Information Technology Reference
In-Depth Information
undergoing mutation with only one viable individual in the founder popula-
tion (52% and 92%, respectively).
Also worth considering is that the computational resources required to
guarantee the creation of large founder populations are very expensive. Thus,
systems such as GEP capable of evolving efficiently with minimal initial
diversity are extremely advantageous. Furthermore, for some complex prob-
lems like, for instance, the discovery of cellular automata rules for the den-
sity-classification task of section 4.4, it is very difficult to generate randomly
a viable individual, or even a mediocre one, to start the run. In those cases,
systems such as GEP can use this individual as founder and continue from
there, whereas systems relying on recombination alone will be stuck for a
long time before they gather momentum. Indeed, in gene expression pro-
gramming, due to the varied set of non-homogenizing genetic operators, there
is no need for large founder populations because as long as one viable indi-
vidual is randomly generated the evolutionary process can get started.
12.3 Testing the Building Block Hypothesis
We have already seen in the previous section the limits of recombination of
building blocks. When systems rely exclusively on existing building blocks
and are incapable of creating new ones through mutation or other mecha-
nisms, they become severely constrained. Here, we will pursue this question
further, comparing three different systems: one that constantly introduces
variation in the population, and two different systems that can only recom-
bine a particular kind of building block - GEP genes. Recall that GEP genes
have defined boundaries and, through gene recombination and gene transpo-
sition, it is possible to test new combinations of these building blocks with-
out disrupting them.
For this analysis, we are going to work with the same sequence induction
problem of the previous section, using the general parameters presented in
Table 12.3. For the first experiment, we are going to use a mix of all the
genetic operators; for the second, we are going to use solely gene recombi-
nation at
p
gr
= 1.0; and for the last experiment, we are going to allow a more
generalized shuffling of building blocks by combining gene recombination
(
p
gr
= 1.0) with gene transposition (
p
gt
= 0.5).
In this analysis, instead of creating a founder population as was done in
the previous section, we are going to study the progression of success rate
