First, we show the inductive learning potential of
PDL2, and the benefits gained by the addition of
precepts on several datasets from [15]. PDL2's
measure of similarity (i.e., node activation) is designed
to deal primarily with symbolic-valued (i.e., nominal)
attributes. Consequently, the selected datasets have
nominal attributes. All don't-know values are treated
as don't-care values, except in the voting3-84 dataset
in which they are treated as a distinct input value. The
results are shown in Table 1. Each of the reported
results is an average over 10 runs of PDL2. In each
run, the training set and the test set are randomly
regenerated. Each set contains one half of the
examples in the complete data set.