how can find the optimal delays and number of hidden nodes in narnet for forecasting task?

I have tried with these code but I haven't found solutions:

1/ Optimal timelags?

N = length(Target)

zy = zscore(Target,1);

autocorry = ifft( abs(fft(zy)).^2 )/N

%AUTOCORRELATION

ZT=zscore(Target,1);

autocorrT = nncorr(ZT,ZT,N-1,'biased')

figure(3)

plot(autocorrT)

%title('ACF')

2/Optimal hidden nodes???

http://www.mathworks.com/matlabcentral/answers/111521-how-to-plot-the-nar-predicted-values-in-matlab

%To find H

[I ,N]=size(Input);

[O ,N]=size(Target);

Neq=N*O;

Hub=floor((N-1)*O/(I+O+1)) %max H for Neq>=Nw

Thank you in advance.

[coqui]

Optimal hidden nodes???

 I have 1413 observations: the targetseries size

 trn/val/tst ratios = 0.8/0.10/0.10 

 ID = 0:2

 Ntrneq = No. oftraining equations = prod(size(ttrn)) = Ntrn*O

where

 [O ,Ntrn]=size(trainTarget);

 O=1 
and

 Ntrn=(1413-2)*0.8= 1129

therefore

 Ntrneq = Ntrn*O=1129*1=1129

and

 Nw = No. of weights to be estimated 

 Nw = (MXFD*O +1 )*H + (H+1)*O 

 Nw = (10*1 +1 )*H + (H+1)*1 % is H =1413*0.8=1130??? and why we employed 

   MXFD=10???

Supposing that H=1130 and MXFD=10:

 Nw = (10*1 +1 )*1130 + (1130+1)*1=13561

and the requirement Ntrneq > Nw yields the upper bound

 Hub = -1 + ceil( (Ntrneq-O) / ( MXFD*O + O +1) )

 Hub = -1 + ceil( (1129-1) / ( 10*1 + 1 +1) )=93

 Hmax = floor(Hub/10)=9

As 9 the number of hidden nodes???

[Greg Heath]

If your calculations are correct, I would perform my typical double loop search to find the smallest value of H <= 10 to give an acceptable solution.

[coqui]

By using Narxnet, if I found zero is the significant delays of the input-target cross correlation function how I must put the code:

inputDelays= [ ] ???????

[coqui]

can you help me by putting some code to determine inputDelays in narx.

thanks

[coqui]

we can use:

cor_seq=xcorr(X,T)% X is input and Y is the target

after, we can plot the cor_seq vs lags and we can find the lag which can define the symmetric part of the plot?????

Is right??????

[Greg Heath]

Why are you posting questions about NARXNET in a NARNET thread ?

Search the NEWSGROUP and ANSWERS

                 NEWSGROUP   ANSWERS
narxnet nncorr       7         42
narxnet xcorr        2          8
narxnet fft          1         13

Read the most recent NEWSGROUP posts first.

I searched the NEWSGROUP threads listed above. Only the "narxnet nncorr" search was fruitfull. Of the 7 threads, only 3 were useful. However, 8 messages in those threads are useful:

 1. Subject: SIGNIFICANT AUTOCORRELATION DELAYS OF THE SIMPLENAR_DATASET
 Date: 31 May, 2015 
 Messages: 12, 15, 19, 23
 2. Subject: NARNET TUTORIAL ON MULTISTEP AHEAD PREDICTIONS
 Date: 17 Dec, 2014 
 Messages: 1 , 7, 11
 3. Subject: open-loop to closed-loop Narx NN in Matlab Help 
 Date: 21 Oct, 2012
 Message: 4

Hope this helps.

Greg

[coqui]

Thank you very much Greg.

Answer 1

[Greg Heath]

I have many posts regarding this.

Search both the NEWSGROUP and ANSWERS using subsets of the following search words

 greg narnet nncorr siglag95 Hub Ntrials tutorial

Hope this helps.

Greg

[coqui]

Thank you Greg,

By using narxnet, I have found that the input/target crosscorrelation function is symmetric about zero lag. how I can put the code of the inputDelay.

Thank you.

[Greg Heath]

NARXNET???

I think you mean NNCORR ...right?

Crosscorrelation functions are not symmetric. That is a NNCORR BUG.

Use the nonnegative lag portion.

Use the smallest subset of significant lags that will yield sufficient results. It is a trial and error process.

Which of my posts are you following?

 NEWSGROUP:
 ANSWERS:

[coqui]

Thanks Greg. We can not use this:

cor_seq=xcorr(X,T)% X is input and Y is the target

after, we can plot the cor_seq vs lags and we can find the lag which can define the symmetric part of the plot?????

Is right??????

[Greg Heath]

No.

1. Find all NONNEGATIVE lags where the absolute value of the input/target crosscorrelation function exceeds a specified threshold.

2. For a threshold I use an estimate of the 95 per cent significance level of Gaussian noise.

3. Then choose a subset of the smallest significant lags to include in the design. For example, if the smallest significant input lags are 0 2 3 6 8, using ID = 0:3 might be a reasonable first try.

4. Similarly for the target autocorrelation function and significant feedback lags EXCEPT, the feedback lags must be POSITIVE.

5. Finally, given the choice of delays, use the double for loop (h = Hmin:dH:Hmax, i = 1:Ntrials) approach to determine the smallest number of hidden nodes that yields either an acceptable, or the best design.

Hope this helps.

Greg