Landscape analysis of gaussian process surrogates for the covariance matrix adaptation evolution strategy

Abstract

Gaussian processes modeling technique has been shown as a valuable surrogate model for the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) in continuous single-objective black-box optimization tasks, where the optimized function is expensive. In this paper, we investigate how different Gaussian process settings influence the error between the predicted and genuine population ordering in connection with features representing the fitness landscape. Apart from using features for landscape analysis known from the literature, we propose a new set of features based on CMA-ES state variables. We perform the landscape analysis of a large set of data generated using runs of a surrogate-assisted version of the CMA-ES on the noiseless part of the Comparing Continuous Optimisers benchmark function testbed.