Process cost modelling using neural networks

One method of establishing cost-estimating relationships (CERs) that appears to offer advantages is that of neural networks as they are ‘model-free estimators’, hence providing key advantages over traditional techniques, e.g. regression analysis. However, the complexity of neural network (NN) structures and the availability of a wide range of structural alternatives may make it difficult to select an appropriate network for a specific cost-modelling requirement. The current work is aimed at resolving this issue by examining the use of the Taguchi methodology for identifying NN structural elements. The Taguchi methodology has been used to identify ‘best’ and ‘poorest’ NN structures and the models developed using these structures compared with regression-based models in terms of their estimating accuracy. The robustness of these models is then examined in terms of the effects on their estimating accuracy of varying the number of data points used in their development and varying the number of predictor variables within the models.