Page 267 - Deep Learning

250 Adaptation

of the target skill is inferred from his task behavior and, finally, the system
uses artificial intelligence techniques to compute when (at which moments
during practice) to intervene, what information to convey and how to convey
it. It is widely assumed that one-on-one instruction is superior to classroom
instruction because a human tutor adapts on a moment-to-moment basis to
the needs of the individual student, especially to the gaps and faults in his or
her representation of the target skill. Consequently, much research on tutor-
ing systems aims to invent programming techniques that allow a computer to
similarly individualize its instruction in general and its response to student
errors in particular.
The constraint-based theory implies a particular blueprint for intelligent
tutoring systems that is known as constraint-based modeling (CBM). The cor-
rect knowledge of the target domain is to be implemented as a constraint
base. A student is represented by the subset of constraints that he has violated.
Learning opportunities are identified as constraint violations. A tutoring mes-
sage should state the violated constraint, formulated in colloquial but general
language, highlight the specific features of the student’s answer or solution that
cause the constraint to be violated and explain how those features violate the
constraint. Constraint-based modeling is an elegant and practical recipe for
computer-based tutoring in response to student errors.
The proof of a recipe is in the pudding. In this case, the pudding was
cooked in New Zealand, but it is enjoyed worldwide. Computer scientist
Antonija Mitrovic at Canterbury University in New Zealand and the members
of her research team have implemented, deployed and evaluated multiple con-
straint-based tutoring systems. The performance of these systems constitute a
test of the constraint-based blueprint for tutoring systems, and, indirectly, of
the underlying learning theory.
59
The first constraint-based tutoring system was SQL-Tutor. It teaches
students to formulate queries in SQL, a database query language. The task of
the student is to fill in fields in a query schema in such a way that the resulting
query poses a meaningful and answerable question with respect to a particu-
lar database. The student fills in the fields and submits the query. The system
evaluates it and provides feedback; the student has the opportunity to revise
the query; and so on. The task of formulating SQL queries is quite difficult.
The mature version of the system contains more than 600 constraints. The
SQL-Tutor was followed by two other constraint-based systems that teach
other database skills: database design (KERMIT) and database normalization
(NORMIT). The three tasks of designing, normalizing and querying databases
60
are quite different from one another. Normalization is a multistep, sequential