Poster De Conférence Année : 2024

Numerical study of the calibration of predictive distributions from Gaussian process interpolation

Résumé

When a function $\mathcal{f}$ : $\mathbb{R}$d → $\mathbb{R}$ is observed at points x1,...,xn, Gaussian Process (GP) interpolation is a standard method to infer the value of f at a new point x. However, what degree of trust can be placed in a prediction from a GP model? The natural way to deal with this question is to study the prediction intervals built from the posterior distribution of the GP model at x. These prediction intervals may be too optimistic or too pessimistic compared to the actual observations (in the precise sense described below). These considerations are crucial in the industry since prediction intervals that are too optimistic can be misleading and have disastrous effects. In contrast, overly pessimistic prediction intervals leads the user to increase the number of observations, resulting in an increasing approximation cost.

Fichier principal
Vignette du fichier
poster_mascotnum2024.pdf (352.34 Ko) Télécharger le fichier
mascotnum2024-abstract_Aurelien_Pion.pdf (272.51 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05299837 , version 1 (06-10-2025)

Licence

Identifiants

  • HAL Id : hal-05299837 , version 1

Citer

Aurélien Pion, José Alves, Emmanuel Vazquez. Numerical study of the calibration of predictive distributions from Gaussian process interpolation. MASCOT-NUM 2024, Apr 2024, Hyères, France. . ⟨hal-05299837⟩
68 Consultations
75 Téléchargements

Partager

  • More