Page 300 - Deep Learning

Error Correction in Context 283

Reported Near Miss Rates
(US 1987–1998 Canada 1989–1998 UK 1990–1999)
6
US NMAC
Canada Airprox
UK Airprox
5 NMAC learning curve model
Data Sources: FAA, CAA and TSB
Rate per 100,000h (IR) 3
4

1 2

0
0 50 100 150 200 250 300 350
Accumulated Experience (MFhrs)
Figure 8.7. The error reduction curve in the airline industry, 1987–1999. The errors
are events in which one airplane comes too close to another. Reprinted with permis-
sion from Duffy and Saull, 2003, Figure 2.6.

change how each one of them operates. For example, the engineers and scien-
tists working at NASA and the employees of a corporation are better thought
of as organizations than as populations. An organization has a central coordi-
nating agency that imposes a division of labor on its members and assembles
their partial solutions into the output of the organization. In organizations,
the successful completion of one person’s task is typically a prerequisite for the
successful completion of another person’s task. If so, additive models are insuf-
ficient because they do not capture those interactions. So what happens to the
shape of change when we move from a population to an organization?
Studies of learning curves in business organizations by economists tend
to confirm that learning curves for economic organizations follow the same
negatively accelerated curve that characterizes individuals and populations.
27
To the best of my knowledge, there are no empirical learning curves published
for manufacturing plants or other types of organizations that depart signifi-
cantly from the negatively accelerated type of curve. It appears that even when
operators are hooked into each other via prerequisite relations, the learning of
the collective follows this type of curve. The reason is that even in this case,