Geoscience Reference
In-Depth Information
various sources for a given problem domain in order to transfer this knowledge to a computer (Liou
1998). A variety of knowledge acquisition methods have been suggested in the past (Breuker and
Wielinga 1983; Grover 1983; Hart 1986; Kidd 1987; Jackson 1999), which tend to be a combination
of a few basic approaches. From the literature, the following four main knowledge acquisition meth-
ods are suggested: documentation (e.g. manuals, guidelines and legislation), studying past cases
(e.g. past projects) and their subsequent shortcomings, discussing cases with experts via personal
or collaborative interviews and watching/observing experts applying their knowledge to current
problems. The selection of an appropriate knowledge acquisition method depends on many factors,
namely, the problem domain, the availability of knowledge resources and the time/cost constraints.
As noted earlier, a major obstacle in building an ES is the so-called knowledge acquisition bottle-
neck. Researchers have tried to automate the knowledge acquisition process by employing specific
software (Angeli 2010), and a great variety of approaches have been proposed in the past by Gaines
and Boose (1998), Michie (1982), Boose and Bradshaw (1987), Quinlan (1986) and Clement (1992).
Other studies have focussed on employing AI techniques other than ES for knowledge acquisition
such as neural networks (Cooke 1992) and genetic algorithms (Odetayo 1995). More recent research
efforts have focussed on extending the knowledge acquisition techniques and processes to include
a wider array of participants and knowledge sources (Gale 1990; Wu et al. 2003; Chen and Rao
2008). In addition, knowledge discovery techniques, that is, data mining, can be a solution to this
problem (Yang et al. 2012) through automated mechanisms or via the introduction of technologies
for building ontologies (Breuker 2013) or through graphical representation techniques (Dong et al.
2012). Once knowledge acquisition has been carried out, a methodology is then needed to represent
this knowledge in an appropriate manner. Herein lies another crucial problem with ES and AI more
generally (Vamos 1998). Decision trees are a popular and efficient method that are frequently used
for representing the logic of the problem and the reasoning for solving it (Giarratano and Riley
2005). Building the knowledge base involves the transformation of the knowledge (e.g. the decision
trees) into a format that computers can process, for example, rules in the case of rule-based systems.
Facts and other system inputs/outputs are then defined and combined with the control processes,
namely, forward or backward chaining.
Another significant issue that may constitute a problem in designing ES is how to handle poten-
tial errors and uncertainty. Errors in ES have been classified by Giarratano and Riley (2005) into
seven basic categories: ambiguous, incomplete, incorrect, measurement, random, systematic and
reasoning, all of which may contribute to uncertainty. As a result, the final decisions may not be
the best or they may even be wrong. The most popular approaches for dealing with uncertainty are
Bayesian probability and fuzzy set theory. The concept of a fuzzy ES has already been mentioned
previously. Bayesian probability is one interpretation of the concept of probability that belongs to
the category of evidential probabilities. The Bayesian interpretation of probability is in essence an
extension of the branch of mathematical logic known as propositional logic which enables reasoning
with propositions whose truth or falseness is uncertain. To evaluate the probability of a hypothesis,
the Bayesian approach involves defining some prior probabilities, which are then updated in the
light of new, relevant data. Both the fuzzy and Bayesian methods are capable of handling inex-
act reasoning because the facts, data and/or knowledge are not known precisely. In these situa-
tions, an appropriate mechanism for reasoning under uncertainty should be chosen where the aim
is not to reach the best solution but to find a good or acceptable solution within a reasonable time
limit. Successful examples of systems operating under uncertainty are those noted earlier, namely,
MYCIN and PROSPECTOR, which are capable of reaching a solution even if these solutions cannot
be absolutely proven by the facts and the data.
11.5.2 S
ySteM
d
eVeloPMent
This stage involves the selection of the appropriate development tools, definition of the implementa-
tion strategy and coding/debugging of the system. The available software for developing ES, which
