Page 268 - Deep Learning

Error Correction: The Specialization Theory 251

choice task, similar to the tasks traditionally investigated by tutoring research-
ers in domains such as algebra, physics and programming. The database
design tutor KERMIT presents the student with a complex graphical interface
that requires multiple actions but does not impose any particular sequence on
the steps. The task of formulating database queries is different from either of
these. The second generation of constraint-based systems to come out of the
Intelligent Tutoring Systems Group at Canterbury broke out of the computer
science domain to encompass such varied instructional topics as the rules of
capitalization in English and group interaction skills. The variety of topics
taught by these systems demonstrates the versatility of the constraint-based
modeling philosophy.
There is no doubt that students learn effectively while interacting with
these systems. Multiple empirical studies carried out in Mitrovic’s laboratory
have shown that the probability of violating a constraint decreases as a func-
61
tion of the number of past opportunities to violate that constraint. In particu-
lar, studies that compared tutoring messages based on intuition with messages
that were in strict accord with these theory-based prescriptions have found a
small advantage for the latter.
The most impressive demonstration of the usefulness of these systems
is that several of them have entered the commercial educational market, an
unusual success in the field of intelligent tutoring research. Web access to the
Canterbury suite of database tutors – SQL-Tutor, KERMIT and NORMIT –
was packaged with a commercially successful series of textbooks in computer
science. Figure 7.4 shows the growth of the online user population. Thousands
of students worldwide have benefited from the advantages of constraint-based
tutoring systems.

From CBM to Multiple Tutoring Modes

The success of constraint-based tutoring systems demonstrates the useful-
ness of the design principles derived from the theory of learning from error.
However, that theory only describes one learning mechanism. But as argued
in Chapter 6, there is no reason to believe that every learning event that occurs
during skill acquisition can be attributed to one and the same learning mecha-
nism. Instead, there are good reasons to believe that people learn in multiple
ways during skill practice. Students can benefit from verbal instructions, espe-
cially at the outset of practice; they can reason from their declarative knowl-
edge about the task; they can draw analogies to already mastered skills; they
can make good use of solved examples; and they can learn from the outcomes