Why the mean square error value is changing for different training functions ?

I have used the training function "traingdm" in training my neural network.

_

%%%%%%%%%%%%%%%%%%%%%

net=newff(minmax(in'),[20,1],{'tansig','purelin'},'traingdm');

%%%%%%%%%%%%%%%%%%%%%

_

What will happen if I change the training function from "traingdm" to "traindx" ?

Does the change affect the mean square error ?

Answer 1

[Greg Heath]

If the problem is not a trivial one do not expect the weights and MSEs to be the same.

To make a valid comparison you have to reset the random number generator to the same intial state(e.g., seed = 0, rng(seed)) to obtain the same default random data division and random weight initialization.

Different minimization algorithms cannot be expected to reach the same local minimum when started from the same initial point.

When you reset the random number generator and make new trial runs, please post a summary of the results. I'm interested in the % of times the results are the same.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Mohan]

Thank you for your reply Greg. When I reset the random number generator to the same initial state and make new trials, I am getting the same result again and again, even if I run it N number of times.

Only when I change the training function, I could see the change in the result. Even then, if I reset the random number generator to the same initial state, then the results are again the same. ( for that training function ).