Figs. 2 and 3 illustrate the variations of the MSE and the CCP
with the neuron number of the hidden layer, respectively. The
maximum CCP and minimum MSE were observed as the neuron
number of the hidden layer was 5. Therefore, our ANN model was
run with 5 neurons in the hidden layer. The ANN system could
correct the weights to minimize training error at each iteration. The
training error consequentially decreased as the system learned. Our
training result showed that the convergence of the BP-ANN model
was very fast. After ninety-seven iterations, the training error
descended to below 0.01, the stopping criterion set by the authors.
As a result, the BP-ANN model for grade classificationwas obtained
with the architecture parameters as