Fig. 3.
An example of the evolution of the training-construction and training-validation errors with training iterations. This example corresponds to i=1 and j=2 in Algorithm A for constructing NNall. The training-construction error continuously decreases with increasing number of training iterations. The training-validation error, however, initially decreases and then increases beyond approximately 500 iterations. To avoid overfitting, network training was stopped at the point of the minimum training-validation error.