narx network producing different outputs for same input

Hi, i have used narx network for multistep ahead prediction. each time i run the code with same input i get different predictions. how can i solve this problem? please can u suggest me some tips to improve the performance of narx network?

Answer 1

[Greg Heath]

If you are reusing the same trained network, the difference is probably caused by different initial delayed values.

If you are retraining new networks, the difference is probably caused by a different initial state of the random number generator.

Hope this helps.

Greg

[shradha yelshetty]

i have created an narx network with series parallel architecture and then i'm training it . once its trained i'm converting it to parallel architecture with the help of closeloop command and using it for multistep ahead prediction. according to your answer i think the problem is caused by the first reason. can you suggest something for solving this problem......

[shradha yelshetty]

apart from this i'm getting very sharp peaks which have very high magnitude when compared to original targets in my output.....how to optimize my output....

[Greg Heath]

I'm on vacation until June 5th. Therefore I don't expect to be able to spend time on your problem.

Very sorry,

Greg