How does removedelay work in NARXNET?

Hi,

In the description of NARX modeling of time series shown in the following address, http://www.mathworks.com/help/nnet/gs/neural-network-time-series-prediction-and-modeling.html , it says for 1 time step early prediction, it removes one delay. I am trying to understand how removing one delay from each of the tapped delay lines makes the output of the network for y(t + 1) instead of y(t). Also, in the prediction result ys, is the last predicted value is the predicted target value for last external input?

Answer 1

[Greg Heath]

When creating timedelaynet(ID), narnet(FD) and narxnet(ID,FD)

 ID >= 0 and FD > 0

otherwise, an ERROR will occur.

Using removedelay(net,n) subtracts n from ALL of the components of ID and FD. The syntax removedelay(net) assumes n = 1.

As expected, an ERROR will occur if n > min(ID).

Similarly, one would expect that an ERROR will occur when n>= min(FD). However, there is no error when n = min(FD) !!!

Since n=min(FD) results in a forbidden zero feedback delay, I assume THIS IS A BUG AND SHOULD BE INVESTIGATED.

Hope this helps.

Thank you for formerly accepting my answer

Greg