PCA-ENHANCED METAMODEL-ASSISTED EVOLUTIONARY ALGORITHMS FOR AERODYNAMIC OPTIMIZATION

This paper deals with evolutionary algorithms (EAs) assisted by surrogate evaluation models or metamodels (Metamodel-Assisted EAs, MAEAs) which are further accelerated by exploiting the Principal Component Analysis (PCA) of the elite members of the evolving population. PCA is used to (a) guide the application of evolution operators and (b) train metamodels, in the form of radial basis functions networks, on patterns of smaller dimension. Compared to previous works by the same authors, this paper also proposes a new way to apply the PCA technique. In particular, the front of non-dominated solutions is divided into sub-fronts and the PCA is applied "locally" to each sub-front. The proposed method is demonstrated in multi-objective, constrained, aerodynamic optimization problems.