Information Technology Reference
In-Depth Information
PP is either estimated, measured directly, or ignored. In most cases it is ignored
because the task is unique enough to the learner. In other cases, such as taking up a
familiar task, the equation does not fit as well.
Constant2, the minimum time due to the limitations of the environment, is
computed from equations describing the world or using physical equipment. For
example, you might measure how long it takes a ball to fall to the ground with a
camera if the task involves catching a falling ball, or you might record how fast an
interface can accept keystrokes when driven by a computer program.
a is thus found typically by plotting the data on a log-log plot, and fitting a
straight line. This fitting is typically done by fitting a linear equation (which has a
known closed form solution) to the log of the trials and the log of the time. The
more accurate but more computationally expensive approach is to fit the power
law equation including the constants that are appropriate using an iterative
algorithm.
Having an equation to predict learning is important for several reasons. For
science, it helps summarize learning, and comparison of the constants is a useful
way to characterize tasks. It also has practical applications. For engineering,
design, and manufacturing, it predicts how fast users will become with practice.
These
equations
are
used
in
manufacturing
to
predict
factory
output
and
profitability.
Others believe that the match to a power law is an artifact of averaging the data
from multiple people and multiple series, and that the curve is best described as an
exponential when the data is examined in its purest form (Heathcote et al. 2000 ).
Both cases have basically the same implications for users with limited lifetimes
and physical equipment, but have different implications for the details of how the
underlying cognitive architecture is implemented.
The improvement in performance time itself does not appear to delineate the
stages of learning noted earlier. This may be because the first stage of learning,
where performance might be quite difficult, has to reach a nearly complete level
before the task can even be performed. There are hints of this in Fig. 5.9 b, where
the initial slope is fairly shallow in the log-log plot, perhaps more shallow than
would be expected. In complex tasks, the transition of rule learning and rule tuning
within task might lead to steady improvement, and be masked by the large number
of rules and sub-tasks. Looking at individual users working on well-measured
tasks may allow these stages to be seen. When this has been done, strategy changes
can be seen (Delaney et al. 1998 ).
In addition to time reducing with practice, several other aspects of performance
improve as well (Rabbitt and Banerji 1989 ). Errors decrease with practice. The
variance in the time to do a task also decreases. That is, with practice the range of
expected times decreases. Some think that this decrease in variability is what leads
most to the speedup because the minimum time to perform a task generally does
not decrease with practice (although there are clearly notable exceptions to this, for
example, for tasks that cannot be completed by novices).
There appear to be two places where learning does not get faster. Users cannot
get faster when the machinery they are working with cannot keep up with them.
Search WWH ::




Custom Search