Information Technology Reference
In-Depth Information
models designed with the MCS-MO algorithm classify correctly 149 sample
cases and, therefore, these models are as good as the model (4.14) designed
using the three-step approach of section 4.2.3. Let's now analyze the struc-
ture of these intricately connected models.
One of the best solutions discovered in this experiment has a fitness of
149.037 and was created in generation 13819 of run 12 (for compactness,
ADFs 0-4 are, respectively, represented by the numerals 0-4):
Normal Genes/ADFs:
-.d1.*./.d2.d0.d0.d1.d0.d3.d2.d2.d0.d3.d1
d2.*.d1.d1.d3.d3.d0.d2.d3.d3.d2.d1.d1.d3.d1
-.*.*.d2.d2.d0./.d0.d3.d0.d0.d3.d1.d1.d3
+.*.d1.-.+.*.d2.d0.d3.d3.d1.d0.d0.d2.d0
-.d1.d2.d3./.d3.+.d3.d2.d2.d0.d3.d2.d3.d1
Homeotic Genes/Cells:
4.1.1.-.4.1.-.1.1.0.3.2.0.0.0
-.*.+.4.2.-./.4.4.3.0.3.3.4.3
-.*.+.0.*.0.3.0.2.4.4.4.3.0.3
(6.10)
In terms of number of hits, this model classifies correctly 149 out of 150
fitness cases. This corresponds to a classification error of 0.667% and a clas-
sification accuracy of 99.33% and, therefore, is an almost perfect solution to
the iris problem.
As you can see by its expression in Figure 6.13, Cell
0
(the cell that distin-
guishes the
setosa
variety from the other two) invokes just one ADF (ADF
4
)
and again we can see that just the difference between sepal width and petal
length is enough to distinguish
Iris setosa
from
Iris versicolor
and
Iris
virginica
. Also worth pointing out is that ADF
4
is also invoked from Cell
1
(three times, although two of the calls cancel themselves out); ADF
0
is used
to design not only the sub-model that is responsible for distinguishing the
versicolor
variety (ADF
0
is used once in Cell
1
) but also the
virginica
variety
(Cell
2
calls this function from three different places); ADF
2
and ADF
3
are
both of them part of the sub-models that classify
Iris versicolor
and
Iris
virginica
, each being invoked just once from each of these cells.
So, apart from some neutral motifs that can be easily removed by hand, it
is worth noticing how compact these models are thanks to code reuse. In-
deed, small building blocks are discovered in the ADFs and then these build-
ing blocks are used as many times as necessary by the different main pro-
grams, creating an extremely compact and intricately connected computer
program.
