Information Technology Reference
In-Depth Information
GEP systems. It was created in generation 837 of run 29 and looks like this:
CAAabCbBAFbDDHFGbHbDGGAI...
...aabaabaaabbabaaaaababbba...
...bbb2453417195768259147257308
C = {0.229217, 0.48703, 0.287506, 0.091735, 0.808899,
0.652405, 0.977082, 0.931214, 0.118988, 0.145019} (9.4)
The decision tree of this individual involves a total of 37 nodes and, given
the bulkiness of the attribute nodes, we won't be able to show it here (you
should be able to draw it easily on paper though). However, and because I
believe it is important to see the structures of complex real-world models to
make things a little bit more palpable, a much smaller model was created in
another experiment using a head size of 10:
0123456789012345678901234567890
ABbBbaFDCIbababaaabbb3894830502
C = {0.297699, 0.667999, 0.392853, 0.896088, 0.149322,
0.39212, 0.183807, 0.289062, 0.461792, 0.23587} (9.5)
This decision tree involves a total of just 15 nodes (see Figure 9.15) and has
a fitness of 341 on the training set and a fitness of 171 on the testing set,
again showing that GEP decision trees with numeric attributes do indeed
generalize outstandingly well.
9.3.2 Classification of Irises
The iris dataset is also an interesting one to use with the EDT-RNC algorithm
as it would allow us to compare the performance of this algorithm both with
the cellular system with multiple outputs and the GEP-MO algorithm.
You know already that there are four numeric attributes in the iris dataset
and three different classes of irises. The set of attributes will consist in this
case of A = {S, T, P, Q}, corresponding, respectively, to SEPAL_LENGTH,
SEPAL_WIDTH, PETAL_LENGTH, and PETAL_WIDTH, all of which will
obviously divide into two different branches. The terminal set will consist of
T = {a, b, c}, corresponding, respectively, to
Iris setosa
,
Iris versicolor
, and
Iris virginica
. Furthermore, a set of 10 random numerical constants drawn
from the rational interval [0, 10] will be used and, as usual, they will be
