A Comparison of Artificial Neural Network and Kriging Model for Predicting the Deterministic Output Response

Abstract

The aim of this paper is to compare the prediction accuracy between two popularapproximation model methods namely, artificial neural network (ANN) and Kriging modelfor modeling the output response from computer simulated experiments (CSE). The naturesof CSE are time-consuming and computationally expensive to run. Hence, many efforts havefocused on developing inexpensive and reliable surrogate models to replace the CSE.Kriging model along with Latin hypercube designs (LHD) have been widely usedfor developing an accurate surrogate model in the context of CSE. The performance ofKriging model is based on the estimation of the unknown parameters. The most popularmethod to estimate these parameters is the maximum likelihood estimate (MLE) method.The MLE method is normally time consuming and fails to obtain the best set of parametersdue to numerical instability and ill-conditioning of the model structure. Due to the popularityof ANN in modeling high and complex problem, this paper presents an application of ANNin the context of CSE and the comparison with Kriging model is employed. The resultsindicate that ANN performs well in terms of prediction accuracy and can be used as analternative of Kriging model in some features of problem under this study.Keywords: computer simulated experiments, artificial neural network, Kriging model,optimal Latin hypercube designs.