An analysis of the impact of the initial sample on evolutionary metamodel-assisted optimization

The modern engineering design process often employs computer simulations to evaluate candidate designs, a setup that results in computationally expensive black-box optimization problems. An established framework to solve such problems is with evolutionary metamodel-assisted algorithms, in which the metamodel provides the evolutionary algorithm with approximate function values at a lower computational cost when compared with the simulation. Such evolutionary optimizers require an initial sample of points, which are then used to train an initial metamodel and to enable the main optimization search. This setup implies that the initial sample can significantly impact the effectiveness of the optimization search. Two main approaches for generating the initial sample are statistical sampling, in which the sample points are generated by sampling a statistical distribution, and the more recent search-based sampling, which uses a population-based algorithm to generate the sample points. Leveraging on the importance of the initial sample, this study presents a detailed comparison of methods for generating the initial sample, and analyzes in detail their impact on the evolutionary metamodel-assisted search in a variety of optimization scenarios. It concludes with a set of recommendations for selection of the sampling method.