next up previous contents
Next: 5 Sampling of Unique Up: 4 Discussion of the Previous: 3 Binomial-3   Contents

4 Remarks

Is increasing diversity beneficial to genetic programming? The results have shown that increasing the genetic differences in populations allows for more global search and local optima avoidance. The results have also shown that higher genetic diversity leads to less code growth, and that increasing genetic diversity adds longer periods of entropy increase. While higher entropy indicates a more explorative search, lower entropy results in less selection pressure as more individuals have the same fitness value. Also, the slower increase of entropy and higher genetic diversity appears to decrease the hill-climbing behaviour that previous research has shown to be effective in solving these problems. Future methods used to increase diversity to improve fitness should clearly state the motivation for such an increase and why that type of diversity would be beneficial. Diversity methods may not be justified in their own right, but they work together with more elitist strategies or as a supplier of programs to a local search method.

Genetic programming searches for solutions to a given objective using program-like representations. However, the task of evolving both the structure and content of a solution is complex and difficult to understand [Daida et al., 2001,O'Reilly, 1998,Hu et al., 2002]. The research presented in Chapter 4 and earlier in this chapter has focused on understanding the relationship between diversity and search, particularly on the kind and level of diversity that encourages good performance. So far this chapter has showed how increased diversity negatively and positively effects performance on several problems. A metaphor of hill-climbing search helped explain the results, where deception appeared to be a cause of poor performance. To further understand diversity and search, particularly with respect to the solutions the population contains during search, the type of structures (tree shapes) and behaviours that genetic programming samples during the evolutionary process will now be examined.


next up previous contents
Next: 5 Sampling of Unique Up: 4 Discussion of the Previous: 3 Binomial-3   Contents
S Gustafson 2004-05-20