Skip to Main content Skip to Navigation

Gaussian Process Modelling under Inequality Constraints

Abstract : Conditioning Gaussian processes (GPs) by inequality constraints gives more realistic models. This thesis focuses on the finite-dimensional approximation of GP models proposed by Maatouk (2015), which satisfies the constraints everywhere in the input space. Several contributions are provided. First, we study the use of Markov chain Monte Carlo methods for truncated multinormals. They result in efficient sampling for linear inequality constraints. Second, we explore the extension of the model, previously limited up tothree-dimensional spaces, to higher dimensions. The introduction of a noise effect allows us to go up to dimension five. We propose a sequential algorithm based on knot insertion, which concentrates the computational budget on the most active dimensions. We also explore the Delaunay triangulation as an alternative to tensorisation. Finally, we study the case of additive models in this context, theoretically and on problems involving hundreds of input variables. Third, we give theoretical results on inference under inequality constraints. The asymptotic consistency and normality of maximum likelihood estimators are established. The main methods throughout this manuscript are implemented in R language programming.They are applied to risk assessment problems in nuclear safety and coastal flooding, accounting for positivity and monotonicity constraints. As a by-product, we also show that the proposed GP approach provides an original framework for modelling Poisson processes with stochastic intensities.
Document type :
Complete list of metadata

Cited literature [153 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Wednesday, June 10, 2020 - 4:43:15 PM
Last modification on : Tuesday, January 5, 2021 - 10:21:17 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02863891, version 1


Andres Felipe Lopez Lopera. Gaussian Process Modelling under Inequality Constraints. General Mathematics [math.GM]. Université de Lyon, 2019. English. ⟨NNT : 2019LYSEM020⟩. ⟨tel-02863891⟩



Record views


Files downloads