Feature extraction for surrogate models in genetic programming

Abstract

We discuss the use of surrogate models in the field of genetic programming. We describe a set of features extracted from each tree and use it to train a model of the fitness function. The results indicate that such a model can be used to predict the fitness of new individuals without the need to evaluate them. In a series of experiments, we show how surrogate modeling is able to reduce the number of fitness evaluations needed in genetic programming, and we discuss how the use of surrogate models affects the exploration and convergence of genetic programming algorithms.