How to improve the close-loop net performance for NARXNET?

Hi all, I'm very curious that even the training and testing data performance are very good, the close-loop net performance is still bad. I compared both the correlation and mse. The situation is the same. Anybody got any idea? Thank you.

[Greg Heath]

Search

 greg closeloop

Answer 1

[Greg Heath]

The bottom line is to drive the OL error as far down as possible.

Then closeloop and train on the same data. CL error will never be lower as OL error.

However, that is the best you can do.

Hope this helps.

Greg