[From Rick Marken (920318)]
Well, I guess things have been as quiet as they seemed. I'll
let Bill respond to Martin's interesting claims about GAs -- if
he wants. I'll try to respond to the Beer posts. I'm glad to
see that Randy is hangin' in there. I'm sure that he will find
his attempts to get us to understand his point of view just as
frustrating as we find our attempts to get him (and just about
everyone else) to understand ours. I doubt that we will convert Randy --
but its nice to have someone participating on CSGNet who has such
a pure understanding of the conventional point of view of behavior.
To begin:
Randy (to Powers):
I continue to encourage your interest in models of insect locomotion.
However, it seems that you prefer to make everything purely
sensory-driven.
No, we prefer to make purposive behaviors have outputs that
drive sensory inputs to internally specified reference levels.
We don't think that those sensory inputs are really smart enough to
drive anything anywhere but in random directions.
This is, of course, a perfectly legitimate approach,
but I must repeat that it does not appear to be the way biology does
it.
Well, if "it" is purposive behavior then I think it is biologists, not
biology, that is mistaken.
A number of experiments have demonstrated that the neural circuits
underlying many rhythmic behaviors (e.g. walking, swimming, chewing,
breathing) can generate the basic oscillatory pattern IN THE COMPLETE
ABSENCE OF SENSORY FEEDBACK.
Sure it can. But can it CONTROL the pattern (if that is what is
controlled). If the "basic oscillatory pattern" is not a controlled
variable then there is no reason to sense it. Did anyone test to
see if the pattern is controlled? See my comments on "the test for
the controlled variable" below.
Of course, this central rhythm must be
reinforced and fine-tuned by sensory feedback in order to exhibit
completely normal output patterns.
I think we're talking "lower level control systems" here. And the
"normal output patterns" suggest that these patterns are under
control. The words "reinforcing" and "fine tuning" imply roles for
these sensory inputs that they could not possibly play and result
in control. If sensory inputs are in the control loop then they
are controlled -- they don't "strengthen" or "guide" output.
In addition, work on cockroach
locomotion by Sasha Zill has suggested that even when sensory feedback
is intact, it may come in too slowly to play any role in fast walking
insects (the cockroach is capable of stepping frequencies in excess of
24 Hz!).
This "speed of feedback" stuff is one of the BASIC MISCONCEPTIONS
that led psychology (and biology and all the other life sciences)
away from an understanding of the nature of feedback control in
behavior. It is based on an S-R conception of how a feedback control
system works. The idea is that a stimulus causes a response that
has some sensory consequence (feedback) -- but if that sensory
consequence doesn't register fast enough then it can't be of any
use. But feedback loops don't work that way -- there is a continuous loop
and feedback is ALWAYS their -- feedback is what is controlled. There
are dynamic contraints on the operations of the loop --slowing factors
and transport lags. But there are ways of dealing with transport lags
(which is what Sacha thinks are too long) when they are long relative
to the bandwidth disturbances to the controlled variable and it is also
highly unlikely that the transport lag in the leg position control system
of a roach is anything close to 60msec, it's probably about 1/10 of that --
plenty fast for control (if position is controlled).
In response to my query:
Would the acceleration of the bug as it falls off a ledge count as a>
behavior to be modeled? If not, why not. If so, why so?
Randy says:
It certainly could. Some species of moths are preyed upon by bats
who, as we all know, navigate by echolocation. These moths have
evolved an interesting escape mechanism. Whenever they detect
vibrations of a certain frequency (namely, that used by bats searching
for prey), they simply fold up their wings and drop like a stone.
.This certainly counts as a behavior in the ethological sense, though I
couldn't say whether it is a behavior in the specialized technical
sense of PCT.
You would have to test to see if the acceleration is controlled. I believe
that the acceleration is simply an output that is part of the loop that
controls another variable -- sensed intensity of sound at a particular
frequency. This answers your question about what might be the controlled
variable. The point I was making is that the conventional approach to
behavior makes no distinction between controlled and uncontrolled
consequences of neural output. This confuses the modeling process; it
makes it impossible to tell when you need a control model vs a
response generation model. I'm afraid that this confusion,
combined with a strong bias toward output-generation models,has made it
impossible for conventional life scientists to understand that purposeful
behavior is the control of perceptual variables. The problem is that control
can look like generated output on casual inspection; the study of "behaivor"
must start with the test for controlled variables. Otherwise, you don't
really even know what you are modeling. The "test" is described in Powers'
Behavior: The control of perception. It is very important to understand
the test. Without it, you have no way of knowing what an organism is
doing (see my article "Behavior is the first degree" in the book "Volitional
Action" edited by Wayne Hershberger).
The basic idea of CT seems to be that behavior is the consequence of
negative feedback control of selected sensory inputs.
Well, er, sort of. Controlled variables are behaviors that are a consequence
of negative feedback control. It is also important to note that many
variables that we call behavior (such as the spatial position of e. coli)
are NOT controlled -- they are side effects of the control of other
variables (for e.coli spatial position is not controlled -- just the
perceived gradient of certain chemicals -- a unidimensional variable, not
a three dimensional spatial variable). A control theorist would not call
these behaviors "behavior". We would call them "irrelevant side effects"(IREs).
I think many models of behavior are, in fact, models of IREs.
However, if CT is making the much stronger claim that negative
feedback control is universal and ALL behavior can be understood in
its terms, then I am extremely skeptical.
No - the problem is the word "behavior" again. PCT says that ALL PURPOSEFUL
BEHAVIOR (controlled variables) is the result of closed loop, negative
feedback control of perceptual inputs. Other variables that might qualify
as behavior but that are not controlled (not purposeful) are IREs --generated
by the good old cause-effect processes that watchmakers have been familiar
with for centuries. Cause-effect models don't work with controlled variables;
you need to understand circular causality to understand control.
Many consequences that
are of utmost importance to an animal are not controlled in any
negative feedback way.
Well, there are ways of testing this -- the test for the controlled
variable. I think the "cockroach escape response" example of "open loop"
control is actually a good example of a system busy controlling
something very important to it -- its perception of forces, smells, etc
that might be from predators (though all the cockroach knows is the
percpetual veriables -- not "predator" -- that's our perception).
Finally, PCT has no organizational preconceptions. We are not interested
in modeling a phenomenon with a control model until we know that what the
system is doing is controlling. I would suggest that your approach actually
does begin with an observational preconception -- it is assumed that
"behavior" is any visible (to you, the observer) consequence of an organisms
actions (where "actions" are any output of the organism that can be seens
as a cause of the consequence called "behavior" -- ie. it could be a
neural impulse, muscle tension or limb movement, etc). PCT starts by
recognizing that their are two different kinds of visible consequences
of actions -- those that are controlled (produced on purpose) and those
that are NOT (the IREs). We are aware of both kinds but only apply control
models (closed loop, negative feedback systems) to the former.
Regards
Rick
ยทยทยท
**************************************************************
Richard S. Marken USMail: 10459 Holman Ave
The Aerospace Corporation Los Angeles, CA 90024
Internet:marken@aerospace.aero.org
(310) 336-6214 (day)
(310) 474-0313 (evening)