don't use random train of som network

train for som network use random method (the results differ for the same input data) thought its net.trainFcn = trainbu (not trainr). How to get the same results? I tried chaneged to trains but failed.

Answer 1

[Greg Heath]

If you initialize the random number generator to the same state, the results will be duplicated.

 help rng

 doc rng

Hope this helps.

Thank you for formally accepting my answer

Greg

[Alexander]

Thanks, I saw your answer but I want to know more. 1. Why trainbu use random learning? 2. How to use trains? With train or adapt? 3. How to use trainc? 4. For random numbers is there convergence if we use much more epochs? (I used 2000 but the results for 14x14 som network are different.

[Greg Heath]

If you enter the command

net = selforgmap %NO SEMICOLON

you can obtain all of the default properties and settings.

The result depends on the order of processing the points. Randomness mitigates the possibility of not finding a good solution provided you create enough independent designs. Just keep training a single design will probably not improve performance because you may be lying in a deep local min that is much higher than the global min.

Again, multiple designs with different random orderings is, in general, the best learning technique.