利用神经网络实现复杂结构的多目标优化设计

结构优化设计中常常包含大量的有限元计算。现代多目标优化设计的发展趋势是以 Pareto遗传算法为代表的随机搜索方法 ,能够搜索到整个 Pareto最优解集 ,但计算量相当大 ,如果每次迭代都要涉及有限元计算 ,将是非常耗时的工作。本文在利用 Kolm ogorov多层神经网络映射存在定理的基础上导出的用神经网络进行结构近似分析的方法 ,用均匀试验设计方法选取特征样本点供神经网络训练 ,将神经网络与 Pareto遗传算法有机地结合 ,使多目标优化的计算效率进一步提高

Multiobjective Optimization of Complex Structure Using Neural Networks

There is large amount of calculation with FEM in complex structure optimal design. The tendency of multobjecitve optimization(MOP) is making full use of evolutionary algorithm(EA), but the cost of time which combines FEM with EA for MOP is very high. We present a flexible technique to optimize complex structure, based on the combined use of radial basis functions, which provide an analytical approximation of the true FEM calculation, and the use of even design of experiments to select typical sample points. The in detail technique is described, with an example to prove its efficiency.