Page 272 - DCAP310_INTRODUCTION_TO_ARTIFICIAL_INTELLIGENCE_AND_EXPERT_SYSTEMS

Introduction to Artificial Intelligence & Expert Systems

Notes
Example: In 1987, Brian Ross found that giving learners analogical examples to illustrate
a probability principle facilitated their later use of the probability formula to solve other
problems. In classroom studies from 1998, Daniel Schwartz and John Bransford found that
generating distinctions between contrasting cases improved students’ subsequent learning. As
reported in 1993, John Clement used a technique of bridging analogies to induce revision of
faulty mental models. Learners were given a series of analogs, beginning with a very close
match and moving gradually to a situation that exemplified the desired new model.
Another line of inquiry focuses on the spontaneous analogies people use as mental models of
the world. This research generally begins with a questionnaire or interview to elicit the person’s
own analogical models. For example, Willet Kempton in 1986 used interviews to uncover two
common analogical models of home heating systems. In the (incorrect) valve model, the
thermostat is like a faucet: It controls the rate at which the furnace produces heat. In the (correct)
threshold model, the thermostat is like an oven: It simply controls the goal temperature, and the
furnace runs at a constant rate. Kempton then examined household thermostat records and
found patterns of thermostat settings corresponding to the two analogies. Some families constantly
adjusted their thermostats from high to low temperatures, an expensive strategy that follows
from the valve model. Others simply set their thermostat twice a day – low at night, higher by
day, consistent with the threshold model.
Retrieval of Analogs: The Inert Knowledge Problem

Learning from cases is often easier than learning principles directly. Despite its usefulness,
however, training with examples and cases often fails to lead to transfer, because people fail to
retrieve potentially useful analogs. For example, Mary Gick and Holyoak found in 1980 that
participants given an insight problem typically failed to solve it, even when they had just read
a story with an analogous solution. Yet, when they were told to use the prior example, they were
able to do so. This shows that the prior knowledge was not lost from memory; this failure to
access prior structurally similar cases is, rather, an instance of “inert knowledge” – knowledge
that is not accessed when needed.
One explanation for this failure of transfer is that people often encode cases in a situation-
specific manner, so that later remindings occur only for highly similar cases. For example, in
1984, Ross gave people mathematical problems to study and later gave them new problems.
Most of their later remindings were to examples that were similar only on the surface, irrespective
of whether the principles matched. Experts in a domain are more likely than novices to retrieve
structurally similar examples, but even experts retrieve some examples that are similar only on
the surface. However, as demonstrated by Laura Novick in 1988, experts reject spurious reminding,
more quickly than do novices. Thus, especially for novices, there is an unfortunate dissociation:
While accuracy of transfer depends critically on the degree of structural match, memory retrieval
depends largely on surface similarity between objects and contexts.

Analogical Encoding in Learning

In the late 20th century, researchers began exploring a new technique, called analogical encoding,
that can help overcome the inert knowledge problem. Instead of studying cases separately,
learners are asked to compare analogous cases and describe their similarities. This fosters the
formation of a common schema, which in turn facilitates transfer to a further problem.

Example: In 1999, Jeffrey Loewenstein, Leigh Thompson, and Gentner found that
graduate management students who compared two analogical cases were nearly three times

266 LOVELY PROFESSIONAL UNIVERSITY