Why advanced population initialization techniques perform poorly in high dimension?

Abstract

Many advanced population initialization techniques for Evolutionary Algorithms (EAs) have hitherto been proposed. Several studies claimed that the techniques significantly improve EAs’ performance. However, recent researches show that they cannot scale well to high dimensional spaces. This study investigates the reasons behind the failure of advanced population initialization techniques in large-scale problems by adopting a wide range of population sizes. To avoid being biased to any particular EA model or problem set, this study employs general purpose tools in the experiments. Our investigations show that, in spite of population size, uniformity of populations drops dramatically when dimensionality grows. The observation confirms that the uniformity loss exist in high dimensional spaces regardless of the type of EA, initializer or problem. Therefore, we conclude that the weak uniformity of the resulting population is the main cause of the poor performance of advanced initializers in high dimensions.