Initialization Functions

  The goal in initializing a radial basis function network is the optimal computation of link weights between hidden and output layer. Here the problem arises that the centers (i.e. link weights between input and hidden layer) as well as the parameter p (i.e.\ the bias of the hidden units) must be set properly. Therefore, three different initialization procedures have been implemented which perform different tasks:

1. RBF_Weights: This procedure first selects evenly distributed centers from the loaded training patterns and assigns them to the links between input and hidden layer. Subsequently the bias of all neurons (parameter p) inside the hidden layer is set to a value determined by the user and finally the links between hidden and output layer are computed.

2. RBF_Weights_Redo: In contrast to the preceding procedure only the links between hidden and output layer are computed. All other links and bias remain unchanged.

3. RBF_Weights_Kohonen: Using the self--organizing method of Kohonen feature maps, appropriate centers are generated on base of the teaching patterns. The computed centers are copied into the corresponding links. No other links and bias are changed.

It is necessary that valid patterns are loaded into SNNS to use the initialization. If no patterns are present upon starting any of the three procedures an alert box will occur showing the error. A detailed description of the procedures and the parameters used is given in the following paragraphs.

• RBF_Weights
• RBF_Weights_Kohonen
• RBF_Weights_Redo