Collaborative Research: Short Cycle Surrogate Based Design Optimization

工程系统通常使用昂贵的计算机模拟其对环境输入的响应来设计。设计师通常会执行数十到数百次模拟，并构建一个近似的响应（称为代理），作为设计参数的函数。代理用于搜索候选最优设计，完成一个设计周期。为了提高精度，在接近候选最优的较小设计参数范围内重复模拟、替代构造和优化循环。每个周期可能需要许多天，因此通常只执行很少的设计周期。不幸的是，没有正式的方法来帮助设计师决定何时停止；这项建议的目标是发展这种方法。该建议的主要任务是使用统计分析来估计如果进行另一个周期，系统性能实质性改进的可能性。此外，我们建议使用多个替代替代物来建立这些统计估计的信心。给定一个概率，设计师可以决定是否值得在另一个设计周期中投入额外的时间和金钱。这种方法将适用于改善许多行业的设计过程，因为几乎每个行业都使用基于代理的设计优化。

Engineering systems are usually designed using costly computer simulations of their response to environmental input. Designer typically perform dozens to hundreds simulations and construct an approximation--called surrogate--to the response as function of design parameters. The surrogate is used to search for a candidate optimum design completing a cycle of design. For improved accuracy, the cycle of simulation, surrogate construction, and optimization is repeated in a smaller range of design parameters near the candidate optimum. Each cycle can require many days, and so typically only few design cycles are performed. Unfortunately, there is no formal methodology to help designers decide when to stop; the objective of the proposal is to develop such methodology. The main task of the proposal is to use statistical analysis to estimate the probability of substantial improvement in system performance if another cycle is undertaken. Also, we propose using multiple alternate surrogates for developing confidence in these statistical estimates. Given a probability, a designer can decide whether expected improvement in performance is worth the extra time and money of another design cycle. This approach will be applicable to improve the design process in a wide range of industries, since almost every industry makes use of surrogate based design optimization.