Effect of Representation on the Performance of Neural Networks in Structural Engineering Applications
David J. Gunaratnam, John S. Gero · Computer-Aided Civil and Infrastructure Engineering · 1994
Abstract: The pattern-mapping, pattern-classification, and optimization capabilities of neural networks have been used to solve a number of structural analysis and design problems. Most applications exploit the pattern-mapping capability and are based on the back-propagation paradigm for neural networks. There are a number of factors that influence the performance of these networks. This paper initially discusses these factors and the domain-dependent and -independent techniques presently available for improving performance. The paper then considers the effect of representation, selected for the input/output pattern pairs, on the performance of these networks and demonstrates that representations based on dimensionless terms, derived from dimensional analysis, lead to improved performance. It is shown that dimensional analysis provides a representational framework, with reduced dimensionality and embedded domain knowledge, within which effective learning can take place and that this representational change can be used to enhance the domain-independent and -dependent techniques presently available for improving performance of these networks.