Optimal sampling strategies for learning a fitness model

Abstract

The paper investigates the use of kriging interpolation and estimation as a function approximation tool for the optimization of computationally complex functions. A model of the fitness function is built from a small number of samples of this function. This model is utilized in a model based learning strategy as an auxiliary fitness function. The kriging approach represents a compromise between global models and local models. The model is initially a global approximation of the entire domain, and successive updates during the optimization process transform it into a more precise local approximation. Several approaches for the sampling of the true fitness function are investigated in order to update a fitness model efficiently and at a low computational cost.