3 Genetic Outliers and Survivability

Genetic outliers are defined in a genetic programming population by first dividing the population into the fit and un-fit. The fit, which are more likely to be selected, are further divided into the similar and the dissimilar using a pair-wise distance metric. Figure 7.1 shows these divisions. Individuals which are dissimilar from the rest of the population are unlikely to have much of a role in the evolutionary process if they have low fitness and are unlikely to be selected. Therefore, the definition of genetic outliers considers both fitness and genetic differences. Next, survivability is measured in these subpopulations by counting the number of times members are selected for recombination and the number of times their off-spring are selected in the next generation. This denotes the offspring's relative rank in the population and the original individual's survivability.

Figure 7.1: The phenotype space is divided into the fit and un-fit. The fit genotype space is then divided into the outliers (filled region) and the in-liers (shaded region). Of course, this example is does not represent the true relationship between phenotype and genotype spaces, it is only an example of defining the outlier individuals.

The previous chapters have shown how trends in diversity and convergence are similar between problem domains. At the same time, Chapter 5 highlighted that the type of search genetic programming carries out in each domain can vary. Specifically, domains sample structures and behaviours differently, indicating a varied response to different levels of diversity. This was also seen in Chapter 4, which showed the varied correlations between diversity and fitness on different problems. As constructed problems often focus either on the search for structure or content, but not both, to initially investigate the role of genetic diversity during the evolutionary process, the Tree-String problem is developed.

The Tree-String problem has two objectives: to match a target tree structure and to match a target string of symbols realised on the structure. This problem has explicit goals of searching for structures and contents which are likely to be conflicting, but are representative of the general class of problems genetic programming is applied to. A much smaller study will be performed later using the problem domains from previous chapters. This later study is carried out to provide additional confirmation of the applicability of analysis and to continue to improve the understanding of those domains.