next up previous contents
Next: 3 Regression Remarks Up: 2 Remarks and Problem Previous: 1 Ant Remarks   Contents

2 Parity Remarks

In Chapter 5, lineage selection prevented a hill-climbing-like behaviour that increased genetic diversity but led to worse fitness. Previous research showed how elitist strategies that incorporated an element of diversity control also successfully improved search. The deception in the Parity problem is due to the many different solutions that have the same fitness. This is in contrast to the Ant problem, where deception was also due to the fact that some solutions are not easily improved by genetic programming. The sampling study in Chapter 5 showed the ease of which genetic programming can sample solutions near the random strategy, and the difficulty in sampling better and worse ones. The benefit of elitism would focus the search more consistently on a particular strategy and create a hill-climbing behaviour of search.

Chapter 7 showed a phase change near generation 20 in the Parity experiments, where further populations came mostly from the in-lier population. This is similar to the Ant experiments. The variable and sporadic survival rates in the genetic outlier population suggests the importance of variation when the population loses diversity, but also shows the inability of the genetic outliers to produce viable offspring consistently. The ability to distinguish between the many near random behaviours would help to reduce the effects of deception in this problem and allow a more predictable selection pressure.


next up previous contents
Next: 3 Regression Remarks Up: 2 Remarks and Problem Previous: 1 Ant Remarks   Contents
S Gustafson 2004-05-20