A hybrid neural network with fuzzy rules for categorical and numeric input

This paper is concerned with the architecture and training of a hybrid neural network that may be used to represent a function that has both numeric and categorical independent variables. The numerical variables are separated from the categorical variables. It is assumed that the function to be represented really consists of several functions whose independent variables are the numerical variables. The hybrid network consists of two networks, the FFNN and the CIN (categorical input network). The FFNN accepts the numerical component as input and the CIN accepts the categorical input. Each produces a vector. These vectors are subsequently combined through the dot product to produce the final output of the combined network. The FFNN actually consists of several FFNN´s with an MLP for each function to be represented. In general the output vector of the CIN will be 1-of-n and in that case the categorical component of the independent variables in effect is used to select the output of one of the MLP´s and thus selects one of the functions represented. The approach suggested is shown to be quite effective.