Information Technology Reference
In-Depth Information
sure that the author was really convinced of the optimality of his result, due perhaps to his
limited experience on the specific application domain, which was not enough to judge.
What the author surely did was to try a number of different topologies, paradigms, network
sizes, training algorithms and found that his own network was offering the very best
performance among all those tests. All tests were neural and no test was performed
according to the state-of-the-art using traditional approaches. This method is (partially)
correct to
optimize the performance of a novel intelligent system
(namely, to find which are the
best choices to get the best out of is, among all possible intelligent systems), but
not
to
evaluate the
appropriateness of an intelligent system for the given application
, compared against a
standard one.
What was true for that specific problem, was that the proposed hybrid empirical/analytical
model developed by a team of experienced engineers and biologists offered a much better
performance than the best existing neural network, even for a comparable computation
complexity, not considering the possible performance of the state-of-the-art. The reason for
that (which happens much more frequently than one can even imagine) is that human
experience, knowledge and mental capacities, which are used to develop a given “non-
intelligent” model, boost so much the overall performance of a given system than even an
optimal intelligent system, trained in the best way but without using the available human
knowledge, cannot compensate ignoring human knowledge during its development.
3.6 How to avoid optimism…
An important step towards acceptance is to avoid unnecessary and inapplicable optimism.
Any development, comparison or selection has to be
fair
and based on
real and well proven
data
; never on
hypotheses
. Optimism usually tends to push the designer towards a solution
which then proves less performing than originally expected, therefore convincing even more
the decision-making people that intelligent systems are not yet a viable solution to their
problem.
3.7 Tools and support
An important step towards industrial acceptance (as for many other industrially relevant
items) is the availability of an appropriate support to the development, use, integration,
conduction and maintenance of the system.
An excellent intelligent system will never be applied until its use is straightforward and
user-friendly. The only chance to have an intelligent system applied is therefore the
development, around the intelligent system itself together with its surrounding elements
(e.g. pre-processors, postprocessors, data mining, etc.), of an appropriate user interface and
development tool which supports, in the order:
the decision-making process in helping to choose the intelligent systems instead of any
other traditional system
the preparation phase (e.g. data collection, training, tuning and testing)
the conduction phase, namely the nominal operation of the intelligent system, when
applied to the industrial process under interest
maintenance, to overcome any problem might occur during conduction.
