Debugging, Problem solving, and Einstellung
Over 50 years ago Luchins demonstrated the powerful effect of "mental
set." He found that subjects given N problems that could all be solved
in the same manner would often fail to notice an obvious solution to the
N+1st problem. Instead, they would attempt to apply the same procedures
that worked for the first N problems. This failure to see an obvious solution
was called "Einstellung" which means psychic blindness.
Martin Levine in 1975 expanded on Luchins's work and proposed that hypotheses
can be meaningfully grouped into domains of related hypotheses. Levine proposed
that (a) subjects will test hypotheses from the domain that contained the
solution to previous problems and (b) once subjects start testing hypotheses
from a. domain, they will not switch domains until they have tested all
the hypotheses in the domain. This implies that if subjects are testing
hypotheses from an infinitely large domain that does not contain the solution,
they will never find the solution. Indeed, Fingerman and Levine (1991) showed
that subjects induced to test hypotheses from the domain of position sequences
(e.g., left, left, right, left, left, right, etc.) will overlook a simple
solution to a problem (i.e., alway choose the stimulus with the "A"
on it) and fail to solve the problem even after hundreds of trials.
I believe that real-world problem solving is often very inefficient because
people have a propensity to miss obvious solutions to problems and spend
a great deal of time testing complex and hard-to-test hypotheses. One finding
from my early work in this area (Lane, Bleichfeld, McDaniel, and Rabinowitz,
1976) is that people are more likely to switch from the domain of simple
hypotheses to the domain of complex hypotheses than vice versa. This suggests
that once a person starts testing hypotheses more complex than the solution,
they will have a difficult time solving the problem.
More recently, I have become interested in applying notions of hypotheses
testing theory to trouble shooting/debugging. Debugging is the most time
consuming and costly portion of software development. It is also a very
time-consuming portion of many users' interaction with their computers.
A basic issue in hypothesis testing is the order in which one chooses to
test hypotheses. A very costly error is to test hypotheses that are complicated
and/or take a long time to test when the solution is very simple. There
is ample anecdotal eveidence that this happens frequently. Therefore, the
most efficient strategy is often to test the easiest-to-test hypotheses
first. In a series of studies my students and I have shown that subjects
are typically not attuned to the strategey of testing hypotheses based on
how hard it is to test them. Although in the "real world" there
is often a trade off between the likelihood that a hypothesis is correct
and the ease with which it is tested, we studied simpler situations where
the probability of being correct was held constant.
Brad Ashby, in his master's thesis, presented subjects with a list of algebraic
expressions and asked them to find the one expression that was incorrect.
A typical expression might be: 2a+3b=24. To test whether an expression is
correct, subjects would click on the expression so as to be presented with
a screen containing the values of "a" and "b". The algebraic
expressions varied in length (number of terms). Expressions checked by subjects
were clearly marked so that subjects would not need to remember which expressions
they had already checked. Since all expressions had an equal probablity
of being the one that was incorrect, the most efficient strategy is to check
the shorter expressions first. Few subjects used this strategy at first,
although a majority of subjects learned it over a series of trials. Interestingly,
when the task was modified slightly so that two of the expressions were
in error, subjects failed to learn to use the "test the easiest to
test items first" strategy.
This research demonstrates that even in a very simple laboratory situation,
subjects are not adept at choosing the order to test hypotheses. It seems
likely that people would fair even more poorly in the more complex hypothesis
testing situations that arise in the real world.
More recently Tim Dammon and I (Dammon, C. T., & Lane, D. M.(1993) Transfer
of training on a fault diagnosis taks. Proceedings of the Human Factors
Society's 37th annual meeting, 1272-1276.) investigated the learning
and transfer of the strategy of testing easiest-to-test hypotheses first.
In the first experiment it was found that training on the "one bug"
problems in the task used by Brad Ashby transferred readily to "two-bug"
problems. The second experiment investigated whether training on this task
would transfer to an entirely different hypothesis-testing situation. Subjects
were first trained on the "one-bug" problem and then given a checkbook
balancing task where the error could be something easy to test (such as
an incorrect entry) or something harder to test (such as computing interest
incorrectly. Despite the difference in the surface features, subjects were
able to transfer the underlying strategy of testing easiest-to-test hypotheses
first.
This result suggests that a very general problem solving strategy can be
learned and then applied spontaneously in a variety of situations. This
contrasts with other research on transfer has found that it is difficult
for subjects to transfer learning to situations that do not share surface
features with the situation in which the original learning took place.
A pilot study conducted by Tim Dammon suggests that experiencing Einstellung
has a longer-lasting effect then simply learning about it. Two problems
that have easily-missed simple solutions were used in this study. Half of
the subjects were given one of these problems while the other half were
told about the danger of overlooking simple solutions. Half of each of these
two groups was tested the same day on the other problem; the other subjects
were tested a week later. Subjects tested the same day readily found the
simple solution. The key finding was that subjects tested a week later were
much more likely to find the simple solution if they had been given the
first problem than if they had simple been told about the danger of missing
simple solutions.