Design of a surrogate model assisted (1+1)-ES

Abstract

Surrogate models are employed in evolutionary algorithms to replace expensive objective function evaluations with cheaper though usually inaccurate estimates based on information gained in past iterations. Implications of the trade-off between computational savings on the one hand and potentially poor steps due to the inaccurate assessment of candidate solutions on the other are generally not well understood. We study the trade-off in the context of a surrogate model assisted (1 + 1) -ES by considering a simple model for single steps. Based on the insights gained, we propose a step size adaptation mechanism for the strategy and experimentally evaluate it using several test functions.