Archived Reorganization and Intrinsic error

from [ Marc Abrams (990924.1926) ]

In my continuing, and fascinating, trip through the archives I came upon
this post by Bill on Intrinsic error and Reorganization

[From Bill Powers (920203.1900)]

Joel Judd (920203.0926) and Martin Taylor (920203.0926) --

Joel says:

I have usually seen "Intrinsic Error" used in relation to physiological
needs of the system. Yet learning is often discussed regarding
activities that don't seem to be "life or death."

and Martin says:

I'm still not clear what is meant by "Intrinsic Error."


Intrinsic error, which drives the process of reorganization, is just
error that the organism knows about without any previous experience ?? in
other words, you don't have to have (i.e., the body doesn't need) an
adult hierarchy of control systems (or any hierarchy at all) to recognize
that an error exists. This requirement follows from the necessity that a
reorganizing system exist and work from the start, so that the hierarchy
can be acquired as appropriate to the environment that happens to exist
around the organism.

For the same reason, intrinsic perceptual signals must represent
variables that aren't the outputs of learned perceptual functions.
Intrinsic reference signals must also be built in for the same reasons.
The reorganizing system must work without any knowledge of the world
outside the organism. That knowledge does not cause its operation;
knowledge results from its operation.

Now you know as much as I know. We can guess at some types of intrinsic
variables. There are probably intrinsic perceptions related, for example,
to hunger and thirst, although of course they aren't perceived as hunger
and thirst -- they're just something important, and the reorganizing
system acts when they deviate from the "right" state. I have proposed
that error signals are also sensed as intrinsic variables (if large
enough and chronic enough); this is legal because every control system,
no matter what it controls, contains an error signal. Comparators, being
elementary functions, can be built in, not requiring learning in order to
exist. This means that losing control (and thus producing large error
signals) or not having it in the first place, can constitute an intrinsic
error condition, regardless of what is being controlled. The latter
condition -- "regardless of what is being controlled" -- is the essential
one if reorganization is to be effective from the beginning of life.

It's possible to do research aimed at identifying intrinsic variables.
When an intrinsic variable deviates from its reference state, the
immediate effect on behavior is not a tendency toward behavior that will
correct the intrinsic error, but a change from smoothly organized
behavior to random variations. That's the sign of reorganization. As
reorganization is not systematic, the initial effect could well be to
make the intrinsic error even larger. But that will speed up
reorganization and the next change will come sooner, so temporary
organizations that increase intrinsic error will not last as long as
those that decrease it. In the long run there will be a biased random
walk to the state of zero intrinsic error, if the organism doesn't die
first. The last organization present at the time intrinsic error goes to
zero is the one we say has been "learned." It is better to think of it as
the one that prevents further reorganization. At this point, the proposed
intrinsic variable would be at its reference state.

There is no logical connection between the control system that results
from reorganization and the intrinsic error that causes reorganization.
This is what makes reorganization so powerful. If it's necessary to learn
how to align a pointer an inch to the left of a moving target in order to
reduce the state of hunger to zero, that control system with that
reference level will be acquired, if reorganization beats out starvation.

Skinner found that "shaping" gives reorganization a much better chance of
working than simply withholding reward (maintaining intrinsic error)
until the requisite pattern of behavior appears. If the change in
organization required to restore control is small, the chances are much
better that a random process will discover it than if the required change
is drastic. So the experimenter allows a new behavior to reduce intrinsic
error (produce reinforcement) when the new behavior represents a partial
step toward the ultimate form that the experimenter wants to see. This
greatly increases the chance that a blind variation will succeed at each
step, so that very great changes can be brought about through a series of
small steps.

When intrinsic error is zero, the rate of reorganization does not
necessarily drop to zero. I have postulated, furthermore, that focusing
attention on some part of the hierarchy directs to that part whatever
amount of reorganization is going on. If the zero?error rate of
reorganization is not zero, then wherever you focus your attention, some
degree of reorganization will be occurring. If the systems involved are
already organized for a minimum of intrinsic error, reorganization will
just cause variations that keep returning to the same state. If the
environment has changed its properties slightly (inducing error and thus
attracting attention), the reorganizing will converge to a new form not
far from the old one and then vary around that form if the zero?error
rate of reinforcement is not zero.

A lot of what we call "learning" is not reorganization, but memorization.
When I show you the order in which to press the buttons to open the door,
all you have to do is remember what you saw and create that sequence?
perception again. This doesn't require any new capabilities or any
changes in existing control systems. Learning vocabulary is probably
learning of this non-reorganizing kind. A person who is a "natural mimic"
probably doesn't even have to reorganize to produce the sounds of a new
language -- there aren't any sounds this person doesn't already know how
to produce. All that has to be "learned" is which sounds to make, and
that requires only memory.

There are probably learning algorithms too, which, once acquired through
reorganization, simply have to be applied to work out the correct action
in a new circumstance, like a computer program. The algorithm, of course,
is a systematic process, and so is not reorganization.


... it has also been suggested and promoted by Gary that Blind Variation
and Selective Retention (BVSR) process outlined by Campbell seems to be
a promising description for the reorganization process

Yes, I've suggest it too. The "selective retention" part doesn't require
a separate "retention" mechanism; an organization is "retained" when it
is effective in correcting intrinsic error and stopping the process of


Date: Wed Feb 05, 1992 4:42 am PST
Subject: Reorganization