An analysis of two-population coevolutionary computation

Coevolutionary computation (CoEC) is the subfield of evolutionary computation (EC) centered around the notion of interaction among simultaneously evolving entities. While it promises important problem-solving advantages, coevolution also brings many challenges. A practitioner trying to use a coevolutionary algorithm (CoEA) to solve a problem is generally interested in how the choices made in designing the algorithm affect its performance on that particular problem. In other words, they would like to have information about the dependency: problem properties + algorithm properties → performance. For traditional evolutionary algorithms (EAs), our understanding of this dependency has reached reasonably satisfactory levels. For CoEAs it has not, as they have proven notoriously more complex and less intuitive. The main contribution of my dissertation is advancing the understanding of this dependency for traditional two-population coevolutionary algorithms. The way I achieve this is through extensive analysis that connects algorithm, problem and performance through one key aspect: dynamics. While the importance of understanding the dynamics of coevolutionary systems has been pointed out by previous research, this dissertation is the first study that "glues" all four pieces together. Additionally, an important feature of the analysis is that it spans subareas of CoEC that were previously studied independently (compositional cooperative and test-based competitive). It bridges them by identifying a problem property and introducing tools for analyzing this property that are applicable across subareas, thus providing a more holistic perspective of the field of CoEC. The analysis is performed both for previously unstudied aspects of CoEAs and for ones that have been investigated using other techniques. The power of the new analysis approach is particularly visible in the latter case, where it is shown to explain prior "mysterious" results.