A Case Study on Target Cost Estimation Using Back-Propagation and Genetic Algorithm Trained Neural Networks

Cost estimation of new products has always been difficult as only few attributes will be known. In these situations, parametric methods are commonly used using a priori determined cost function where parameters are evaluated from historical data. Neural networks, in contrast, are non-parametric, i.e., they attempt to fit curves without being provided a predetermined function. In this article, this property of neural networks is used to investigate their applicability for cost estimation of certain major aircraft subassemblies. The study is conducted in collaboration with an aerospace company located in Montreal, Canada. Two neural network models, one trained by the gradient descent algorithm and the other by genetic algorithm, are considered and compared with one another. The study, using historical data, shows an example for which the neural network model trained by genetic algorithm is robust and fits well both the training and validation data sets.