The elephant has 11 sides

[From Bill Powers (941206.0920 MST)]

Avery Andrews (941206)--

Then again, at higher levels, it seems to me that many activities are
repetitions, with only occasional results (trying to get that program
to run: test it, change something, test it, ....), and it's amazing how
time flies when you're doing this sort of thing. I think Skinner did
identify a significant aspect of the structure of activity.

The problem is always with focusing on a single level of control and
mistaking it for the whole thing -- the modeler's version of synecdoche
(which my dictionary tells me is pronounced "sin-EK-duh-kee"). No doubt,
control of repetition rate of simple events does occur, whatever label
one puts on that level of control. But that sort of behavior is simply a
means to higher ends, and at the same time employs lower levels of
control which have nothing to do with controlling repetition rates.

The worst aspect of this problem occurs when people get stuck at one
level and essentially refuse to recognize either higher or lower levels
of organization. This is where Skinner went astray in analyzing
language. It's like watching an orchestra conductor and saying that all
he's doing in waving his arms in repetitive patterns. He is certainly
doing that, but it's not ALL he's doing -- it's not even the most
interesting aspect of what he's doing. Ditto for the programmer getting
a program to run. If you narrow your viewpoint enough, it seems that the
programmer is just typing for a while, then hitting Ctrl-F9, then
typing some more, then hitting Ctrl-F9 again, and so on until satiation
sets in. A simple repetitive pattern. But of course it's not a
repetitive pattern; if the programmer were typing the same thing each
time, it would be, but the crucial fact is that something different is
being typed each time, in a new relationship to what is already there,
and furious activity is going on at the computational level in anything
but a repetitive pattern. But all of that is invisible if you just
characterize the behavior as "typing" and "hitting a function key."

Perhaps I have downplayed my definitions of levels of control too much,
just because I know that the particular scheme I offered needs a lot of
work. People keep forgetting about the _concept_ of hierarchical
control. Behavior never takes place at just one level, and any model
that tries to handle all of behavior at one level is just wrong. In some
quarters, the mistake is in trying to represent functions at all levels
as if they were logical processes; in others, trying to represent them
as events, or relationships, or reactions to sensations, or abstract
conceptions. Each level of analysis addresses a certain level of
behavioral organization, but is wrong for any other level of
organization. Unfortunately, once the analyst gets absorbed in one
viewpoint, that viewpoint tends to be applied to all levels, from spinal
reflexes to system concepts. So you end up with things like the TOTE
unit, where a sequential control system is used as a model of all
control systems, even spinal control loops.

In the HPCT model, each level is thought of as specializing in a certain
level of operation. It treats its world ALWAYS as consisting of the
kinds of perceptions with which it deals, and NEVER in terms of any
other kind of perception. To a logic-level system, the whole world
consists of variables which are either true or false, and all processes
consist of applying logical operations to logical variables. This level
knows NOTHING about sequences, or categories, or relationships, or
events, or configurations, or sensations, or intensities. It doesn't
know what the variables it handles stand for. That kind of knowledge is
_distributed_ up and down the levels; only in awareness is it all
brought together into a single picture of reality. At any one level, the
whole world consists of variables of the type constructed by and
controlled by that level. The other levels simply don't exist.

I think it's critical for PCT modelers to practice seeing behavior at
many levels, whether they're the ones I have defined or not. When you
get stuck at one level, you start trying to account for aspects of
organization that don't belong at that level, that require some sort of
capability that doesn't exist at that level. We've run into this problem
many times in the last four years on the net. Don't all control systems
control the "relationship" between a perceptual signal and a reference
signal? Sure, if you're thinking in terms of relationships. They all
also create a sequence of error signals: first a large one, then a
smaller one, and so on to the smallest one. They all follow a certain
logic: if the error is positive, generate one output; if negative,
generate the opposite output. They all deal in intensities: even a
perceptual signal representing the truth of a logical proposition has to
have a nonzero magnitude, and magnitude is intensity, isn't it?

All that's happening here is that the observer, the analyst, is adopting
different levels of perception as a viewpoint, and applying the same
viewpoint to everything. What is needed to model with HPCT is an
awareness of ALL the levels, and an ability to see which is the
appropriate level for the organization in question. The analyst has to
become acutely aware of thinking in levels, or the whole process will
degenerate into a confusing mess. The question is always whether the
system being examined is capable of handling variables of a certain
level, or whether it is only the analyst's point of view that makes it
seem so.

I sit near the middle of a web consisting of people from a wide variety
of specialties in the sciences of life. Each specialty, it seems to me,
tends to focus on a particular level of perception and control, and to
analyze all behavior from that point of view. To some, all of perception
is a systematic or probabilistic computation; to others, it is a set of
stimuli; to others, a set of social situations; to others, a structure
of words and syntax. From my point of view, the elephant we are all
talking about has 11 sides and there are representatives from other
disciplines who speak mostly about one or two of the sides and ignore
the others. The most difficult and the most important aspect of HPCT to
get across is that the elephant can't be understood from a single point
of view.


Best to all,

Bill P.

Re: Higher levels

What are some experiments demonstrating the existence and characteristics
of specific higher levels?

Lars Chrisian Smith