[NEURAL NETWORK] How can I make MSE be the same in different trying times ?

The Mean Square Error (MSE) turns out to be different each time I train the network.

_net = feedforwardnet(10,'trainlm');

net.divideParam.trainRatio=1;

net.divideParam.valRatio=0;

net.divideParam.testRatio=0;

net.trainParam.epochs=10;

[net,tr]=train(net2,x,t);

MSE=sumsqr(net2(x)-t)_

I guess it come from net2.initFcn caused this problem. How do I change it to make MSE becomes the same ?

Answer 1

[Greg Heath]

Initialize the state of the random number generator ONCE AND ONLY ONCE before the random data division and random weight initialization.

In the following example, The design is not repeatable if either or both the first two commands in the loop are commented out.

 clear all, clc
 [ x, t ] = simplefit_dataset;
 MSE00    = mean(var(t',1)) %  8.3378 Reference MSE
 net      = feedforwardnet; % Defaults used

 for i = 1:2
    rng('default')                     % Reset RNG state
    net          = configure(net,x,t); % Weight initialization
    [net tr y e] = train(net,x,t);     % y = net(x), e=t-y
    NMSE     = mse(e)/MSE00            % 1.7558e-05
    ttrn     = t(tr.trainInd);
    MSEtrn00 = mean(var(ttrn',1))      % 8.8219
    NMSEtrn  = tr.best_perf/MSEtrn00   % 1.386e-05
    NMSEval  = tr.best_vperf/MSEtrn00  % 1.0019e-05
    NMSEtst  = tr.best_tperf/MSEtrn00  % 3.6061e-05
end

% P.S. Look at the results of the following command

 tr = tr % No semicolon

Hope this helps.

Thank you for formally accepting my answer

Greg

[Son Nguyen]

Appreciate very much. It really helps. Could you please help me a favor? How can I make it returns MSE per iterations ? The current MSE is applied only when training if finished.

[Greg Heath]

Obviously, you did not follow my advice to see what goodies tumble out of

tr = tr

[Son Nguyen]

I apologize that I missed it, I got it already.

I still has one more question. Is there anyway to take out the initial weights of training ? I'm going to compare my algorithm with LM and BFGS. In order to do that, I have to make they have the same initial weights.

If you feel not comfortable to answer my question here, please let me know, I will make a new question for it.

Appreciate very much for your help.

Answer 2

[Greg Heath]

Remove the underscores at the beginning and end.

Use net.divideFcn = 'dividetrain' to automatically get the (1/0/0) data division

I hope you don't use 10 epochs for serious work

net2 in never defined

MSE depends on net, not net2

sumsq yields SSE, not MSE

For repeatability you have to set the initial state of the random number generator before data divisions and weight initializations.

MSE is scale dependent. Better to normalize by the average target variance

Hope this helps.

Thank you for formerly accepting my answer

Greg