Decision support with neural networks in the management of research and development: Concepts and application to cost estimation

Despite the small number of applications to date, neural networks are likely to be able to contribute to decision support in selected fields of R&D management. We identify the potential of neural networks in the application domain and compare it to 'classical' applications, such as the recognition of hand-written characters. Typical neural network architectures for R&D management tend to be simple, having law complexity, and only a small number of training samples are generally available. As an example, we carry out experiments for a typical R&D management application where neural networks have to estimate the final cost of a new product under development. It turns out that neural networks based on the standard backpropagation learning algorithm perform reasonably well when the ratio between highest and lowest cost is small, even for relatively small training set sizes. Otherwise the learning algorithm tends to undervalue low cost levels, so that deviations between estimated cost and real cost are intolerably high. Future research will have to investigate a modification of the error definition of the backpropagation algorithm. Finally, a number of general statements are derived from our experience, and examples are provided where neural networks are appropriate or inappropriate in the domain of R&D management. (C) 1998 Elsevier Science B.V. All rights reserved.