Assessing the suitability of surrogate models in evolutionary optimization

The paper deals with the application of evolutionary algorithms to black-box optimization, frequently encountered in biology, chemistry and engineering. In those areas, however, the evaluation of the black-box fitness is often costly and time-consuming. Such a situation is usually tackled by evaluating the original fitness only sometimes, and evaluating its appropriate response-surface model otherwise, called surrogate model of the fitness. Several kinds of models have been successful in surrogate modelling, and a variety of models of each kind can be obtained through parametrization. Therefore, real-world applications of surrogate modelling entail the problem of assessing the suitability of different models for the optimization task being solved. The present paper attempts to systematically investigate this problem. It surveys available methods to assess model suitability and reports the incorporation of several such methods in our recently proposed approach to surrogate modelling based on radial basis function networks. In addition to the commonly used global suitability of a model, it pays much attention also to its local suitability for a given input. Finally, it shows some results of testing several of the surveyed methods in two real-world applications.