A Comparative Study on Artificial Neural Network and Radial Basis Function for Modelling Output Response from Computer Simulated Experiments

Abstract

Computer simulated experiments (CSE) have been widely used to investigate complex physical phenomena, particularly when physical experiments are not feasible due to limitations of experimental materials. The natures of CSE are time-consuming and the computer codes are expensive. Therefore, experimental designs and statistical models approaches play a major role in the context of CSE in order to develop the approximation model for use as a surrogate model. Many researchers have attempted to develop various predictive models to fit the output responses from CSE. The purpose of this paper is to compare the prediction accuracy of three models namely Kriging model (KRG), Radial basis function (RBF) model and Artificial neural network (ANN) model, respectively. These three models are constructed by using the optimal Latin hypercube designs (OLHD). The prediction accuracy of each model is validated though non-linear test problems ranging from 2 to 8 input variables and evaluated by the root mean squared of error (RMSE) values. The results show that RBF model performs well when small dimension of problem with small design run is considered while KRG model is the most accurate model when the design run is large. For larger dimensions of problem, KRG model is suitable for small design runs while ANN model performs superior over the other models when the design runs are large.