Keywords
Sensitivity analysis, Uncertainty analysis, High-dimensional environmental models, Grouping algorithm
Start Date
27-6-2018 10:40 AM
End Date
27-6-2018 12:00 PM
Abstract
Global sensitivity analysis (GSA) has proven useful and necessary in parameterization, calibration, and uncertainty analysis of the advanced Earth and Environmental Systems Models (EESMs) that are nowadays essential tools in planning and decision making under uncertainty and non-stationarity. However, the EESMs typically involve many input factors (resulting in high-dimensional response surfaces), which are seldom known to high precision, leading to a huge leap in computational cost when performing GSA (a manifestation of curse of dimensionality). This issue precludes effective implementation of the current-generation GSA techniques because a comprehensive sensitivity analysis usually requires a prohibitively large number of model runs.
To overcome this challenge, we develop a GSA methodology enabled with an automated “factor grouping” strategy that is based on a bootstrap-based clustering analysis. Our proposed method is designed to robustly categorize input factors into a certain number of groups of different sizes using information gained throughout the GSA. The algorithm utilizes two efficient methods to determine the optimal number of groups. Furthermore, to monitor convergence of the GSA, we introduce a measure of reliability based on factor grouping, which facilitates efficiently performing GSA with a limited number of model runs. Here, we demonstrate the performance of this methodology with a variogram-based GSA algorithm, known as Variogram Analysis of Response Surfaces (VARS), to conduct a sensitivity analysis for two high-dimensional complex problems.
Addressing Curse of Dimensionality in Global Sensitivity Analysis of Large Environmental Models: An Automated Grouping Strategy
Global sensitivity analysis (GSA) has proven useful and necessary in parameterization, calibration, and uncertainty analysis of the advanced Earth and Environmental Systems Models (EESMs) that are nowadays essential tools in planning and decision making under uncertainty and non-stationarity. However, the EESMs typically involve many input factors (resulting in high-dimensional response surfaces), which are seldom known to high precision, leading to a huge leap in computational cost when performing GSA (a manifestation of curse of dimensionality). This issue precludes effective implementation of the current-generation GSA techniques because a comprehensive sensitivity analysis usually requires a prohibitively large number of model runs.
To overcome this challenge, we develop a GSA methodology enabled with an automated “factor grouping” strategy that is based on a bootstrap-based clustering analysis. Our proposed method is designed to robustly categorize input factors into a certain number of groups of different sizes using information gained throughout the GSA. The algorithm utilizes two efficient methods to determine the optimal number of groups. Furthermore, to monitor convergence of the GSA, we introduce a measure of reliability based on factor grouping, which facilitates efficiently performing GSA with a limited number of model runs. Here, we demonstrate the performance of this methodology with a variogram-based GSA algorithm, known as Variogram Analysis of Response Surfaces (VARS), to conduct a sensitivity analysis for two high-dimensional complex problems.
Stream and Session
Stream E: Modeling for Planetary Health and Environmental Sustainability
E3: Complexity, Sensitivity, and Uncertainty Issues in Integrated Environmental Models