In a general sense, dissimilar individuals allow for exploration to take place. However, as seen in Chapter 7, dissimilar individuals are not consistently effective in populations where they are likely to breed with very different individuals. However, as the population becomes more equivalent in fitness, dissimilar individuals provide a chance for escaping local optima. The previous literature suggests that diverse individuals would be most effective when breeding in separate populations, where they compete and recombine with more like individuals. One of the most important areas of improvement that can be made to genetic programming, or any population-based search method, is the precise characterisation of the population components to allow for new methods to improve dynamics, efficiency and the rate of success. The research presented here makes a contribution to that effort.
The majority of experiments used in this thesis involved three problem domains commonly used in the genetic programming literature. The reason for this was twofold. Firstly, it was necessary to use the same problem domains, and algorithm parameters, on which to build sound explanations across the different chapters and experiments. Secondly, as these problems are frequently used by the community to develop new methods and theoretical arguments, it is useful to apply to them the full range of the diversity experiments and analysis for future reference.