Developing optimal neural network metamodels based on prediction intervals

Finding optimal structures for neural networks is remains an open problem, despite the rich array of literature on the application of neural networks in different areas of science and engineering. The stochastic nature of operations common in complex systems makes point prediction performance of neural network metamodels an additional challenge. We propose a method for selecting the best structure of a neural network metamodel. For selecting the network structure, the new method uses interval prediction capability of neural networks and chooses a topology that yields the narrowest prediction band for targets. This is an improvement on traditional criteria, such as mean square error or mean absolute percentage error. As a case study, the interval prediction method is applied to a metamodel of a complex system composed of many inextricably interconnected entities and stochastic processes. The demonstrated results expressly show that selecting the network structure based on the proposed method yields more reliable estimates.