Empirical comparison of gradient-based methods on an engine-inlet shape optimization problem

With the development of increasingly sophisticated adjoint flow-solvers capable of providing objective function gradients at reasonable computational costs, modern deterministic gradient-based search methods have come to be regarded as amongst the most powerful tools in aerodynamic shape optimization and MDO problems. However, their performance can be disappointing when the objective function landscape features multiple local optima, long valleys, noise or discontinuities. Equally, stochastic global explorers, such as Genetic Algorithms (GAs), while less affected by these problems, are relatively slow to converge. In this paper we propose GLOSSY (Global/Local Search Strategy), a generic hybrid approach, which combines a global exploration method with gradient-based exploitation. We analyze the performance of two optimizers based on the GLOSSY framework (fusing a GA with a quasi-Newton local search method) and we show through a set of comparative tests that on the moderately noisy objective landscape of a jet-engine inlet shape optimization problem the hybrid outperforms both of its components used individually. We also look at the issue of what global / local search effort ratio gives the hybrid the best performance