A study of surrogate models for their use in multiobjective evolutionary algorithms

Evolutionary Algorithms (EAs) are bioinspired meta-heuristics that have been successfully used to solve multiobjective optimization problems (MOPs). However, when EAs need to perform several objective function evaluations in order to reach a subobtimal solution and each of these evaluations are computationally expensive, then, these problems can remain intractable even by these meta-heuristics. Therefore, it is necessary to employ an additional strategy in order to reduce the response time of EAs when optimizing these expensive problems. Replacing the original problem with a surrogate model has been an usual strategy for time reduction. However, despite its success, few comparison among surrogate models for multiobjective optimization problems have been reported in the specialized literature. In this paper, we compare four meta-modeling techniques: Radial Basis Functions, Support Vector Regression, Polynomial Regression and Kriging-DACE in different aspects such as accuracy, robustness, efficiency, and scalability with the aim to identify advantages and drawbacks of each meta-modeling technique in order to choose the most suitable one to be combined with multiobjective evolutionary algorithms.