There are several inexpensive methods that should be suitable for
initialization of neural classifiers, including statistical discriminant
functions, decision trees such as C4.5 or CART, and clusterization
methods. We will consider statistical discriminant and clusterization
techniques here, although the use of decision trees is equally worth
investigating. Classical statistical discriminant techniques include
Fisher, linear, logistic, multiple discriminant analysis, and more
modern methods like DIPOL92 [10]. The Fisher discriminant analysis
[11] is based on a projection on a single line passing through the
origin, trying to increase the separation of projected points between
the classes and decrease the separation of projected points inside the
same class. The error function becomes: