Genetic Programming - GeoComputation - page 197

of around 90% (Beriro et al., 2013; Moreno et al., 2010). Without knowing that you are modelling
calculated data you may conclude, like Shiri and Kişi (2014), that an evolved GP program outper-
forms traditional E PA N models by a specified amount and should therefore be employed in water
resources planning. The right answer for the wrong reason? Another good illustration of the rel-
evance of what is being modelled can be found in Beriro et al. (2013). They concluded that the
quality of the testing data can be more influential in terms of discriminating between models than
the quality of the original data that was used to develop such models, notwithstanding the fact that
development data must contain sufficient information about the natural system in the first instant.
The key message in such examples is that close attention should be paid to what is being modelled
as it has a direct relationship with what we might reasonably expect in return from our investment
in model development and testing.
8.5.4 h ere B e d ragonS ?
Here be dragons refers to dangerous or unexplored territories, in imitation of the medieval carto-
graphic practice of putting dragons, sea serpents and other mythological creatures in uncharted
areas of maps. Such geographical phrasing nicely extends our earlier concept of a two-headed mon-
ster, and such an analogy may be well suited to describe what is actually happening in the field.
GP is a well-established data-driven modelling tool that over the past 10 years has been used suc-
cessfully by geographers and environmental scientists alike. Truly practical applications are still
few and far between, despite the fact that many papers purport to having evolved new and excit-
ing equations. Why is this? Perhaps it is because authors fail to provide the missing link between
the evolved models and the conceptual model of the system being examined. This might be true.
However, another reason that the current authors feel prevails among potential users is that GP
is scary: an unknown entity, which is apparently so complicated, that it is feared. Likewise, this
may also be the reason why some existing users fail to go beyond using goodness-of-fit statistics
to select a preferred model. Is this happening because GP is too high tech? Is it too challenging? Is
artificial intelligence something that electrical and computer engineers can use to design new wash-
ing machines, as opposed to a tool that is readily accessible to modern graduates of GC? Indeed,
unless we bridge the gap between theory and practice, GP may well end up being a slightly esoteric
choice for modelling studies, one that fails to properly link in with current environmental policy and
decision-making activities.
We should also consider whether GP is being used as a toy or a useful tool. Looking at some
recently published research suggests the former. For complex hydrological problems, it has certainly
been used extensively, and to a lesser extent, in spatial analysis, but this does not really allowing
us to manoeuvre it towards mainstream applications or deeper scientific enquiry. Computer tech-
nology and software is such that complicated modelling operations can now be performed on a
home computer by researchers with little to no background in computer science, programming or
environmental modelling, but how are we responding to this? Recent literature reflects two distinct
lines of enquiry: (1) GP is a curve fitting black-box data-driven modelling tool where findings are
heavily weighted on goodness-of-fit statistics and intermodel competition; and (2) GP is a novel in
silico laboratory method able to generate evidence that can be used to test known relationships or
derive new knowledge. If the second, and if GP is as prevalent as Scopus literature searches sug-
gests, then should we not be seeing it used by the masses, in a similar way to traditional statistical
techniques, or taught regularly in undergraduate classes? After 15 years of progressive develop-
ment, one might expect so. Are people really too afraid to use it or is it not as great as its propo-
nents suggest? Is it really a two-headed monster or scary dragon? The authors of this chapter are
strong proponents of the technique, and we urge you towards experimentation and discovery, and
to bravely go where others fear to tread!
One of the reasons for writing this chapter was an attempt to demystify GP and present it as a
simple and accessible tool that provides an exciting and novel way of examining geographically