Neural networks for cost estimation: simulations and pilot application

Neural networks in a multilayer perceptron architecture are able to classify data and approximate functions based on a set of sample data (curve fitting). These properties are used to investigate experimentally the applicability of neural networks for cost estimation in early phases of product design. Experiments are based on pilot cost data from a manufacturing company. In addition, artificially created simulative data are used for benchmarking. The cost estimation performance is compared to conventional methods, i.e. linear and non-linear parametric regression. Neural networks achieve lower deviations in their cost estimations. Beyond the use of standard neural architectures, simple modifications for a performance improvement are suggested and tested. Finally, a profile for situations where neural networks are appropriate is derived from the results.