In the ANN analysis carried out for the four models (Table 7),
the smallest maximum relative error values and RMSE’s in the testing
sets are obtained from the network for Model 1 with a = 1 and
g = 1.0, for Model 2 with a = 0.25 and g = 0.50, for Model 3 with
a = 0.75 and g = 1.0, and for Model 4 with a = 0.75 and g = 0.75.
The smallest average relative errors of the testing data sets in these
cases are found from the networks as 1.955%, 1.222%, 2.721%, and
1.577%, respectively. The maximum relative error may be reduced
if the stopping criterion, which is the epoch number, is increased.
Besides, the conjugate gradient or the scaled conjugate gradient
methods may be used to reduce the maximum relative error instead
of the generalized delta rule in learning. Also different network
structures with one or more hidden layers or nodes with
different learning rates and momentum terms may produce smaller
error. Relative error is calculated as