Neural Network Toolbox NARX

I trained a data set with NARX and generated matlab matrix-only function for the trained network. So far so good!

Yet, I'm curious why the generated function still requires the "y" data set. Except the first a few delay, doesn't it use its own y output? In the for-loop of the generated function, there is this line:

% Input 2 (this input2 is the y label that I trained)

xd2(:,xdts) = mapminmax_apply(x2(:,ts),x2_step1);

Is it predicting the y based on actual previous y rather than previously predicted y value? If so, is there a way that the network uses its own previous prediction except for the first few delay?

Thank you

===========================================

Ok. I realized there's a closed loop net as well. However, its performance is drastically poorer than open-loop based on the actual label. Is there a possible way to augment this?

Answer 1

[Greg Heath]

MATLAB does not have an answer for what to do when the loop is closed and the resulting CL performance is OFTEN VERY MUCH WORSE than the OL peformance.

I have tried to continue training with the new CL configuration. This only works a very small percentage of the time.

I think I have posted in the NEWSGROUP which of the MATLAB EXAMPLES are easily trained with a simple CL command.

The only suggestion I have is to try to train the CL configuration from the beginning.

It would be nice to find a non-MATLAB discussion.