Generalization:

When a test set has been set aside, one may check whether the behavior learned by the DTRNN from the learning set is consistent with the test set. This is called the generalization test.

Some learning algorithms (Giles et al., 1992) partition learning sets in a small starting set and a number of test sets (see (Giles et al., 1992)). Once the starting set is learned, a test set is used to check the DTRNN. If the test fails, the test is added to the learning set to be relearned. If, after some relearning runs, the remaining test sets are correctly classified without having to relearn, the learning algorithm terminates.