Geoscience Reference
In-Depth Information
application of EAs was unlimited and largely untapped (Diplock 2000). It is clear from review-
ing the literature over the last 10 years that the potential of EAs has not yet been realised within
GC. Why is this? EAs have the advantage over other optimisation methods in that they are black
box, few assumptions are made about the underlying objective functions and they do not require
insight into the nature of the problem. Could the lack of uptake be due to the limitations and
technical difficulties or problems that arise from the implementation of EAs? For example, one
major issue with the GA approach that has emerged from related work is the large amount of
computational time that can be required to run the model. Malleson et al. (2013) developed an
ABM linked to a GA for optimising parameters in a crime model, specifically the operation of
behavioural rule sets of burglars. Even after some simplifications from the original configura-
tion (Malleson et al. 2010), a single model run still required approximately 10 h to complete on a
normal desktop machine. Malleson et al. (2013) used a 16-core virtual machine provided through
Amazon Web Services (Expósito et al. 2013) to generate their results, but even with this hard-
ware, each GA iteration - with a population of only 20 chromosomes - required approximately
20 h to complete. An associated side effect of the computation time is that it is not feasible to run
each individual model configuration multiple times, which would be preferable to give a more
comprehensive assessment of the model error (especially if the simulation is probabilistic so each
run will lead to slightly different results). More powerful computer systems are required if this
GA is to be run with a larger population for a larger number of iterations and with multiple runs
per model configuration.
In addition to the computational time required, when constructing an EA, there is a time invest-
ment in the general setup. Although there are freely available
off-the-shelf
packages as mentioned
previously, most researchers tend to use a programming language or modelling environment to
build their own programs. Careful consideration must be given to the way in which parameters
are represented, and an appropriate fitness function must be chosen for evaluating if the solution is
successful: if these criteria are poorly defined, then the EA may be unable to find a solution to the
problem or may end up solving the wrong problem! Initial experimentation is also required to find
the appropriate values for general EA parameters such as mutation rate, recombination method and
size of population, although the literature can provide some guidance.
The hybridisation of EAs with other established methodologies in GC for model calibration is an
area that shows the greatest potential for the uptake of this approach. There has already been consid-
erable work in linking together neural networks with GAs (Abrahart et al. 2007; Heppenstall et al.
2007b) to improve the accuracy of forecasting in rainfall-runoff modelling. The work of Fischer
and Leung (1998) is illustrative of a similar vein of research in modelling SI data; see also Reggiani,
Nijkamp and Sabella (2001).
Another developing area of EAs that can potentially offer a new methodological avenue for
researchers is the use of EAs for optimising model structure. This approach is of particular use,
for example, in finding solutions for difficult open-ended problems. Here, an automated method
is required that can estimate the right number of dimensions for a particular solution regardless of
the spaces within which it exists. For example, the structural complexities within a neural network
model are often related to the data sets used in the model development. The standard approach
to building a neural network requires a structure to be selected
a priori
and thereafter trained;
the usual format comprises a single layer of hidden units, with each hidden unit in that hidden
layer being connected to each unit in the input and output layers. Thus, the goal of networks with
fixed topologies is to optimise the connection weights that determine the overall network function.
However, their structure will also influence their behaviour, and topology is important because it
determines the size of the solution and therefore the size of the space in which that solution can
be found. There are no guidelines that relate network topology to particular problems or solutions.
This is an area that is still in early development but clearly using EAs (such as the JNeat package:
http://www.cs.utexas.edu/users/ai-lab/?jneat) could potentially produce models with both optimal
structures and parameters.
