Data driven surrogate-based optimization in the problem solving environment WBCSim

摘要: Large scale, multidisciplinary, engineering designs are always difficult due to the complexity and dimensionality of these problems. Direct coupling between analysis codes optimization routines can be prohibitively time consuming underlying simulation codes. One way tackling this problem is by constructing computationally cheap(er) approximations expensive simulations that mimic behavior model as closely possible. This paper presents a data driven, surrogate-based algorithm uses trust region-based sequential approximate (SAO) framework statistical sampling approach based on design experiment (DOE) arrays. The implemented using techniques from two packages—SURFPACK SHEPPACK provide collection approximation algorithms build surrogates three different DOE techniques—full factorial (FF), Latin hypercube sampling, central composite design—are used train surrogates. results compared with obtained directly an optimizer code. biggest concern in SAO generation required database. As number variables grows, computational cost generating database grows rapidly. A driven proposed tackle situation, where trick run if only nearby point does not exist cumulatively growing Over matures enriched more optimizations performed. Results show methodology dramatically reduces total calls runs during process.