[From Bill Powers (930409.1145 MDT)]
Greg Williams (930409) --
Briefly, he [Maier] trained rats to do a task, then made the
outcome random (NOT contingent on doing what would have
otherwise been the efficacious action). The rats ended up
behaving rather peculiarly, too say the least (perhaps some
netters might want to guess how they behaved before learning
all of the details).
OK, I'll give it a stab since we do have a model of this kind of
situation.
The random output effects suggest the e. coli situation, where
there is no systematic relationship between any one action and
any controlled variable. Under these conditions, if there is any
control at all it can be achieved only by altering the RATE of
behavior, not by altering the acts in any systematic way. In a
rat, which is far more complex than e. coli, we would also expect
reorganization of behavior to be going on if the effects of any
behavior involve intrinsic variables. The behavior should begin
to lose its normal connections with its effects.
If only the rate of behavior can have any systematic effects (but
see below), then we have to consider the sorts of variables
affected only by rate of behavior and not by the specific
behavior itself. In a choice situation, for example, the only
bias on behavior (arrived at in the manner of e. coli) would be
toward the statistical distribution of rewards among the choices.
However, in the relationship between the distribution of random
actions and random reinforcements, there may be local maxima of
control of reinforcement, so behavior would do a random walk
around some form, in the way that e. coli moves randomly around
the source of a chemical gradient. This could trap the animal in
some irrelevant pattern of rates of behavior or even kinds of
behavior.
It would be hard, however, for an experimenter to make sure that
the only effects of behavior that matter to the animal are those
having to do with the specific contingencies that the
experimenter is presenting at random. The very act of behaving
creates fatigue, for example. The appearance of the surroundings
changes with every change in the rat's actions. Other goals such
as exploration or light/dark preferences are always active. So
the most we can say is that a few continencies out of all those
that actually exist have been randomized; the remainder are as
systematic as ever. Among the rats' actions, even in the midst of
reorganization, there will be systematic components aimed at
controlling perceptual variables irrelevant to the experimenter's
purposes. The components of behavior used to control these other
variables might appear quite bizzare if one is thinking that they
are attempts to control for the randomized consequences. Given
what they actually accomplish for the rat, however, they might
make perfect sense.
The main effect that I would predict would be periods of rapid
behavior interspersed with periods of little behavior. This is on
the basis of guessing that rapid behavior indicates an increase
in intrinsic error, while little or no behavior indicates that
intrinsic error is decreasing or at least not getting worse --
the e. coli phenomenon.
I checked through MARMOT, and Ft. Lewis does have the Maier book.
Mary or I will pick it up later today.
Oh, I just realized that there is one more effect that can be
predicted. You say that Maier first trained the rats, by
implication using non-random contingencies. So the rats had
learned how to control for contigent rewards by means of certain
actions.
When randomizing commences, the immediate result will be an
increase in error in each control system involved with each
contingency. The result should be an immediate intensification of
the actions: instead of pressing a bar less when the reward fails
to appear as usual, the animal will press the bar more. It might
not even notice that the other contingency has occurred, as it
focusses on producing enough output to make that damned reward
appear where it has always appeared before.
The bizzarre behavior will develop only after some time has
passed, so that intrinsic error has had a chance to build up and
start reorganization going.
But it's hard to predict behavior when reorganization starts, so
almost anything could happen.
···
--------------------------------------------------------
Greg Williams (930408) via Gary Cziko:
To all those who appear SINGLE-minded in their view of science
(i.e., population description isn't "real"), I say: just try to
use generative modeling of individuals to schedule bread flour
deliveries to Subway shops across the country.
You're beating a dead unicorn. I have always agreed that to study
mass phenomena, mass statistics are appropriate. My objections to
mass statistics in the practical sense is that in most published
studies the statistics are lousy even in statisticians' terms. It
is almost unheard of for a statistical study actually to be
replicated. Some such studies -- like longitudinal twin studies
-- are practically impossible to replicate anyway.
Good studies of population phenomena are done, and are applied
appropriately to predict population effects. The basis of the
prediction is only that what has happened before is likely to
happen again, so I would call it engineering rather than science,
but there's nothing illegitimate about such studies. One just has
to remain aware of who the beneficiary is.
Even in the application of the results of population studies to
individuals, the degree of scientific error involved depends on
the quality of the results. If one finds, over a population, that
the effect correlates 0.999 with the treatment, there will be
very few individuals in the population who will not have
exhibited the effect. It's not a bad guess, then, to say that one
may have discovered a property common to all the individuals.
It's then worthwhile to investigate the exceptions and see if the
deviations can be explained, making the observation universal. In
fact, such a population study with N individuals would be
essentially equivalent to N individual studies.
But if the correlations are down about 0.8 or less, there are
simply too many exceptions to the hypothesis to permit saying
that anything common to the individuals has been found. To go
ahead anyway and pretend that you have discovered something about
ALL the people is simply to commit a scientific fraud. Such
frauds abound in the literature.
Please try to keep straight what I object to and what I do not
object to. I'm quite consistent about it.
--------------------------------------------------------------\
Best,
Bill P.