Page 269 - Deep Learning

252 Adaptation

ER-Tutor
SQL-Tutor
8000 Normit

6000
Cumulative Number of Users 4000

2000

0
0 10 20 30 40 50 60 70 80
Months on Web
Figure 7.4. Growth of the user pool for constraint-based tutoring systems.

of trial and error behavior. Once practice is under way, students learn primar-
ily from the information residing in the outcomes, positive or negative, of their
own actions. In the long run, information laid down in memory during the
execution of a learned strategy supports the discovery of shortcuts and mul-
tiple forms of strategy optimization.
Each mode of learning is a potential target for tutoring. The research
and development process used in the case of learning from error potentially
applies to every mode of learning: select a psychologically plausible learning
mechanism; specify the type of information that mechanism needs to receive
as input; identify the conditions that mark a learning opportunity; specify how
to compute what information to convey and how to convey it; implement the
tutoring system; and evaluate. The full pedagogical power of intelligent tutor-
ing systems will be realized when those systems become capable of supporting
all nine modes of learning described in Chapter 6. Eventually the produc-
tion of effective tutoring systems will become routine. Educators will develop
systems for every conceivable topic and make them available electronically,
thereby providing individualized instruction for every person on the planet
at virtually no cost, over and above the initial development costs. The race
between the need for citizen training and society’s ability to pay for it is on,
and we can still win.