# How to use ftinet to create neural network

1 view (last 30 days)
Das Bably on 21 Nov 2013
Answered: Greg Heath on 1 Dec 2013
Hi, I am doing a regression analysis using matlab Neuralnetwor for the verson2009, where I need to correlate 3 input data sets of(3*10) matrix with a target data set of (1*10)matrix. At first I have tried to use nntool and nftool. Now I am trying to do using command line by creating network with fitnet, but I am getting an error "??? Error using ==> load Unable to read file InputNeg: No such file or directory. Error in ==> math1 at 14 inputs = load ('InputNeg');" When I am trying to load the Input and Target Mat file. Then I tried to write the values directly in the code then I got the error "??? Undefined function or method 'fitnet' for input arguments of type 'double'. Error in ==> math1 at 17 net = fitnet(hiddenLayerSize);" Please If any one have any Idea about this, suggest me what should I do. Thanks in Advance
Greg Heath on 30 Nov 2013

Greg Heath on 1 Dec 2013
This is a poor example: N = 10 is not large enough to obtain robust practical solutions. With [I N ] =size(x) = [ 3 10 ], [O N ] = size(t) = [ 1 10], Ntst = round(0.15*N) = 2, Nval = Ntst =2, Ntrn = N-Nval-Ntst = 6, Ntrneq = Ntrn*O = 6 is the number of training equations.
With H hidden nodes, the number of unknown weights is Nw = (I+1)*H+(H+1)*O. A sufficient condition for a robust practical solution is Ntrneq >> Nw. However, Notice that, Ntrneq > Nw only when H <= Hub where Hub = -1+ceil( (Ntrneq-O) / (I+O+1)) = 0.
There is quite a bit more that I could say about this poor example. However, the main points are
1. I was able to run your code
2. Be careful of overfitting (H too large and/or N too small) with your real data set.
Hope this helps.
Thank you for formally accepting my answer
Greg