Geoscience Reference
In-Depth Information
But should GC researchers really be worried about whether industry can continue to stay on
target with Moore's law during the next decade? With increasing access to parallel, distributed
and cloud computing (Adnan et al. 2014; Birkin and Malleson 2014; Sui 2014) which provides
solutions to speed, memory and storage, what problems do researchers in GC have that need
more powerful computers? Four different areas are discussed in the following text, yet most of
these have caveats or associated limitations that surpass those of the need for more computing
power:
1.
Agent-based or individual-based modelling
: There is currently a limit to the number
of agents or individuals that can be simulated and/or the number of characteristics and
behaviours that can be captured. Yet there is a desire to build more realistic models with
a greater number of agents, which poses some computational challenges. Referred to
as massive agent-based systems (MABSs), they contain very high numbers of agents,
that is, on the order of millions or more. Parry and Bithell (2012) present a number of
potential solutions for how to build MABS. Their suggestions include investing in more
computer power on the one hand while suggesting other approaches such as moving to a
parallel computing environment or restructuring the software to build super-individuals,
which are like cohorts of aggregates of individuals. These latter two solutions are illus-
trated using an agent-based model (ABM) of insect populations. Farmer (2011) talks
about his ambitious plans to build a MABS of the entire United States in order to predict
events like the 2008 economic crisis before they happen. Although many challenges
have been identified, computing power is not considered to be one of the limiting factors.
Instead, more pertinent issues are identified including how to calibrate such a model,
what elements to include and at what level, limits to prediction due to uncertainty in the
stock market and how to build good decision rules for the agents. In this case, the limit-
ing factor is really the model complexity and sufficient understanding of the behaviour
of the agents.
2.
Methods of search and optimisation including model calibration and sensitivity analy-
sis
: At present, computational power limits the size of the parameter spaces that can be
explored either in searching, optimisation or model calibration as well as the number of
parameters that can be tested in a sensitivity analysis if the procedure is to be carried out
in a time that is considered to be acceptable to the researcher or project. As the model
becomes more complex, the possibility for exhaustive analyses becomes less likely, so
researchers need to make decisions about what is good enough in order to have confidence
in their results. Greater computational power would allow GC researchers to evaluate more
solutions or parameters as originally advocated by Kirkby (2000).
3.
Increasing the resolution of grid-based modelling
: Large-scale physical models such as
global climate and numerical weather prediction models are currently restricted to the
size of the grids on which they operate. With improvements in computing power already
witnessed over the last four decades, there has already been an increase in the resolu-
tion of these models, which is a trend that will continue in the future. But does improved
resolution alone ensure better forecasts? The answer is no. Although a finer resolution
would allow more local conditions to be taken into account such as the influence of local
topography or vegetation, the physics that govern these models also needs to be improved.
Thus, developments in understanding local processes are a fundamental part of improving
models with higher resolution.
4.
Our ability to process, store and visualise vast quantities of data
: The amount of
georeferenced data has increased phenomenally during the last two decades. We may need
faster computers and more storage to handle these datasets, but the real issues are much
more complex. The data themselves represent a limitation to GC, which is discussed in
more detail in the next section.
