Iterative RBF neural networks as metamodels of stochastic simulations

Abstract

Research into emerging technological approaches to make computer simulations more effective and efficient is an essential ingredient to developing successful manufacturing models. This study is a premiere study in using neural networks in metamodeling stochastic simulation in manufacturing domain. A new iterative RBF neural network was developed rather than the baseline ANN models which were used in stochastic simulation metamodeling in domains such as combat simulations in the military, service industries, and transportation companies. Given the fact that typical stochastic simulation metamodeling approaches involves the use of regression models in response surface methods, RBF become a natural target for such an attempt because they use a family of surfaces each of which naturally divides an input space into 2 regions and the n patterns will be assigned either class X+ or X-. This dichotomy of the points is said to be separable with respect to the family of surfaces if there exists a surface in the family that separates the points in the class X+ from those in the class X-. In fact, for the evaluation of the quality of a ball steel, RBF metamodel trained on 1521 training examples from a set of 13000 different simulation runs and was able to outperform direct simulation on 120 additional test examples which were not included in the training set.