Information Technology Reference
In-Depth Information
3.3 Gathering data for network training
Many intelligent systems (mostly, those based on
neural languages
) rely on the availability
of
empirical data
, which is usually gathered into large
training sets
. Unfortunately, these are
often too expensive to obtain, as each data point is usually an appropriate
measurement
of
a mechanical or chemical or biological or economical process. Several processes are so
slow that each point may require up to several days to be acquired. In some cases, if an
accurate numerical model is available, computer simulations can substitute direct
measurements.
Some soft computing techniques (in particular, those based on
fuzzy
and
Bayesian languages
)
may require much smaller training sets, as they rely on a predefined model, described in
linguistic terms
according to
previous human experience
. This is the main reason why fuzzy
logic has been accepted more quickly and extensively by industry than neural systems.
An industrial manager has to consider attentively the trade-off between the cost of gathering
a large training set and the reliability of the trained neurofuzzy network. As already said,
this trade-off often pushes towards the use of
fuzzy languages
whenever possible and bounds
the use of
neural languages
to applications which have enough (historical) data available.
3.4 Analytical vs. empirical methods
As already said, one of the advantages of intelligent systems is that a given analytical/
empirical model is by definition specific and cannot be tailored to a different problem, while
neural networks are. Furthermore an analytical/empirical model usually comes after years
of improvements, while neural networks are trained in a short time. Yet, a purely analytical
model can be developed without any field measurement, while an empirical model requires
a limited amount of field measurement. Instead most intelligent systems always require a
huge amount of field measurements which, in several cases, can take years to gather.
Last but not least, the amount of field measurements which is required (that is roughly the
development time) is a function of the reliability which is asked to the model. A large
training set is in fact mandatory in industry to offer an adequately high reliability, while
reliability of analytical models is often independent of field measurements but relies on
designer's experience.
3.5 Performance is always optimistic...
Virtually any paper published in literature shows that, for a “wide range of applications,
neural networks and fuzzy systems offer tremendously good performance”.
Unfortunately, more than 90% of them do not even try to afford a fair performance
comparison with other
state-of-the-art
techniques and it becomes difficult to feel how good
such performance really is. Just as an example, a paper (not cited) claimed that
the proposed
neural model of a biochemical process is 90% accurate
and the author was
enthusiast
of that
incredible result
. Since the reviewer had little experience on modelling that specific process,
he could not do anything else than blindly accept author's statement. But, when the paper
was read by an experienced colleague, he pointed out that state of the art had already
achieved about 95% since a few years, making those results useless for industry. It is quite
