MATLAB: Finding adequate matlab neural criterion function

neural network

i have a feed forward neural network that is simulating some kind of mathematic function, can someone give me examples of matlab criterion function that will force results to be below targets. I know that this can be more effectively accomplished by multiplying error in trainlm but i want to do this with criterion function as much as possible and than to use multiplying in trainlm. Using mse gives 50%, 50% for results below target and above target, i would like to make it 90%, 10%.

Best Answer

>i have a feed forward neural network that is simulating some kind of mathematic function,
Curvefitting; not pattern recognition or classification. Correct?
What is the function?
>can someone give me examples of matlab criterion function that will force results to be below targets.
NO, NO, NO!
ANY typical training error criterion can be driven to 0 by just using enough hidden nodes, H.
HOWEVER, the most useful design goal is to minimize the expected error on NONTRAINING data that can be assumed to be randomly drawn from the same probability distribution as the training data! For most real-world problems, that minimum is nonzero.
Let Ntrn be the number of input/target training pairs with I dimensional inputs and O-dimensional outputs. The number of training equations is Ntreq =Ntrn*O and the number of unknown weights to estimate is Nw = (I+1)*H+(H+1)*O
A successful FFMLP with I-H-O node topology can usually be obtained in one of several ways. For example,
1. Use a relatively large amount of training data so that any random draw of nontraining data will have a similar probability distribution. Usually, Ntrneq >> Nw is sufficient, This results in the following limitation on the size of H.
H << floor( (Ntrneq-O)/(I+O+1) )
A satisfactory value for the ratio of the two sides of the inequality depends on the data. I usually search for the minimum value of H that yields a degree-of-freedom adjusted MSE that is 100 times lower than the average target variance.
I have posted many such examples.
2. Use a separate validation set to characterize the salient features of all nontraining data. Monitor the validation error during training and stop when the validation error is minimized.
Most, if not all, of the MATLAB examples include this option as a default.
3. Use regularized training with the training function trainbr.
Hope this helps.
Thank you for formally accepting my answer.
Greg
Related Question