Page 208 - Deep Learning

The Growth of Competence 191

The hiatus in theorizing came to an end with a 1979 Psychological Review
42
paper by Simon and Yuichiro Anzai. They presented a computer program
that modeled the successive strategy changes of a single person who solved
a problem-solving task multiple times. The paper demonstrated the feasi-
bility of simulating the acquisition and not only the execution of cognitive
skills. The paper was soon followed by the initial versions of J. R. Anderson’s
ACt model and the Soar model proposed by newell, paul S. Rosenbloom
and John E. Laird. 43, 44 The 1983 version of the ACt model included six dif-
ferent learning mechanisms (proceduralization, rule composition, rule gen-
eralization, rule discrimination, strengthening and weakening). The success
of the initial models established computer simulation as a useful theoretical
tool.
The story since 1979 is one of proliferation. A wide variety of theories
have been proposed and embodied in computer simulation models. typically,
a model consists of a cognitive architecture that follows some theory of sta-
ble behavior like the one presented earlier in this chapter, plus a repertoire
of learning mechanisms. Models differ in the details of their performance
mechanisms, but, more important, they incorporate different learning mech-
anisms. The fact that theories are embedded in computer programs have the
peculiar linguistic consequence that they tend to be known by the proper
names of those programs instead of descriptive titles. Almost all serious mod-
els are multimechanism models. For example, Ron Sun’s CLARION model
learns through both bottom-up generation of rules and rule generalization,
while VanLehn’s Cascade model learns from solved examples as well as by
analogy. 45, 46 All in all, the emergence of computer simulation as a theoret-
ical tool triggered an unprecedented explosion of the theoretical imagina-
tion. More new hypotheses about the mechanisms behind the acquisition of
cognitive skills were proposed in the period 1979–1999 than in the previous
century. 47
The journey from the Law of Effect to Cascade and CLARION represents
a century of scientific progress. Theories of skill acquisition are more precisely
formulated, more responsive to the complexity of human skill acquisition and
more explanatory than they were a century ago. nevertheless, the historical
trajectory through the space of theories looks in retrospect like a drunkard’s
walk. The choice of mechanisms to include in any one model is not grounded
in any principled argument about which repertoire of mechanisms is most
likely to be the one that has in fact evolved in the human brain. For example,
why did the ACt model of 1983 vintage not include any mechanism for learn-
ing from examples, and why does Cascade lack a discrimination mechanism?