Upon reading in the rule base, the knowledge primitive measurements of the
training examples, and all training example goals, OMLET begins by
learning the membership functions of all level 1 categories. The first
learning epoch is used to make initial estimates of the membership functions, and
then OMLET iterates for 1000 additional training epochs. A learning rate of
0.15 is used during the 1000 training epochs, so that 15 percent of the actual
error for each training example is propagated to the adjustable ranges on each
epoch. After the 1000 training epochs, the best range parameters (those that
resulted in the lowest overall error) for level 1 categories are restored and
frozen. The 1000 training epochs are then repeated for the level 2 categories,
followed by the level 3 categories, and so on until all ranges in the category
definition tree have been learned
.
The performance task of the OMLET system is evaluated by how well the trained system recognizes objects that were not used in the training phase. One measurement of system performance is the error observed on the test examples. The error for a test example is computed as the absolute value of the difference between the desired and actual evaluation measures. Training/Test sets are configured two ways: random partitioning of all labeled data into training and test sets, and leave-one-out testing. In the first case, for a given size training set, 10 train/test set pairs are created by randomly partitioning all the labeled data. The error for a single test set is the average error of all test examples. The results for a given size training set are reported as the average error of the 10 partitions. In leave-one-out testing, one example in the data set is used to test while all remaining samples form the training set. This is repeated using each example in the data as the test set, and results are reported as the average error of all test examples. The average error per example versus the training set size is plotted for training sets of 10, 20, 30, ... , N-1 samples. The point with N-1 training examples represents the leave-one-out test results.