7 Diversity, Survivability and a Niche for Island Models

Population diversity effects many aspects of the evolutionary process. For example, the number of unique fitness values in the current population effects selection pressure. Also, a high amount of genetic differences, indicated in previous chapters by high edit distance diversity, is likely to produce slower code growth. Low amounts of genetic differences, or low edit distance diversity, describes convergence and can lead to lower selection pressure. Diversity is obviously an important topic in genetic programming. Without adequate diversity the population may be unable to produce variations to improve solution quality. However, increasing diversity can prevent sufficient exploitation and convergence. With respect to search, how do diverse individuals contribute to the evolutionary process? That is, how successful are the members of the population, which account for the most genetic dissimilarity, in producing new solutions during search?

The survivability of dissimilar and fit individuals is examined in this chapter to establish their role in the search process. The Tree-String problem is introduced and used in an empirical study. This problem has many similar attributes to other domains and constructed problems. The results from this study are compared with a survey of distributed and related methods that use diverse and fit individuals differently. Chapter 4 surveyed previous measures and methods of diversity. In this chapter, that survey is complemented with a survey of distributed models and related methods. Based on the analysis of the following study, a detailed proposal is made to improve genetic programming search by encouraging the recombination of genetically similar individuals by means of a speciation island model.