MATLAB: Try to simulate neural network in Matlab by theself

cneural network

hi every one, i tried to create a neural network that estimate y = x ^ 2 so i create a fitting neural network and give some sample for input and out put. then i tried to come this network to c++. but the result is different i tried to find why this happened. i wrote this command in matlab :
purelin(net.LW{2}*tansig(net.IW{1}*in+net.b{1})+net.b{2})
and the result was 16.0828 for : in = 3
my network has 2 neurons. and other information are :
net.IW : 0.344272596370387 0.344111217766824
net.LW : 31.7635369693519 -31.8082184881063
b : -1.16610230053776 1.16667147712026
b2 : 51.3266249426358
so is any body have any idea why this has happened? thank you

Best Answer

I used fitnet(2) with all defaults except for
rng(0)
net.trainparam.goal = MSE00/100;
where MSE00 is the mse for the constant mean value (=1704) output model.
It converged resulting in
Nepochs = tr.epoch(end) % 5
NMSEtrn = tr.perf(end)/MSE00 % 0.0018431
NMSEval =tr.vperf(end)/MSE00 % 0.00068653
NMSEtst = tr.tperf(end)/MSE00 % 0.0016467
whereas
y3 = net(3) % 70.409 (instead of 3^2 = 9)
sqerr3 = (y3-3^2)^2 % 3771.1
Nsqerr3 = sqerr3/MSE00 % 0.0016237
So, to get a good feel for how well the net is performing
look at the normalized test set NMSEtst.
Hope this helps.
Greg
Related Question