Conventional numerical optimization tools use the gradient information. So, their convergence processes seem to be consistent. However, it may converge to different solutions with the different initial design. Especially, it is nearly failed when the function is noisy or non-smooth.
AutoDesign uses a sequential approximate optimization based on meta-models. Thus, it can overcome the noisy or non-smoothness of performance indexes (PIs). Nevertheless, its’ convergence may be different according to the initial DOE methods or meta-model methods. Empirically, in AutoDesign, this phenomenon occurs when the performance indexes are noisy or discontinuous.
However, the meta-model based optimization is less sensitive to local noisy and discontinuity than the gradient-based optimization algorithm. As mentioned in the guide of initial sampling, AutoDesign have possibility to converge a global optimum nearly if the initial sampling points cover design space properly and the performance indexes are smooth. So, it is very important in dynamic response optimization that one formulates his design problem insensitive to noise.