DataLab is a compact statistics package aimed at exploratory data analysis. Please visit the DataLab Web site for more information....

RBF Network - Calculate Model

Command: Math -> Neural Netzwork -> RBF Network -> Calculate Model...

The calculation window for RBF networks shows three areas (from left to right): the parameter section, the plot of the residuals, and the plot of the estimated vs. the actual target values. The user may set up the following parameters:

Hidden N. The number of the hidden neurons. Note that the number of hidden neurons should be 10 % or less of the number of observations in the training data set.
R The parameter R determines the flatness of the top of the basis function. For continuous models R should be zero, for classification purposes R may be set to values greater than zero (typically between 10 and 1000).
S The parameter S determines the slope of the basis functions. The smaller the value of S the more the basis functions are overlapping. Typically S is set to values between 0.01 and 0.5 (S must not be zero).

The command Training starts the training of an RBF-network. The number of hidden units and certain other design parameters must be set up prior to the start of the training. After activating Training the user has to specify the variables which should be used as inputs and as target values of the net. The network can be of any input size. The user first has to select all input variables, followed by choosing the target variable (the variable which should be estimated by the neural net).

After the target variable has been selected, the network is trained, which may last between several seconds and a few minutes (depending mainly on the number of training data, the number of hidden neurons, and the type of computer DataLab is running on).

The results of the training are shown in graphical from. After the training has been completed, the user may read the protocol, or may store the trained network on the disk. The trained and stored network can be reloaded at a later time and applied to other data (see command Apply Network).

Last Update: 2012-Jul-25