Usage: |
{inp,net,err} = rbftrain2(x,y,clusters,learn,epochs,mMSE{,activ})
|
Input: |
| x | (n x p) matrix, predictor variables
|
| y | (n x q) matrix, response variables
|
| clusters | scalar, number of clusters to be built in the hidden layer
|
| learn | (3 x 1) vector, minimum learning rate and maximum learning rate
for building the clusters respectively, learn[3] is the learning
rate for training the output layer. learn must be from (0,1)
|
| epochs | (2 x 1) vector, number of epochs to train the cluster and output
units respectively
|
| mMSE | scalar, minimum value of the mean squared error to stop the training
|
| activ | optional scalar, 0 for binary (default) or 1 for bipolar activation sigmoid function
|
Output: |
| inp | list of input parameters which define the network; it consists of elements
inputs (scalar, number of predictors),
clusters (see Input),
outputs (scalar, number of responses),
samples (scalar, number of samples),
learn, epochs, mMSE and activ (see Input). |
| net | list of network characteristics containing:
clustersWeights ((clusters x inputs) matrix, weights of the hidden layer),
trFonDev ((clusters x 1) vector, deviance of the transfer function),
outputsWeights ((outputs x clusters) matrix, weights of the output layer) and
bias ((outputs x 1) vector, weights of the bias). |
| err | list of output error functions containing:
MSE ((epochs[2] x 1) vector, mean squared error for the network in each training epoch),
maxDiff ((epochs[2] x outputs) matrix, maximum difference between the real
and predicted output in each output unit for each training epoch) and
meanDiff ((epochs[2] x outputs) matrix, mean differences in every training epoch). |