Page 191 -

5.5 Multi-Laver Percevtrons 179

Train
0.0
o 500 1000 1500 2000 2500 3000 3500 4000 - Verify
Epoch
Figure 5.26. Learning curve for the cork stoppers data (3 classes) with a
MLP7:5:3. Notice the error degradation for the verification set after approximately
500 epochs.

Choosing the best neural net solution can be quite a tedious job since one has to
try several architectures and starting conditions. As previously mentioned,
Statistics has a helpful IPS tool that allows the user to perform a series of
experiments and retain the best solutions, which are displayed at the end in a sorted
way. Using this tool it was possible to find an MLP2:2:3 solution with features N
and PRT and a much lower number of weights (15), which after training with 200
epochs performed similarly. This is shown in Table 5.5 for the training,
verification and test sets. In this and other examples the logistic sigmoid was
always used as activation function. Selection of cases for training, validation and
test sets is always performed randomly.

Table 5.5. MLP2:2:3 classification matrices of three classes of cork stoppers. True
classification along the rows; predicted classification along the columns.