Evolutionary algorithm characterization in real parameter optimization problems

This paper deals with the constant problem of establishing a usable and reliable evolutionary algorithm (EA) characterization procedure so that final users like engineers, mathematicians or physicists can have more specific information to choose the most suitable EA for a given problem. The practical goal behind this work is to provide insights into relevant features of fitness landscapes and their relationship to the performance of different algorithms. This should help users to minimize the typical initial stage in which they apply a well-known EA, or a modified version of it, to the functions they want to optimize without really taking into account its suitability to the particular features of the problem. This trial and error procedure is usually due to a lack of objective and detailed characterizations of the algorithms in the literature in terms of the types of functions or landscape characteristics they are well suited to handle and, more importantly, the types for which they are not appropriate. Specifically, the influence of separability and modality of the fitness landscapes on the behaviour of EAs is analysed in depth to conclude that the typical binary classification of the target functions into separable/non-separable and unimodal/multimodal is too general, and characterizing the EAs' response in these terms is misleading. Consequently, more detailed features of the fitness landscape in terms of separability and modality are proposed here and their relevance in the EAs' behaviour is shown through experimentation using standardized benchmark functions that are described using those features. Three different EAs, the genetic algorithm, the Covariance Matrix Adaptation Evolution Strategy and Differential Evolution, are evaluated over these benchmarks and their behaviour is explained in terms of the proposed features.