Geoscience Reference
In-Depth Information
18.5 COMPUTATION-LED LIMITS
This limit to GC deals with the question of whether computers can ever replace humans and thereby
be used to help solve some currently difficult and complex GC problems. The idea of a thinking
machine is relevant to GC, since AI is one of the cornerstones of this discipline. As AI is ultimately
about designing machines with human intelligence, will we ever be in a position to use computers as
if they were real GC scientists? According to predictions by Kurzweil (2005), the answer would be
yes as he predicts that machines with human intelligence will be developed by 2029. He views com-
putational power as the limiting factor in developing such a capability. However, the development
of a thinking machine is much more complex than simply having adequate computational power
and represents another real limit to GC research. Intelligence is still a concept that is hotly debated.
One of the original ideas for the development of intelligent machines was put forth by Alan Turing
who wrote a seminal paper in which he outlined his famous Turing test (Turing, 1950). The idea was
simple: if a human engages in a conversation with both another human and a machine but cannot
distinguish between the two, then the machine is said to exhibit intelligent behaviour. The ELIZA
and PARRY programs were examples of initial attempts to pass the Turing test (Weizenbaum
1966; Colby et al. 1972). However, the Turing test has been criticised by many researchers (Searle
1980; French 1990) since it merely determines whether a computer can imitate the behaviour of a
human and not whether it exhibits intelligence or whether it can think in any reasoned capacity.
Despite these criticisms, the Turing test is still the subject of development as encouraged through
the Loebner Prize. This annual contest has been held since 1991 (http://www.loebner.net/Prizef/
loebner-prize.html) and continues to the present day to judge attempts by conversational programs
(or chatterbots) to pass the Turing test.
Research in AI over the last 50 years has concentrated on addressing individual aspects
of human intelligence, for example, the ability to learn through research into artificial neural
networks and the ability to reason through the development of expert systems and fuzzy logic
(Demetriou et al. 2014; Fischer 2014; Fisher and Robinson 2014), but the sum of these have not
resulted in a real thinking machine. Perhaps the closest current example to an intelligent machine
is IBM's Watson (Baker 2011), which was able to beat any of the best human competitors in the
game of Jeopardy. Although it relies on brute force computational power and access to terabytes
of information, it also contains cognitive computing or functions for analysing relationships in
natural language and the capability to learn from its mistakes. As with the Turing test, Watson
has its critics (Searle 2011), but IBM is now adapting this technology to analyse big data in the
healthcare, financial and business sectors.
18.6 UNCERTAINTY
There is an increasing need to provide uncertainty estimates with model predictions. Not only does
this provide confidence in the model, it also addresses the issue of trust, which is an important com-
ponent of modelling, especially when the results are generated for stakeholders or the public. This
becomes particularly relevant when providing estimates of risk for critical systems as evidenced by
recent events in Italy (Nature 2012).
There is already a considerable literature on how to estimate and visualise different types of
uncertainty in geographical models and information (Phillips 1999; Goodchild and Zhang 2002;
Maceachren et al. 2005; Montanari 2007). Although the uncertainty estimation itself is not always
straightforward, especially when trying to account for uncertainty from different sources, the
discussion here is concerned with the situation where the uncertainty estimates become so large
that model outputs essentially become meaningless. Maslin and Austin (2012) discuss this issue
in the context of whether global climate models have reached their limits in terms of predicting
the impacts of climate change. For example, they cite outputs from the UK Met Office's HadCM3
model where monthly discharge is predicted to fall by 16% on one end of the uncertainty band and
