About parameters for NARX

Hi all:

I have built a NARX with 2 time-series as inputs. I have seen that changing Number of Neurons and Number of Delays, accuracy changes from time to time (sometimes for better, sometimes for worst). Instead of trial & error approach, is there a way to estimate/to start with a good combination of Delays v/s Number of Neurones in order to improve the accuracy of the network?. Thanks for your suggestions.

Best regards.

Answer 1

[Greg Heath]

FD: Find the significant delays in the autocorrelation function of the target

ID: Find the significant delays in the cross-correlation function of the input and target.

Search for some of my example code

 greg nncorr thresh95

Hope this helps.

Thank you for formally accepting my answer

Greg

[Fran Mat]

Hi Greg: Thanks for this.

With your search parameters i got this one: http://www.mathworks.ch/matlabcentral/answers/77142-future-value-prediction-with-neural-network-method-and-right-input-and-target-format-data

which I think is not exactly the topic I am looking for. Remember: I have 5 series as input and 1 as output. Should I get the correlation five times? for 5 inputs versus 1 output?. Please clarify.

Thanks.

Fran.

[Greg Heath]

Yes. You will get 6 sequences of significant lags; 5 for ID and one for FD. However, you also have the magnitudes of those correlations. So, if your inputs and target are standardized (zero-mean/unit-variance), you can make a reasonable choice of which ones to keep. But remember that the sequence of ID lags you choose will be applied to all of inputs.