Gaussian processes for computer experiments with monotonicity information

Abstract : Recently, statistical researchers have shown increased interest in Gaussian process modeling with mono- tonicity constraints (see [2], [4] and [5]). In computer experiments, the true function (scalar output) may be known to be monotone with respect to some or all input variables. We propose a new methodology based on the Bayesian Gaussian process metamodeling to sample from posterior distribution including monotonicity information in the monovariate case. Let y = f (x) be a monotonic increasing function where the input x is assumed to be scalar and in the domain [0, 1]. We consider a set of computer experiments {(xi , yi ) | i = 1, * * * , n} of size n and assume that yi = f (xi ), 1 ≤ i ≤ n. (1) Also suppose that M is the space of increasing functions and (Yx )x∈[0,1] is a zero-mean Gaussian process (GP) with kernel k(x, xt) given by a priori knowledge about the relationship between the input x and the output y. The following experimental results (see figures below) are obtained with the classical Gaussian kernel. We are interested in the simulation of the conditional (or posterior) distribution of the GP Y given data and monotonicity information...
Type de document :
Poster
MascotNum Annual Conference, Apr 2014, Zurich, Switzerland. 2p., 2014
Liste complète des métadonnées

https://hal-emse.ccsd.cnrs.fr/emse-00960345
Contributeur : Florent Breuil <>
Soumis le : mardi 18 mars 2014 - 09:37:02
Dernière modification le : jeudi 11 janvier 2018 - 06:16:31

Identifiants

  • HAL Id : emse-00960345, version 1

Citation

Hassan Maatouk. Gaussian processes for computer experiments with monotonicity information. MascotNum Annual Conference, Apr 2014, Zurich, Switzerland. 2p., 2014. 〈emse-00960345〉

Partager

Métriques

Consultations de la notice

306