Automatic surrogate modelling technique selection based on features of optimization problems

A typical scenario when solving industrial single or multiobjective optimization problems is that no explicit formulation of the problem is available. Instead, a dataset containing vectors of decision variables together with their objective function value(s) is given and a surrogate model (or metamodel) is build from the data and used for optimization and decision-making. This data-driven optimization process strongly depends on the ability of the surrogate model to predict the objective value of decision variables not present in the original dataset. Therefore, the choice of surrogate modelling technique is crucial. While many surrogate modelling techniques have been discussed in the literature, there is no standard procedure that will select the best technique for a given problem. In this work, we propose the automatic selection of a surrogate modelling technique based on exploratory landscape features of the optimization problem that underlies the given dataset. The overall idea is to learn offline from a large pool of benchmark problems, on which we can evaluate a large number of surrogate modelling techniques. When given a new dataset, features are used to select the most appropriate surrogate modelling technique. The preliminary experiments reported here suggest that the proposed automatic selector is able to identify high-accuracy surrogate models as long as an appropriate classifier is used for selection.