Fig. 1. Use the AutoTest function to select out-sample for automatic testing of networks.
By default, all the data points in the training are used for training, and all in the testing range for testing. However, you can use the AutoTest function to further select the training and testing data from their ranges.
The selection pattern (Test 1/4, Learn 3/4) means take the first 3 data points for training and the subsequent fourth data point for testing.
You can select the training and testing data such that they do not overlapped. The non-overlapping combinations are: (Test 1/4, Learn 3/4), (Test 1/3, Learn 2/3) and (Test 1/2, Learn 1/2). With these selections, those data used for training will not be used for testing. The significance of these selection strategies is that they exclude the test data from the training data, so the network performance is measured against the test data but not the training data. In the case of those GENETICA-based networks, the entire network population will be evaluated and ranked according to the test data rather than the training data, and the possibility of an over-trained population is reduced.
Copyright © 1993-96, NIBS Pte Ltd.