Model-based optimization revisited: Towards real-world processes

The application of empirically determined surrogate models provides a standard solution to expensive optimization problems. Over the last decades several variants based on DACE (design and analysis of computer experiments) have provided excellent optimization results in cases where only a few evaluations could be made. In this paper these approaches are revisited with respect to their applicability in the optimization of production processes, which are in general multiobjective and allow no exact evaluations. The comparison to standard methods of experimental design shows significant improvements with respect to prediction quality and accuracy in detecting the optimum even if the experimental outcomes are highly distorted by noise. The universally assumed sensitivity of DACE models to nondeterministic data can therefore be refuted. Additionally, a practical example points out the potential of applying EC-methods to production processes by means of these models.