SOM training/t​esting-sim​-trainr

I want to differentiate 3 sets of data, using SOFM. I am getting very absurd results. Sometimes the sample hits reverses. Are we supposed to use the 'sim' command to test a network? if yes, then even the data that is being used to train the network, when simulated, doesn't give more than 20-40% accuracy. How to go about it? Where could have I gone wrong?Does topology make a difference?I am using dimensions as [3] only, since 3 sets are there Also, I used wavelet decomposition to process my data, and selected some coefficients. Please Help.

PS: I am using trainr(by default from GUI - nntool). Thanking you in advance.

[greatest]

Please also explain me the role of learning fnc, ordering phase etc. Why should they be important? Is Hit-and-trial the best method to set these parameters?

Answer 1

[Greg Heath]

SOFM is not meant for classification. Therefore Kohonen extended it to LVQ (lvqnet).

However, the MLP (newff, patternnet) and RBF (newrb) and are based on universal approximators and, for me, are preferrable.

Hope this helps.

Greg

[greatest]

Thank you sir.

This shall benefit me. The only thing that bugs me is that I am following a paper on this project, and they have used newsom fnc. to classify the data. I am goin to follow your technique though.