Learning-guided exploration in airfoil optimization

Abstract

A learning-based exploration approach is proposed to escape from the basins of attraction of converged-to optima, by selecting on what is termed the interestingness of a solution. This interestingness is based on the modeling error made by a surrogate model that is trained on all solutions encountered earlier during the search. Compared to multiple standard optimization runs, a learning-guided restart scheme that alternates between a quality optimization phase and an exploration phase directed by interestingness finds solutions that are more diverse and of higher quality. © 2013 Springer-Verlag.