Train more a Neural Network to have more consistent results every time?

Hello

Before using Matlab, i was programming my own Neural Networks, and when i trained them long enough, i had the same simulation outputs everytime (approximatively).

Now i use Matlab and on my predictions, i don't always have the same consistent results.

So do i need to train the NN longer?

And how do i do that properly?

thanks!!

Jeff

Answer 1

[Greg Heath]

Results are different because the initial state of the RNG is different. Therefore you will get different random data divisions and initial weights.

For nontrivial problems this often results in a wide variety of results.

It is not unusual for me to designs 100 nets in a double for loop:

 1. Initialize the RNG

 2. Outer loop over 10 trial values for number of hidden nodes: 

 h = Hmin:dH:Hmax

 3. Inner loop over Ntrials = 10 different initial RNG states.

 i = 1:Ntrials.

I have posted zillions of examples in both the NEWSGROUP and ANSWERS. Search using "greg" with the above code snippets.

Hope this helps.

Thank you for accepting my answer

Greg