Closedloop Recurrent Neural Network problem

My closed loop network response is random numbers between +ve and -ve values while the open loop network performance reaches to 1e-12.
I have a model with 9 outputs and 7 inputs. I generate the outputs using series of sine functions with different frequencies. I measure the input from the inverse model (which I have), and the model has first/second order differentiations.
Is the problem in my data? cause I read many answers here about problems in closedloop responses differ than the openloop?

I appreciate any help.

Answer 1

[Greg Heath]

I have 3 suggestions.

1. Only use significant delays obtained from the target autocorrelation function and target/input cross-correlation function.

2. Minimize H given ID, FD, and MSE <= 0.005*mean(var(target',1))

3. Do not use the default dividerand. Instead, use divideblock or divideind.

Hope this helps.

Greg

[Ala Abd]

1. Is there a reference on how to do this, sorry my background on this is not that much? I really appreciated.I saw a previous answer for you on how to find the autocorr but don't know how to use it to find the significant delays.
2. I learned that from your answers.
3. Also, using it based on your answers.

Answer 2

[Greg Heath]

I know this is not a good answer, but I and others are also having closeloop performance problems. So far I haven't found out why.

Greg

Answer 3

[Greg Heath]

Assume a 1-dimensional training target of length N

 1. xcorrtn = nncorr(zscore(ttrn,1),zscore(randn(1,N)),N-1,'biased');
 2. Sort the absolute values
 3. Find the value 95% of the way to the end.
 4. Repeat 100 times and take the average

Use that value as the 95% confidence threshold value for the significant auto and cross correlations of ttrn.

Greg