An analysis of evolutionary-based sampling methodologies

A common approach for solving simulation-driven engineering problems is by using metamodel-assisted optimization algorithms, namely, in which a metamodel approximates the computationally expensive simulation and provides predicted values at a lower computational cost. Such algorithms typically generate an initial sample of solutions which are then used to train a preliminary metamodel and to initiate optimization process. One approach for generating the initial sample is with the design of experiment methods which are statistically oriented, while the more recent search-driven sampling approach invokes a computational intelligence optimizer such as an evolutionary algorithm, and then uses the vectors it generated as the initial sample. Since the initial sample can strongly impact the effectiveness of the optimization process, this study presents an extensive comparison and analysis between the two approaches across a variety of settings. Results show that evolutionary-based sampling performed well when the size of the initial sample was large as this enabled a more extended and consequently a more effective evolutionary search. When the initial sample was small the design of experiments methods typically performed better since they distributed the vectors more effectively in the search space.