Training Patterns - Quantity

The number of cases that are used for training is important. For backpropagation neural networks, the more training patterns that are used, the better the resulting model will normally be. If we picture a neural network as producing an optimal N-dimensional nonlinear surface fit based on the training set provided, the better we define this N-dimensional surface with our training cases, the more accurate the model will be. That includes both quantity and diversity of training cases. The result is a network that better generalizes relationships. Increasing the number of patterns and their coverage has the additional advantage of permitting more complexity in the network's hidden structure, allowing the model to handle increasingly complex relationships between input and output sets.

Qnet permits the use of test sets during training. Test sets are patterns set aside from the training set to test for network overtraining and to check the integrity of the model. If a model cannot respond intelligently to patterns outside the training set, then the model will be of little value (unless training set memorization is the goal). Qnet allows the user to indicate the number of patterns to allocate for the test set and then to monitor the network’s test set response error during training. The test patterns should be contained in the same file(s) as the training patterns. You can instruct Qnet to select a certain number of records randomly or a specified number of patterns can be used from the beginning or ending of the training file(s). If the patterns in the training data file(s) are in some type of systematic order, it is best to select the test patterns on a random basis so that test set is not a limited subset of the training cases. Normally, the training set will contain many more patterns than the test set. It is common for test sets to make of 10-20% of the total available training set.