Information Technology Reference
In-Depth Information
6.6 Multiple Cells for Multiple Outputs: The Iris Problem
We have seen already in this chapter some of the advantages of the cellular
systems, especially the discovery of modular solutions, code reuse, and to-
tally unsupervised linking. Now we are going to see how the multicellular
system can be successfully used to solve problems with multiple outputs.
In all the problems we have solved so far with the multicellular system,
the different cells were engaged in finding the same kind of solution and
only the performance of the best cell mattered to evaluate the fitness. But,
here,
n
different cells will be used to classify
n
different classes and, there-
fore, the fitness of the individual will depend on the accuracy of the different
sub-models expressed in the different cells.
This is indeed very similar to what we did in chapter 4 (section Multiple
Genes for Multiple Outputs), where the multigenic system of gene expres-
sion programming was used to solve problems of multiple outputs. But there
is a very important difference between these systems: whereas in the
multigenic system we could only have
n
genes for classifying
n
classes, in
the multicellular system we can have as many normal genes as necessary
encoding different automatically defined functions. These ADFs are then
invoked as many times as necessary from the
n
different cells. This means
that, for instance, if there is a sub-task that is very easy, the cell in charge of
solving it might only need to use one or two ADFs; but if one of the sub-tasks
is very complex, the cell in charge might need to combine more ADFs, say
four or five, in order to build a good model. We will see below that the
multicellular system with multiple outputs works exactly like this, creating
extremely compact and intricately connected models.
As a comparison, in the multigenic system with multiple outputs, we don't
have this kind of flexibility: for all the sub-tasks a single gene encoding a
single sub-ET is all the system has to work with, irrespective of the complex-
ity of the sub-task.
So, in this section, firstly, we are going to analyze the performance of the
multicellular system with multiple outputs by solving the already familiar
iris problem of section 4.2.3. Here, we are going to use pretty much the same
settings (number of runs, number of generations, population size, function
set, training set, and fitness function) so that we can compare both algo-
rithms (see Table 6.8 and compare with Table 4.7). And secondly, we are
going to study the structure of these intricately connected models by analyzing
the structure of one of the best-of-experiment solutions.
