addr | creates one hidden layer network |
ann | is a tool to run a feed-forward neural network |
committee |
This macro computes a committee of networks with nets of
the form single
layer feedforward perceptron. The macro can be used alone
or in connection with the library ISTA. The standalone version
also needs the parameter data. Just choose 0 for the input.
The number of nets to build the committee can be chosen.
The data will be splitted with this number to build the
different datasets. The weight for the cases for the training
of the net can be chosen, the numbers of hidden units and
additional information concerning the weights of the units.
Different optional parameters can be chosen to
build the architektur of the network. The choice holds for
every single net. The default values are chosen in
order to solve a linear regression problem. The optional
parameters constits of 8 values. Boolean values for linear
output, entropy error function, log probability models and
for skip connections (direkt links). The fifth values is
the maximum value for the starting weights, the sixth is the
weight decay, the seventh the maximum number of iterations
and the the last value generates the output concerning the
architekur of the net if it is equal to one.
The output consits of the Error and MSE of the single nets
and for all cases. Additionally the R^2 for the average of
the nets and the R^2 of the committee are shown.
|
cv | runs a cross validation over the hidden units |
cvdec | runs a cross validation over the weight decay |
erfkl | Kullback-Leibler criterion for classification |
erfqua | (1-R^2) criterion for regression |
finalshow | shows the final visualization of the network |
gennet | generates interactively a feedforward network |
neuronal | This macro computes different networks of the form single layer feedforward perceptron. The macro can be used alone or in connection with the library ISTA. The standalone version also needs the parameter data. Just choose 0 for the input. It is possible to split the data in a training and a test set. The weight for the cases for the training of the net can be chosen, the numbers of hidden units with ``from, stepwidth, to'' and additional information concerning the weights of the units. Different optional parameters can be chosen to build the architektur of the network. The choice holds for every single net. The default values are chosen in order to solve a linear regression problem. The optional parameters constits of 8 values. Boolean values for linear output, entropy error function, log probability models and for skip connections (direkt links). The fifth values is the maximum value for the starting weights, the sixth is the weight decay, the seventh the maximum number of iterations and the the last value generates the output concerning the architekur of the net if it is equal to one. The output consits of the Error and MSE of the different nets (MSE for test and trainings data separately if chosen) and the R^2. |
nninfo | shows some information about the actual network |
nnlayer | builds a feedforward network |
nnmain | loads the necessary libraries |
nnrinfo | shows information about the net in the output window. |
nnrload | loads a network from different files |
nnrnet | trains a one hidden layer feed forward network. The optional parameter param consists of 8 values. Boolean values for linear output, entropy error function, log probability models and for skip connections. The fifth value is the maximum value for the starting weights, the sixth the weight decay, the seventh the number of maximal iterations and the last value generates some output if equal to one. |
nnrpredict | estimates for a given net and dataset the response. |
nnrsave | saves a network in different names |
optdec | runs for each set of observations a neural network to estimate the generalization error |
readshow | shows the visualization of a feedforward neural network |
resclass | shows the residuals in case of the classification |
resreg | shows the residuals in case of the regression |
runcv | runs a cross validation and estimates the generalization error |
runinit | initializes the training andtest dataset, the errors and the weights in the network |
runnet | runs a network with prespecified optimization method |
runnew | optimize a neural network by a quadratic approximation |
runqsa | optimizes a neural network by a stochastic search |
runsa | optimizes a neural network by Boltzman annealing |
runshow | visualizes a neural network during optimization |
weidist1 | transforms weights in distances (\delta^{(2)}) |
weidist2 | transforms weights in distances (\delta^{(2)}) |
weidist3 | transforms weights in distances |
weinit | initializes the weights of a neural network |
x3matrix | constructs a matrix linke in XploRe 3 |