Chapter Evolving Connectionist and Fuzzy Connectionist Systems: Theory and Applications for Adaptive, On-line Intelligent Systems

5.3. Using positive examples only.
In many practical applications only positive examples are available for a certain
class and no negative examples; or the negative examples may be too many that
creates problems of statistical incorrectness of the training procedure. Should
10,000 class objects (each of them represented by, say 1,000 examples) be used as
negative examples to train a module to recognise just one of these objects? What if
we have already trained a NN module on both positive and negative examples, but
currently data has become available about a new class object and the examples
about the other class objects (negative example) are lost.
Here, three EFuNNs are evolved on the Iris data by using positive examples
only. The following are their characteristics: SThr=0.85; Errthr=0.05; rn(setosa) =
6; rn(versicolor) =16; rn(virginica)=20. Overall correct classification rate: Setosa -
50(100%); Versicolor - 48 (96%); Virginica - 46 (92%).
In this case only rule nodes that support the "yes" fuzzy output node have been
created. That results in less number of rule nodes created (evolved), but also in a
higher false positive activation of the 
EFuNNs when similar data, but from
different classes, are presented. As a general rule, the true positive activation is
higher than the false positive one. By taking the maximum activated module as the

---

125
winning class, when an input vector is presented, even better classification than
the one achieved in 5.1, where both positive and negative examples were used, can
be achieved.
5.4. Cascade eco-learning.
This strategy was explained in section 2. Here we assume that data, once used, is
lost forever. When a new class data arrives, a new class EFuNN module is created.
It starts to evolve on the positive data of this class, as well as on following
negative data about other classes, and on the negative exemplars stored in the W1
connections of the already evolved EFuNNs (before this module was created).
The following are the parameters of the three evolved 
EFuNNs from the Iris
data: SThr=0.85, Errthr=0.1; rn(setosa) = 19 (4 positive); rn(versicolor) =32 (9
positive); rn(virginica)=27 (12 positive). Overall classification: 
Setosa -
50(100%); Versicolor - 50 (100%); Virginica - 50 (100%).