Elman Network with Logsig Output Neuron

I was trying to create a neural network for a classification task. I used the following line;

net=newelm(elman_input,elman_targets,[ 5,5,1], {'tansig','tansig','logsig'})

but this actually seems to create a network with 4 layers not 3.

size(net.layers) = 4,1

It seems matlab wants to autogenerate the output layer in case I get the dimensions wrong. The autogenerated output neuron has a purelin transfer function and works well.

When I change the output neuron to a logsig transfer function (more appropriate for my particular task) the network behaves extremely badly and I cannot train it (I tried traingdx and trainrp).

anyone else tried an elman network with a logsig output function?

I am keen to avoid the linear output because I want to use a custom performance function that will not respond well to negative outputs.

Answer 1

[Greg Heath]

net=newelm(elman_input,elman_targets,[ 5,5], {'tansig','tansig','logsig'});

The output dimension is automatically obtained from elman_targets.