中文摘要：

针对传统的计算机试验优化设计计算费时的不足,一方面改进了原始的随机进化寻优算法,以兼顾深度搜索与广度搜索;另一方面提出了分步试验优化设计的思路,逐步分批地布置试验样本点,化整为零地减小搜索空间;最后采用径向基神经网络替代模型,以均匀试验设计为例,检验分步试验优化设计方法的有效性.计算实践表明,改进算法可平均节省50%左右的机时,寻找到与原始算法结果相差约1%的最优值;分步试验设计较一步试验设计可减少40%-60%左右的机时,数值实验表明,随着试验样本点数量的递增,由不同试验设计所产生的替代模型的误差将趋于一致,分步试验设计尤其适合基于大规模试验样本点的替代模型。

英文摘要：

Considering the high computational cost of traditional optimal design of computer experiments, the original stochastic evolutionary algorithm was improved by balancing the depth searching and broadness searching. And optimal experiment design by steps was provided, which divided full searching space into subspaces and arranged samplings in batches. The influences of experiment design by steps on RBF network metamodel were also measured by numerical tests. The training sampling set of RBF network wss decided upon by uniform design. The computer practice shows that the improved algorithm can decrease about 50 % time cost at an average to find optimal values, which are about 1% different from what the original algorithm finds. The theory analyses show that the complements of experiment design by steps can decrease about 40% -60% machine time,compared with the traditional experiment design, which is confirmed by computer examples. Numerical tests show that, with the increment of samplings, the errors of metamodel resulted from different experiment designs will be consistent. Numerical tests also show that experiment design by steps will be suitable for the metamodels that need largescale experiment samplings.