Hello,
I am very happy to see that the Rprop neural network training
algorithm is now available in the neuralnet package.
I have a few questions about the package and a few feature
requests also.
According to the documentation the startweights argument
of the neuralnet function is
"a vector containing starting values for the weights.
The weights will not be randomly initialized."
If the weights of the network are not initialized randomly, do
I have to initialize them explicitly using this argument ?
If the answer is yes, how to initialize them? (None of the
examples in the documentation uses the startweights argument.)
The err.fct argument of the neuralnet function may have the
special "ce" value, that stands for cross entropy minimization.
I think that it is useful if the output layer consists of softmax
neurons. Are softmax output units supported?
My feature requests are the following. (I think that it would
be easy to implement them. Actually, I have done it myself.)
The plot.nn function accepts rep="best". It would be convenient
to pass par="best" to the function compute also that currently
accepts only an integer as the value of the rep argument.
Finally, it would be nice to store the number of the best
repetition found as a component in the neuralnet object (for
examlpe, as $rep.best). This can be accessed only as
rep.best <- which.min(net$result.matrix["error",])
currently, which is quite technical.
Best regards,
Peter Jeszenszky