How to customize performance function in MATLAB neural network.

Greetings. I am relatively a beginner in neural network. I started off using MATLAB neural network tool until I was met with some limitations. My problem is as follows: 1.) I would like to be able to customize my performance function such that I don't use the inbuilt performance functions such as mse, mae, etc. 2.) I would also like to be able to modify how many training iteration before any validation. I read somewhere that it takes five (5) training iteration before every consecutive validation check. However, if someone has an idea of a better neural network tool that is good for function approximation using MLP, please recommend.

Thank you.

Answer 1

[Greg Heath]

1. Start with a copy of mse (or other performance function)

   a. Enter the command: type mse 
   b. Copy and save the result under a different name, e.g., myperf
   c. Modify it
   d. Save it
   e. Change the net property
      net.perfornFcn = myperf;
   f. Test it.

2. net.trainParam.max_fail = 10; %(or whatever)

3. Enter WITHOUT THE SEMICOLON!

net = fitnet

You will see a long list of default parameters that can be changed to your liking.

Hope this helps.

Thank you for formally accepting my answer

Greg

PS: This is explained in the website documentation section on custom functions.

[Enrico Camporeale]

this will hardly work if you do not change the backpropagation function as well

[Arygianni Valentino]

Hi Greg..

If you do not mind, would you explain more and give an example of the step Modify the performance function? Since I got stuck there. I could not find the formula where the corresponding performance function does the calculation.

[Chibuzo Nnonyelu]

This approach did not actually work. I ended up writing my own MLP BP from scratch using C#. C# ensured faster training with larger data. And I was able to try my custom error and performance function.

Using MATLAB for ANN is only good if you are just fine with its available functions and all.