Structural Optimum Design Based On Computing Intelligence

It has been very important to find very robust and very efficient algorithms for solving optimization problems at structural optimum design. At the same time, structural analysis occupies an important place in the structural optimum design. But the conventional structural approximate analysis methods based on the first-order or the second-order derivatives of the stresses and the displacements of structures are not global, while an efficient global structural approximate analysis methods is helpful to decrease the number of the times of structural analysis, so the amount of calculation of structural analysis in the structural optimum design is reduced. Under this background, this paper has developed the researches on the structural optimum design based on Genetic Algorithms and on the structural approximate analysis based on artificial neural network. The main work of this paper is as the following:1. Two new GAs-determined GA (suitable for optimum problems with few variables) and half-determined GA (suitable for optimum problems with infinite variables)?are proposed in this paper. The ability to search the neighbor-area of the best solution of half-determined GA is improved with three new strategies: Half determined Hamming-decreasing, Low bit determined mutation and Self-adaptive scaling of design variables. Half-determined GA takes the advantages on the global-searching of genetic operators based on probabilities and the advantages on the local-searching of determined and half-determined genetic operators, therefore, this algorithm is very robust and very fast.2. A new structural optimization algorithm combined with the Half-determined GA and the hybrid approximation techniques is proposed. Numeric test on some classic structural optimum problems shows that this algorithm is feasible.3. Constructive research on the units of the hidden-layer and the corresponding error estimation of Gaussian RBF neural network has been made. This work provides the preliminary theory basis for improving the precise of approximation with RBF NN and reducing the amount of calculation.4. The mechanical background for structural approximate analysis with RBF NN has been researched.5. A new learning algorithm based on the half-determined GA for RBF NN is proposed, which considers the number of the units in the hidden-layer the centers % the width and the weight of RBF NN. Theoretically, the best RBF NN (the number of the units in the hidden-layer is the least and the precise is the highest) can be get with thisIValgorithm.6. A new model of structural approximate analysis based on RBF NN is proposed on the level of global simulation. The numeric tests show that RBF NN has very good calculate characters and it is more suitable for structural approximate analysis than other feed-forward NNs, such as BP NN.