One might expect that higher amounts of diversity generally leads to better search in genetic programming. Diversity and dissimilar individuals could provide key variations that allow new solutions to be discovered and improved by passing on their genetic material to offspring. While the results presented in this chapter showed how dissimilar individuals play an important role in producing offspring when the population is genetically similar, this role is usually unstable. Dissimilar individuals were particularly important when the many fit individuals also had the same fitness. That is, when the most fit subpopulation contains low phenotype diversity, dissimilar individuals are more successful in producing offspring that later contribute toward search. Also, as phenotype diversity is lost among the most fit individuals, the worse fit individuals are more successful in contributing new solutions to later generations.
These results raise questions about the motivation of methods based on dissimilarity mating and blindly promoting high diversity. If dissimilar individuals have a low chance of success producing offspring with high relative fitness, why concentrate effort on producing additional diversity? Lastly, the results also highlighted the possible ineffectiveness of migrants in distributed models to contribute positively toward search. While this was not directly investigated, the results certainly suggest further work should be carried out in validating the effectiveness of migrants.
A niche for island models was proposed that leverages the results from this chapter. The niche consists of identifying structurally different individuals (that are unlikely to mate effectively with the rest of the population) and speciates them to a new island. The new island provides a chance to search more effectively with this individual and the region of the search space this individual represents.
The conclusions of this thesis are presented in the following chapter.