Information Technology Reference
In-Depth Information
It was chosen because it illustrates one of the dangers of applying parsimony
pressure prematurely. As you can see, the system managed to create very
compact decision trees that were far from perfect solutions (for instance, the
decision trees discovered in generations 55 and 99 are both extremely com-
pact but could only solve 12 and 13 instances respectively). And as you can
see, it took a while for the system to find its way around those local optima.
For instance, it took 364 more generations for the system to create a solution
that surpassed the almost perfect and extremely compact solution discov-
ered in generation 99. And not surprisingly, this first perfect solution is far
from parsimonious and totally unrelated to the best of generation 99. As you
can see, it went through six simplification events before ending up with just
eight nodes, which, as we know, is the most parsimonious configuration that
can be achieved to describe the data presented in Table 9.3.
So, no matter how redundant a configuration is chosen, for such simple
problems, it is always possible to find the most parsimonious solutions by
applying parsimony pressure. For more complex problems, this strategy might
be useful to get rid of unnecessary neutral blocks that might be costly in
terms of computational effort. However, as we learned here, it might be more
productive to apply parsimony pressure just after the creation of the final
model in order to avoid unnecessary local optima.
Let's now see in the next chapter yet another interesting application of
gene expression programming - neural network induction - that relies even
more heavily on the swift handling of large amounts of random numerical
constants.
