Neural-Network Modeling

EMPIRICAL METHODS are regarded as less desirable than those based on physical principles, primarily because the former are perceived to be limited to the knowledge base on which they are created and hence may not generalize well. The interpolation or extrapolation of regression equations is associated with unspecified uncertainty, even when the fit with the known domain is good. Yet, when dealing with complicated mechanical properties such as ductility, fatigue, or creep, physical models are less than useful in dealing with the complexities of technology and often are limited to making qualitative and simplistic inferences. Empirical modeling has, however, taken a turn for the better with the advent of neural networks, which permit the discovery of fundamental relationships and quantitative structure within vast arrays of ill-understood data. The significant factor in this success has been in the understanding of the rules for creating robust models and in the treatment of noise and uncertainties. It is now well established that the method is not only capable of representing known data but can also lead to the discovery of novel concepts (Ref 1 to 3). To begin, the method is introduced.