Global Sensitivity Analysis Methodology for Construction Simulation Models: Multiple Linear Regressions versus Multilayer Perceptions

Abstract

In this research, the multilayer perceptron (MLP), also known as error backpropagation neural networks, is made transparent and explainable by contrasting with the commonly applied multiple linear regression (MLR). A novel MLP-based method for performing global sensitivity analysis is formalized to tackle complicated, nonexplainable simulation models or artificial intelligence (AI) models, which were developed to support critical decisions in construction engineering. The sensitivity analysis results serve as further evidence to validate the decision support models and lend new insights into the problems under investigation. The proposed new method was applied in two case studies in construction engineering, they are: precast viaduct installation cycles and concrete strength development. In both applications, the results of sensitivity analysis were represented in straightforward forms and effectively cross-checked with the existing knowledge of the problem domain or the experiences of construction practitioners.