Censoring mutation in differential evolution

In this paper we show how relative characteristics of individuals in context of their population can be used to customize and guide the search process in Differential Evolution, which is a state-of-the art real-parameter global optimization algorithm. Analysis of exploitation phase of the search process shows that probability of creating an offspring, which outperforms its parent and can hence enter the population is strongly negatively correlated with distance between them. We use this property in a censored Differential Evolution variant, which in the exploitation phase saves computational time by rejecting distant offsprings without evaluating their fitness. Comparison of censored and classical variants of DE/rand/1 and DE/best/1 for different dimensions of the search space reveals interesting patterns about appropriate choice of the scaling factor and questionably exploitative character of the DE/best/1 algorithm. Finally, we present the main idea of an ongoing study about basing adaptation on ranks or relative fitness of individuals in their population.