2 Survivability of the Diverse

In the context of search, the recombination of population members would hopefully lead to improved solutions. However, in genetic programming, there are arguments for both preventing and encouraging the recombination of dissimilar solutions. The ability of recombination to improve solutions can be studied by measuring the change in fitness between parent and offspring. As the root parent tends to contribute most of the genetic material, the root parent and offspring relationship is initially considered. However, instead of measuring the absolute value of fitness and its change between parent and offspring, the relative rank of parent and offspring in each respective population will be used to indicate which individuals are guiding the search more effectively (also used in [Luke, 2003]). That is, even if the overall offspring population's fitness is worse, it is still desirable to know which of these individuals contributed to search.

In the following empirical study, the survivability of diverse individuals are studied to understand the guiding forces in search. The phrase diverse individuals refers to those individuals in the population that account for the majority of dissimilarity. These individuals are found by comparing their average dissimilarity to the population's average dissimilarity. Following this study, many adjustments may be possible to improve efficiency or performance. Additionally, there is another important motivation for studying the survivability of diverse individuals in standard genetic programming populations: the role of the migrant in distributed models. Generalising these results to describe the expected behaviour of migrants in a distributed model is not a stretch of applicability. Distributed models are commonly defined in a way to mimic several single population runs with periodic exchange of individuals. The following study, which also considers the most dissimilar individuals as migrants, is probably being conservative as migrants are likely to be much more dissimilar in real distributed models. Next, the genetic outlier definition and the measure of survivability are described.