Information Technology Reference
In-Depth Information
2. Maturity level of intelligent systems
As mentioned in the introduction, all soft computing and intelligent systems techniques
suffered alternating periods of
acceptance
(due to the novelty and the promising preliminary
results) and
rejection
(due to the acquired awareness of limits). Hundreds of algorithms,
topologies, training rules have been: i)
conceived
and
developed
; ii)
tested
,
evaluated
,
tuned
and
optimized
; iii)
temporarily or partially
abandoned
(>>90%); iv)
accepted
and
applied
to real
problems (<<10%).
Most of the original theories have been nearly abandoned (like, for instance,
Hopfield
networks
and
Boltzmann machines
,
glass spin theories
,
stochastic networks
, etc.) either because
they could not offer reasonable performance or because they were too cumbersome to use.
Other theories (like, for instance,
perceptrons
,
radial basis functions
and
fuzzy systems
)
eventually reached widespread acceptance, since they are more viable.
What is then at present the level of maturity of intelligent systems? This can be evaluated
from a series of clues such as:
how many theories and paradigms
have been developed altogether
. This number should be
as high as possible, to ensure that no option has been forgotten;
how many paradigms
have survived after maybe ten years
. This should be low, to
minimize the knowledge one needs to learn (see section 2.2);
the
level of acquaintance
a typical engineer has with these techniques. This should be high
and it should be achieved quickly (see section 2.1);
the
count of accepted industrial applications
should be significant (see sections 3, 4, 5).
Due to the maturity level they reached in about half a century from the preliminary works,
intelligent systems now deserve to be in the
knowledge briefcase
of each engineer, economist,
agronomist, scientist, etc.
together with, and at the same acceptance level of
several other basic
techniques like algebra, statistics, geometry, etc.
A way to reach a widespread industrial acceptance is to avoid using statements like:
I have used/developed a neural network for...
but, instead:
I have just developed a complex system with interacting signal pre-processor, neural network, user
interface, a differential equation solver, a post processor, some sensor and actuator interface, etc.
The major difference between the two approaches is which element(s) of a system receive(s)
more attention by the designer. In the former statement, attention (therefore the design
effort) is stressed on the presence of a neural network, which therefore improperly becomes
the most relevant block. In the latter statement, the neural network takes its proper place,
that is, at the same level as all the other system elements. In many cases, the blocks
surrounding intelligent subsystem are the most complex to design and use.
An example for this is in the field of image processing and handwriting recognition, where a
successful application relies much more on a proper image pre-processing (filtering, contrast
enhancement, segmentation, labelling, skeletonization, etc.) than on the neurofuzzy
processing itself.
