An analog computing approach to structural analysis and optimal design through artificial neural networks

There has been significant research activity in the general area of neural computing, or more specifically, artificial neural networks. This thesis investigates the applicability of distinctly different architectures of neural networks in structural analysis and optimal design. Specific topics investigated in this work include: (1) Adaptive Resonance Theory (ART) based neural networks in automated conceptual design of structural systems; (2) Hopfield neural networks and a variant of this network referred to as the deformable template model in the solution of combinatorial optimization problems; (3) Development of a measure referred to as generalized cross validation for determining the quality of a function approximation obtained through the use of regularization neural networks; and (4) The use of regularization networks in structural damage monitoring. The results of these studies clearly demonstrate the usefulness of the neural network based computing paradigm in problems of structural engineering and design. In addition to this demonstrated potential in providing acceptable solutions to genetically difficult optimization problems, neural networks offer a significant function approximation capability for reducing the computational effort in an iterative design environment. The approach presented in this thesis provides a rational approach for controlling the quality of function approximations from the regularization network. Another important contribution of the present work is to show how the pattern recognition feature of neural networks, particularly the ART network, can be used to create the framework for a conceptual design system, in a manner similar to that available in the rule-based expert systems.