For NARNET and NARXNET to be useful, target input values used in the openloop (OL) design have to be replaced by feedback from the output. This is accomplished via the CLOSELOOP (CL) command.
Very often, however, the resulting CL configuration is either not as accurate or is completely useless.
UNFORTUNATELY THERE IS NOTHING IN THE help OR doc DOCUMENTATION ABOUT THIS.
I have been struggling to devise a successful technique for improving the accuracy of CL designs. If you search both the NEWSREADER and ANSWERS with
you will see my efforts to resolve this issue.
All I can recommend at this time is the following
1. Try to obtain the lowest error rate possible with as few delays and hidden nodes as possible using the OL configuration. The fewer weights you have to estimate, the higher your probability of success.
2. The default values ID = 1:2, FD = 1:2, H = 10 are very often inadequate. Nevertheless, since you may have to make tens or even hundreds of OL designs before obtaining a successful CL design, I suggest that you ALWAYS start with the defaults.
a. Successful input delays tend to be subsets of the significant delays
obtained from the statistically significant peaks of the crosscorrelation
functions of paired inputs and outputs.
b. Successful feedback delays tend to be subsets of the significant
delays obtained from the statistically significant peaks of the target
c. Success usually results from choosing a reasonably sized subset of the delays
found in a and b.
4. For each trial combination of delays, you can use a double loop design search over number of hidden nodes (outerloop h = Hmin:dH:Hmax) and initial random weights (inner loop: i = 1:Ntrials).
For examples search the NEWSREADER and ANSWERS using the search words
5. For unbiased future predictions the test data subset data must occur after the training and validation subsets. The natural datadivision option to use is DIVIDEBLOCK. However, other trn/val configurations using divideind can certainly be used.
6. After achieving success with at least one OL design, use the CLOSELOOP command to replace target inputs with output feedback.
7. Unfortunately, when the loop is closed, very often the predictions are not as accurate or even completely useless.
8. The only other remedies I have to offer are
a. Try CL designs from all of the other successful OL designs
b. Train the CL configuration(s) initialized with the weights of the OL configuration(s).
c. Try designing CL designs from random initial weights.
Hope this helps.
Thank you for formally accepting my answer