Experiences and Obstacles in Industrial Applications of Intelligent Systems - Intelligent Systems - page 118

Yet this has to be done in the most appropriate way, that is, by showing industry
unambiguously if, where and when intelligent systems offer significant advantages or, more
realistically, more advantages than drawbacks and associating the intelligent system with an
appropriate development environment and enough supporting tools , which is often the most
time consuming element to be developed.

This is one of the major reasons for the several Special Sessions on Industrial Applications of
Intelligent Systems which have been held in the last decade. Authors are usually requested
to present their ideas on intelligent systems but, more important, to prove that they are
either comparable or significantly better than other standard techniques. Such a comparison
has to be as fair as possible, as it is not normally the case . In practice, in most papers, intelligent
systems are usually compared among themselves. The expert reader is left with the
question:

Are you sure that other techniques would not be even better or simpler?

Or, when a comparison is attempted with standard techniques, these are usually much
older, that is, the paper demonstrates, for instance, that an up-to-date neurofuzzy network is
much better than an older-than-my-father standard technique , which is rather obvious, as
technology keeps improving, independent of intelligent system.

One of the major reasons for this lack of fair comparisons is that comparing an intelligent
system against an up-to-date standard method requires developing by scratch an appropriate
demonstrator, which often requires either a lot of specific experience or a lot of time, and
usually nobody wants to afford it.

Only those research groups who tightly cooperate with an industrial group can merge industrial and
academic experiences, to develop both techniques appropriately, although these are seldom done
together, due to unaffordable additional costs.

5.3 A critical question

So far very few applications of intelligent systems provided such better performance with
respect to other techniques to really convince even the most sceptical user. In most cases,
they can either offer a slightly better performance (when compared with an alternative well-
designed method) with a shorter design time but, on the other hand, design risks are often
so critical that they definitely impair the advantages. It is therefore time for a critical
question:

In which applications are neural networks have fuzzy logic a higher chance of being accepted?

We think that, at present, the most promising areas are, for instance:

 data mining , knowledge based systems , where information, data, knowledge and models

are valuable items , but they are often hidden in a huge amount of noise, ambiguous,
contradicting data. Data is so wide, contradicting, ambiguous, that no method can be
accurate and predictable, therefore neural networks may provide advantages, without
the need to be 100% correct;

 prediction/classification of partially random processes , like time-series prediction , forecasting ,

complex pattern classification , semantic Web , etc., where the randomness of the