Information Technology Reference
In-Depth Information
For instance, the model of an
electric motor
can model nothing else than an
electric motor
, and
its parameters represent, for instance, winding resistance and inductance, rotor inertia,
friction, etc. which are directly measurable and for which the designer can feel if they
assume reasonable values or not. By appropriately varying these parameters, the model will
be adapted to either
large or small motors
, either
fast or slow
, but it will never be able to model,
for instance, a
chemical process
. The designer can easily become aware that, for instance, an
improperly tuned model has a too large or too narrow winding resistance in comparison
with the size of the motor under examination. He can therefore immediately be aware of
improper tuning or of some motor fault or damage and behave accordingly.
On the other hand, intelligent systems are so generic that they can adapt to virtually any
system, either electrical or chemical or economical or mechanical or agronomic, etc. The
same parameters can therefore mean anything, depending on the actual use of the network
(e.g. pollution of a chemical process, yield of a manufacturing process, winding resistance of
a motor, rate of infection in an agricultural plant, etc.); in addition, parameters are
interchangeable and there is no clue to understand what a given parameter really represents
in practice. Further, nobody will ever be aware that training has not been done correctly and
whether the model really represents a given system or not.
3.2 How to avoid crypticity
The use of modern unification paradigms (Reyneri, 1999) allows to easily convert neural and
wavelet networks into fuzzy systems and viceversa, with several advantages, among which,
for instance:
a given neuro/wavelet network can be converted into fuzzy language, thus interpreted
linguistically by experts, who are then able to “validate” and consequently “accept” an
otherwise cryptic neuro/wavelet model;
human experience, usually expressed as a set of fuzzy rules, can be converted into a
neural network and then empirically tuned by means of an appropriate training set.
Fuzzy (or expert systems) rules are usually understandable by an expert, such as
he/she can understand the “concept” which lays behind them. An appropriate neural
training of the rules therefore allows to fine tune the expert's knowledge based on the
available empirical evidence.
It is therefore mandatory to abandon all the older approaches who were more like “magic
formulae” than real engineering methods and concentrate on modern approaches that
consider neural, wavelet, fuzzy, Bayesian, regressor, clustering techniques, etc. as
interchangeable paradigms
. The ever lasting fight among neural- and fuzzy-people is so
detrimental, as it helps to maintain the level of crypticity high, therefore preventing a
widespread acceptance of intelligent systems.
The choice between, for instance,
neural networks
and
fuzzy logic
should therefore be
converted into a more appropriate selection between a
neural
and a
fuzzy language
, which
should be chosen depending on: i) the available knowledge from human experts; ii) the size
of available training set; iii) the availability of other piece of information on the problem; iv)
the level of crypticity which is accepted; v) if and how the model has to be interpreted by
humans or processed by computers.
