Identifying Important Observations Using Cross Validation and Computationally Frugal Sensitivity Analysis Methods

Abstract

Sensitivity analysis methods are used to identify measurements most likely to provide important information for model development and predictions. Methods range from computationally demanding Monte Carlo and cross-validation methods that require thousands to millions of model runs, to very computationally efficient linear methods able to account for interrelations between parameters that involve tens to hundreds of runs. Some argue that because linear methods neglect the effects of model nonlinearity, they are not worth considering. However, when faced with computationally demanding models needed to simulate, for example, climate change, the chance of obtaining insights with so few model runs is tempting. This work compares results for a nonlinear groundwater model using computationally demanding cross-validation and computationally efficient local sensitivity analysis methods.