| Summary: | Dctennining the large ofsample sia is hig}ly contexr dependent- In genemt, the
larger the din)ension of paramel€r, the larger sample size must be to obtain a given deeree of
approximatio.. In nan! cases ofNN application, the data set is randomly split inio two
mnrually exclusive subs€rs, i.e. lraining and lesriog sets. The firsl is used for model building,
\vhile the second is used ro assess lhe pe.formanca (sefleralization) ofrhe model- Both training
and resting sets are the sarne size. In this paper, frve difTerent datn panirioning is utilized to rest
vheih€r the prcdiciion ability of NN is affecred by rb€ number ofobservalion in taining set.
The purpose oflhis papcl is to cvduate rhe effect of mlio betwe€n rraining nnd t€sring s€ts ro
the ftisclassificalion rate in neu.al networks (NN) model. An e pirical sludy has been done by
using Fishefs
jris
data. The results show ihat the hrhclassification ratc decrease when the
number oftraining set increase. Model with 2 hidden neuron obtains minimum error when the
ratio oftraining set is 20%. Wher$s model with I hidden neuron yieldq rhc minirnum eror ar
th e .atio of rrain ing set 400/0, 50o/a, 600/0 and 80d/o
|
|---|