next up previous contents
Next: 2 Parity Remarks Up: 2 Remarks and Problem Previous: 3 Question 3: What   Contents

1 Ant Remarks

The results from Chapter 5 showed an additional example of how avoiding deception, via increased diversity, improved performance. The results of the sampling study at the end of Chapter 5 also demonstrated how the Ant problem samples many different structures of a small size and fewer different behaviours of increasing fitness.

The correlation between good fitness and high fitness-based diversity in the Chapter 4 Ant experiments also showed the benefit of maintaining more different fitness values, i.e. fewer different behaviours with the same fitness. However, the correlation between good fitness and low edit distance diversity in the later stages of the Ant experiments showed how better runs also converged more. The survivability study at the end of Chapter 7 showed a phase change in the Ant experiments, where in-liers began producing the majority of surviving offspring. This change, around generation 10-15, occurred when more and more individuals became part of the un-fit population. When better runs converge, and in-liers have much higher survivability, it is not clear what the rest of the population contributes. A hill-climbing or local optimisation strategy at this point may be a more effective and efficient way to continue search.


next up previous contents
Next: 2 Parity Remarks Up: 2 Remarks and Problem Previous: 3 Question 3: What   Contents
S Gustafson 2004-05-20