The performance of the pattern classifier depends very much on the relevance of the features used in the input vector. There are various methods to investigate how relevant a feature is for the discrimination of a class, all of them seeking to eliminate irrelevant information in the input vector, consequently reducing the size of the system and the number of calculations involved in the process. This work presents new results obtained using the criteria of relevance described in Ref. [12]. This criterion is based on the search for changes in the answers given by the network when a feature used is substituted by its average value. The larger the difference between the replies given by the network, the larger the relevance of the feature [22]. In our previous works [14,15], this criterion was used to evaluate six features initially
employed, but then the classifiers were linear, implemented only with a neuron of the hyperbolic tangent type. Note that the criteria of relevance is calculated using the equation [14,22]: