next up previous contents
Next: The U-matrix representation Up: Process data Previous: Data preprocessing

Training a SOM

Several Self-Organizing Maps were trained with a different number of codebook vectors and varying teaching parameters. Of all these SOMs, the one with the smallest quantization error was picked. This was the SOM with 25 times 15 codebok vectors. Bubble neighborhood was used as the neighborhood function. The learning rate was chosen to decrease linearly as a function of time. Training was done in two phases. During the first cycle, the initial alpha value tex2html_wrap_inline2199 , the map was taught with 35000 samples and the initial neighborhood had the value 15. The respective parameter values for the second cycle were 0.04, 150000 and 6.



Jaakko Hollmen
Fri Mar 8 13:44:32 EET 1996