Why should I use cross-correlation and auto-correlation to determine the number of delays in a NARX neural network?

I'm working with a NARX network to model the response of a dynamic system. I have the data for both the input signal and the system response. In trying to figure out the appropriate number of delays that I need to use (both input and feedback delays), I have come across several references to cross-correlation and auto-correlation. As I understand it, I would pick the delays that correspond to the highest peaks in the auto-correlation and cross-correlation plots, as they are more statistically significant than the others. What I'm not understanding is why is it appropriate to use that in the context of neural networks?

Answer 1

[Greg Heath]

It doesn't matter if it is a neural network or any other nonlinear regression model.

There tends to be a high probability that inputs with significant linear input-output correlations have a significant effect on outputs when a nonlinear regression model is used.

Similarly, there tends to be a high probability that inputs with insignificant linear input-output correlations have an insignificant effect on outputs when a nonlinear regression model is used.

Hope this helps.

Greg

[chanbeom Bak]

I can't understand that. Could you explain again?