conference on Uncertainty Quantification, Apr 2022, Atlanta, United States. ⟨emse-03657435⟩
178 View
43 Download

Share

Gmail Facebook X LinkedIn More