Recent progress in simulation-driven optimization

Real-world engineering problems are often simulation-driven and involve a large number of design variables, a combination which exacerbates the optimization difficulty. Often metamodels would be used to approximate the simulation, but in high dimensional settings their approximation quality may be poor. To address this challenge, this study proposes a metamodel-assisted framework which adds a dimensionalityreduction component based on variable selection. The latter is used to select a subset of variables and to formulate a sequence of lower-dimensional optimization problems. Modelling and optimization are then performed on these reduced-dimensionality problems resulting in more accurate metamodels and consequently a more effective optimization search. Extensive performance analysis with high dimensional problems (several hundreds of variables) shows the effectiveness of the proposed approach.