Constructing prediction intervals for neural network metamodels of complex systems

A rich literature discussing techniques for adopting neural networks for metamodelling of complex systems exists. The main focus in many studies conducted so far has been on training and utilising neural networks as point estimators/predictors. Uncertainties prevailing within complex systems and dependencies amongst constituent entities are real threats for prediction performance of these types of metamodels. From a practical point of view, an indication of prediction accuracy is necessary before making a decision based on results yielded by a metamodel. In this paper we adopt neural network metamodels for constructing prediction intervals of stochastic system performance measures. Upper and lower bounds of a prediction interval are computed such that the real system performance will lie between them with a high probability. Demonstrated results for a real world case study show that the constructed prediction intervals cover the targets, are more informative and more suited for decision making, when compared with point predictions.