Page 107 - Data Science Algorithms in a Week

Evolutionary Optimization of Support Vector Machines … 91

A two-sample t-test shows that, at generation 20, the average best fitness of the
elitism GA is significantly higher at the 0.1 level with a p-value of 0.054 and a lower
limit for the 90% confidence interval of 0.542557.
Figure 13 shows the standard deviation of the two GAs as a function of the
generation which illustrates another advantage of using the elitist strategy: as the
generation increases, the standard deviation decreases. The standard deviation of the GA
th
with elitist strategy is significantly lower at the 20 generation at the 0.1 level in the F
test for two variances and the Bonferroni confidence interval.

Effect of Crossover Type

We tested four crossover types: one point, two points, uniform, and a 4-parents
diagonal. The comparison is shown in Figure 14 and Figure 15.

Figure 14. Effect of the different crossover type on the fitness function.

Table 4. Average and Variance in the 20th generation as a function of the
crossover type

Crossover Type Average Variance
Diagonal 84.24481 1.015474
Two-point 84.10167 0.456379
Uniform 84.06692 1.105777
One-point 83.71069 1.593839