Improve network generalization NarX

Very good I would do the following: divide my data into 10 parts and each train separately checking with other cells, is crosvalidation guess but I'm a little busy and I'm not sure how. I could explain a little further as performing autocorrelation, cross correlation and other steps to achieve better network generalization NarX and clarify concepts?

Thank you very much.

Answer 1

[Greg Heath]

I just posted an answer to your question on the NEWSGROUP

http://groups.google.com/group/comp.soft-sys.matlab/browse_frm/thread/542545cc25b4c179#

Hope this helps

Thank you for formally accepting my answer.

Greg

[FRANCISCO]

Good the truth that I have read the text several times, and tried raise it but I can not think how. My intention is to break the data into a number of parts and then training them go. Thus I believe that the improved accuracies. I used divideblock for the division of data on the network but when the train does not give me good results, makes me quite wrong. Thank you very much, any feedback will help me.

[Greg Heath]

If you would try your code on the polution_data set we can compare results. I have used the delays ID=1:2, FD=1:2 and H = 16 with dividetrain and MSEgoal = 0.08*Ndof*MSE00a/Neq. A lower goal will cause training to extend to maxepoch (default = 1000; I will change it to 100)). The results are R2a = 0.92 for openloop and 0.88 for closed loop.

I will be experimentinng with this data for some time: Linear trend removal, Significant delays, validation stopping and minimizing H. Not necessarily in that order.

Greg