Revealing the interlevel dependence structure of categorical inputs in numerical environmental simulations with kernel model selection. (May 2022)

• Record Type: Journal Article
• Authors: Rohmer, Jeremy; Roustant, Olivier; Lecacheux, Sophie; Manceau, Jean-Charles
• Is Part Of: Environmental modelling & software. Volume 151(2022)
• Publication Date: 2022-05
• Subjects: Categorical variables -- Computationally intensive simulator -- Metamodel -- Kriging -- Model selection; CS Compound symmetry -- DT Decision tree -- E expert-based -- Gen general -- GP Gaussian process -- LR low rank -- O ordinal -- RF random forest; Environmental monitoring -- Computer programs -- Periodicals; Ecology -- Computer simulation -- Periodicals; Digital computer simulation -- Periodicals; Computer software -- Periodicals; Environmental Monitoring -- Periodicals; Computer Simulation -- Periodicals; Environnement -- Surveillance -- Logiciels -- Périodiques; Écologie -- Simulation, Méthodes de -- Périodiques; Simulation par ordinateur -- Périodiques; Logiciels -- Périodiques; Computer software; Digital computer simulation; Ecology -- Computer simulation; Environmental monitoring -- Computer programs; Periodicals; Electronic journals; 363.70015118
• DOI: 10.1016/j.envsoft.2022.105380
• Languages: English
• ISSNs: 1364-8152

Abstract

Abstract: Model uncertainties are generally integrated in environmental long-running numerical simulators via a categorical variable. By focusing on Gaussian process (GP) models, we show how different categorical kernel models (exchangeable, ordinal, group, etc.) can bring valuable insights into the correlation of the simulator output values computed for different levels of the categorical variable, i.e., the interlevel dependence structure. Supported by two real case applications (cyclone-induced waves and reservoir modeling), we have proposed a cross-validation approach to select the most appropriate kernel by finding a trade-off between predictability, explainability, and stability of the covariance coefficients. This approach can be used effectively to support some physical assumptions regarding the categorical variable. Through comparison to tree-based techniques, we show that GP models can be considered a satisfactory compromise when only a few model runs (∼100) are available by presenting a high predictability and a concise and graphical way to map the interlevel dependence structure. Highlights: Explore influence of categorical inputs with Gaussian process regression. Multicriterion selection of the kernel covariance function for categorical inputs. Applications to real cases (cyclone waves and CO2 reservoir modelling). GP correlation matrix provides a concise and graphical tool to explore dependencies. High predictability compared to tree-based methods.