Biomedical Engineering Reference
In-Depth Information
problems, proofs of convergence of the search
algorithms, and classifying problem spaces.
On the practical side, GP research will target
[44]
:
development or evolution versus those from
learning. Its reason is the less stringent separa-
tion of time scales among evolution, develop-
ment, and learning in GP systems. Whereas
biological evolution can happen over many
thousands or millions of years, development
over the lifetime of an organism, and learning
over phases of that lifetime, all three mecha-
nisms are on similar time scales in GP, usually
tied into single runs of a GP system. In addition,
in most GP systems there is no notion of species
(and their evolution). Rather, the entire popula-
tion is essentially mixed and therefore belongs
to one single species. Finally, the goal under the
influence of evolution is behavior, the same
entity usually associated with learning. First
attempts to examine learning as a separate task
for which evolution/development have to pro-
vide the means have been made
[49]
, yet this
area requires much more investigation.
• theidentiicationofappropriaterepresenta-
tions for GP in particular problems,
• thedesignofopen-endedevolutionary
systems with GP,
• theproblemofgeneralizationinGP,
• theestablishmentofbenchmarksfor
measuring and comparing GP performance,
and
• modularityandscalabilityinGP.
There is also more room for adding bioinspi-
ration to GP. For instance, the relation between
evolution and development has been studied
for decades in biology
[45]
. It was found that
the time-dependent process of gene expression
and gene regulation through both internal and
external cues is the mechanism by which both
processes can be unified
[46]
. Some progress
has also been made in GP to couple evolution
and development. The developmental approach
in GP takes the form of a recipe that, upon its
execution, generates a structure that is sub-
jected to fitness tests. Thus, it is not the GP
program itself that is tested but the result of its
execution.
Similar to the coupling between evolution
and development, the coupling between devel-
opment and learning was considered an
important link for understanding the mecha-
nisms of development and learning processes.
Cognitive neuroscience has presented evi-
dence for this coupling by finding that there
are critical periods in development in which
certain learning tasks are facilitated (and only
sometimes possible). If the critical period is
missed, learning success in a task is substantially
reduced
[48]
.
The coupling between development (or evo-
lution) and learning has only recently been
explored in GP. The problem is to clearly sepa-
rate adaptations or fitness gains resulting from
17.5 APPLICATIONS
The textbook of Banzhaf
et al
. from 1998 lists 173
GP applications from A to Z that already existed
at the time
[24]
. Fifteen years have passed, and
the field has continued to develop rapidly. The
main application areas of GP are (from narrow
to wide)
[24]
:
• computerscience,
• science,
• engineering,
• businessandinance,and
• artandentertainment.
Koza has contributed many interesting applica-
tions to some of these areas, demonstrating the
breadth of the method. However, a more detailed
look is warranted.
In computer science, much effort has gone
into the development of algorithms using GP.
By being able to manipulate symbolic struc-
tures, GP is one of the few heuristic search
methods for algorithms. Sorting algorithms,
