Information Technology Reference
In-Depth Information
Figure 3.2
Subcomponent crossover within a parse tree. This kind of exchange is not
possible with standard subtree crossover.
would have a considerable computational overhead. Moreover, the real problem
does not lie with crossover but with the parse tree representation.
One problem is that the parse tree representation is not very flexible and does
not readily support many forms of editing. Other, more flexible, representations
have been used in GP [4, 26, 32, 35], though not widely. However, inflexibility
is just part of a wider problem of the parse tree representation not being well
suited to evolution. Parse trees were not designed to be evolvable, and therefore
it is unsurprising that they are not ideal candidates for evolution. Evolvability is
not a common property of representations, and it seems likely that it is especially
uncommon among those used by humans to solve problems. Trees, for example,
implicitly capture the needs of humans that a structure be decomposed and
explicitly connected in order to be easily understood. Evolution does not require
these constraints because it does not need to understand what is being evolved.
Introducing constraints to evolution only serves to limit what can be evolved.
A Biomimetic Solution
The obvious solution, therefore, is to find a program representation that is evolv-
able. This is the approach taken by enzyme genetic programming. However,
enzyme GP does not attempt to construct a new representation afresh, but rather
look to biology, an evolutionary system that already possesses an evolvable rep-
resentation, which, in a sense, represents computational structures. The aim of
enzyme GP is to identify principles of biological representations thought to con-
tribute to their evolvability and adapt these principles to improve the evolvabil-
ity of representations in genetic programming. It should be stated that the aim
is not to produce an exact model of these biological representations but only to
mimic those constructs and behaviors that might improve GP. Before introduc-
ing enzyme GP, we review the biological representations on which it is based.
