2 Code Growth and Problem Difficulty

The results in Chapter 6 demonstrated several possible areas of controlling code growth indirectly. While the recombination of genetically similar individuals is likely to cause a consistent amount of code growth in future populations, the recombination of dissimilar individuals, while producing less overall code growth, is more likely to also produce fewer good offspring. Therefore, investigating adaptive recombination methods that are aware of the similarity of individuals could be a solution to minimising unjustified code growth while achieving good fitness.

Chapter 6 also showed the effects of problem difficulty. It may be possible to control population diversity to better deal with harder instances. For example, when a difficult instance causes a population to contain one really good individual, methods which dynamically prevent the over-selection of this individual are likely to improve the overall performance of the algorithm.

Parts of the causal model in Chapter 6 remain to be fully evaluated. An area of future research here would be measuring the dissimilarity of solutions in hard and easy instances. The hypothesis stated that easy instances allow more optimal and different solutions to be acquired quickly. This was based on the notion that easy instances can be solved equally well by different solutions in different ways. Hard instances were thought to be solved by fewer solutions that are more similar. Future work can investigate the actual dissimilarities in solutions for easy and hard instances.