When using the multinom function from the nnet package, how can I control the architecture of the neural networks?

Question

In other words, when I do nnet(...) I can use the size parameter to control the number of units in the hidden layer. My particular model requires outputting probabilities so I wanted logistic units and so I turned to multinom from the nnet package to output type='probs' in my predict function. How can I pass down a size argument? When I call it with, say, size=5 or something I get an error:

 formal argument "size" matched by multiple actual arguments

multinom appears to have hard coded size = 0 in its calls to nnet.default. I suspect there's a mathematical reason for this, but don't feel comfortable speculating further. — joran
library('nnet')..also, yes it seems to be size=0 which sounds like it's identical to just logistic regression! — Palace Chan
@PalaceChan: the difference is that multinom can handle factor responses with 3 or more levels, while logistic regression (as implemented in glm) is for binary responses only. — Hong Ooi

Hong Ooi Hong Ooi · Accepted Answer · 2012-04-02T02:05:33

multinom fits linear multinomial logistic models, which is why the size parameter is hardcoded to 0. If you want an actual neural network with multinomial outputs, just use nnet with a response with more than 2 levels, and set softmax=TRUE.

When using the multinom function from the nnet package, how can I control the architecture of the neural networks?

1 Answers