3 Contributions

This thesis makes the following contributions:

1. A survey and analysis of diversity in genetic programming demonstrates the complexity behind the issues of diversity measures and methods and the relationship between diversity and fitness.

2. An analysis using genetic lineages shows how a search metaphor of hill-climbing can be used to explain and improve genetic programming search. Also, the sampling of unique structures and behaviours by genetic programming demonstrates the low sampling of both complex behaviours and unique structures of large size.

3. A causal model is developed which links increased rates of code growth to non-decreased selection pressure and to increased similarity within the population. Decreased selection pressure occurs when fitness-based diversity is lost, and increased similarity in the population is the result of both faster convergence and non-decreased selection pressure.

4. An analysis using the Tree-String problem shows the inability to produce good offspring by both dissimilar-and-fit individuals and by similar-and-equally-well-fit individuals.

5. A model is proposed that identifies dissimilar individuals and moves them to new islands where they can contribute to search more effectively.