Design of stiffened plates using soft computing techniques

This thesis deals with development of reliable, accurate and efficient soft computing tools for the analysis and design of structures. The efficiency of soft computing techniques in structural design such as, particular algorithms based on genetic algorithm, neural network, gene expression programming and large?scale continuous or discrete structural design problems are studied. The algorithms are studied both in deterministic and reliability based structural design problems. To increase the computational efficiency as well as the robustness of the design procedure, an effort is put forth. Structural optimization process requires the efficient integration of computer assisted geometry modeling, automated mesh generation, structural analysis and soft computing applications. The use of soft computing techniques is motivated from the time?consuming repeated finite strip analysis required during the optimization process. A trained soft computing technique is used to perform deterministic constraints check in the case of reliability based design. The suitability of the soft computing techniques predictions will be investigated in a number of structural design problems in order to demonstrate the computational advantages of the proposed methodologies.Principally, it is desired that this thesis will provide stable bases for further investigation, leading to a more intensive use of structural optimization algorithms with soft computing techniques to solve practical problems. Because of the broad diversity of structures encountered in practice, it becomes clear that this thesis is concentrated on buckling and free vibration analyses of stiffened plates.All soft computing models used in this study are presented in explicit form neural network and gene expression programming models. The most accurate results are obtained by neural network model rather than gene expression programming. Neural networks are treated as black box in general. It should be noted that explicit formulation of neural network models is of significant importance as it will serve for important advantages in the analysis and design of structures. This thesis aims to open the black box and to present the neural network models in its explicit form. An alternative algorithm for the selection of optimum neural network architecture that automatically selects the best architecture is proposed. By using finite strip method large testing and training sets are constructed and high generalization capabilities of the models are obtained.