... steps

Koza (1992) reported the number of 400, but work by W. Langdon suggests this number was 600.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... similarly

Figure 4.10, plotting edit distance Two diversity, shows an initial increase followed by a decrease in most experiments. The decrease comes at a later stage than that seen with edit distance One. While this might seem to imply that their is an increase in diversity with respect to root portions of the trees, which occurs for a longer amount of time, two other issues need to be considered. First, the $K$ value is 0.5, meaning that each depth level in the tree (for binary trees) is capable of contributing the same to the distance measure. Secondly, the Ant problem uses a 3-arity function and the regression problems both contain unary functions. These two issues explain why a more dramatic increase in edit distance Two diversity is seen in the Ant problem (due to the 3-arity function) and less in the regression problems (due to the unary function) and an overall later decrease in all problems (due to the trees still growing toward maximum depths in combination with the $K$ value, as seen in Figure 4.6). The performance of both measures, edit distance One and Two, can also be seen for the same problem domains in Chapter 5, Figure 5.2. In this case, it is more clear what increased diversity under the edit distance Two measure looks like.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... fitness

It is common, when using `wrapped' functions (log and division), for solutions in regression experiments to obtain an extremely high fitness value. Thus, when calculating the average mean fitness of the population in each generation, it is necessary to remove these individuals from the calculation to avoid skewed results. This was performed for the Binomial-3 graphs, which resulted in approximately 1% of the population being ignored during the calculation.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.