Parametrizing Convection Selection: Conclusions from the Analysis of Performance in the NKq Model

Convection selection in evolutionary algorithms is a method of splitting the population into subpopulations based on the fitness values of solutions. Convection selection was previously found to be superior to standard selection techniques in difficult tasks of evolutionary design. However, reaching its full potential requires tuning of parameters that affect the performance of the evolutionary search process. Performing experiments on benchmark fitness functions does not provide general knowledge required for such tuning. Therefore, in order to gain an insight into the link between the characteristics of the fitness landscape, the parameters of the selection technique, and the quality of the best found solutions, we perform an analysis based on the NKq model of rugged fitness landscapes with neutrality. As a result, we identify several rules that will help researchers and practitioners of evolutionary algorithms adjust the values of convection selection parameters based on the knowledge of the properties of a given optimization problem.