NetNews Usenet Archive 1993 #1/NN_1993_1.iso/spool/bit/listserv/csgl/2102

Text File | 1993-01-10 | 7.9 KB | 164 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!zaphod.mps.ohio-state.edu!howland.reston.ans.net!paladin.american.edu!auvm!MCIMAIL.COM!0004972767
Return-Path: <@VMD.CSO.UIUC.EDU:[e-mail]>
Message-ID: <52930110035825/[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Sun, 10 Jan 1993 03:58:00 GMT
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         Hortideas Publishing <[e-mail]>
Subject:      Who's got the generative model?
Lines: 152

From Greg Williams (930109)

>Bill Powers (930108.0800)

I'm almost persuaded by your argument for Newton as a generative modeler. He
certainly wasn't curve-fitting. Still, he didn't try to hypothesize WHY his
particular generative model for gravity should work. I think it was a good
idea at the time NOT to make such hypotheses. So I'm trying to make room in my
conceptual apparatus for the case where a scientist doesn't just fit a
polynomial (often of degree 1!) to the data, but, as you say, hypothesizes a
particular form for a mathematical model of the relationships among
observables, and yet doesn't try to include any nonobservables in the model.
Bear with me.

>[Here there was a 17-hour power failure while we finished getting
>2 feet of snow]

Good thing the CSG meeting is in the summer! I hope everyone stayed warm and
comfortable for the duration of the AC outage.

>But he [Skinner] was fixated on environmental control of
>behavior, and was forced to conclude that behavior is controlled
>by its consequences, even though the only CLEAR relationship he
>could see was that of consequences being controlled by behavior.
>I have always considered this to be his most intellectually
>dishonest ploy.

You have to remember that he construed current control of consequences by
behavior (which he freely admitted) as itself having "come under the control"
of the organisms' history -- of consequences in the past.

>From another standpoint, the behaviorist COULDN'T characterize
>the experimental setup correctly. To do so would be to see that
>the stimulus is not an independent variable. The assumption is
>that the stimulus varies, and as a consequence of that the
>response varies. To measure the response, one arbitrarily varies
>the stimulus, so the stimulus has a known value or pattern that
>is independent of the behavior. If the stimulus is defined so it
>depends on the response, it's impossible to perform this
>manipulation (without breaking any actual feedback loop that's
>present).

I think you are putting words into the mouths of at least some behaviorists
(including Skinner), but if you can produce some documentation to support your
claim about what they say, I'm ready to be corrected. A while back, I posted
some quotes from Skinner which contradict the notion that he thought a
stimulus could not be affected by the responding organism. Note that even in a
feedback situation, the stimulus can be manipulated to a degree, because
control isn't perfect.

>Skinner saw the reinforcer as a consequence of behavior. But
>being unable to give up the idea that the environment controls,
>he then treated this consequence as an independent variable, and
>said that it controls the behavior. To be sure it controls only
>FUTURE behavior, but with his blind spot he never saw the obvious
>implication: that the BEHAVIOR which produces this consequence
>controls ("controls" meaning influences) the future behavior via
>the apparatus. To see this loop whole would have meant giving up
>the concept that the environment determines behavior, and that,
>above all, he was unwilling to do.

Seeing the whole loop, with BOTH environmental and organismic influences on
output, is EXACTLY the middle way between environmentalism and organismism
which I was arguing for some time back.

>If making the best predictions one can is not true scientific
>work, I don't know what is. Do you expect a full blown science of
>life as complete as that of physics to spring into being
>overnight?

No.

>Even worse, do you really think I expect it to?

To be truthful, I must say that sometimes you sound as if wishing can make it
so. I think it is important to be circumspect about one's enthusiasm, even
enthusiasm for a revolution. One important reason for this is so others won't
be tempted to mistake true advances for zealotry.

>>... maybe your model is already "noise"-limited. If so, then
>>adding a noise term should improve prediction of cursor
>>position IN A STATISTICAL SENSE, AVERAGED OVER MANY RUNS.

>No, it will make the fit worse. You will never do better with a
>noisy model than with a noiseless one. Noise adds in quadrature.

That's right. Chalk that mistake up to zealotry.

>No, the behaviorist will hell "Foul!" when I point out that the
>so-called stimulus is not an independent variable.

That's what I'm not so sure about. I wish we had a real behaviorist on the
net. Maybe Dennis could speak to this issue?

>Rick Marken (930108.1400)

>This is NOT the PCT problem with S-R models. We KNOW that they do not
>contain reference to only observables. In the model R=f(S) f (and probably
>also S if it is the sensory variable) is unobservable.

f is supposed to USE observables only, like cursor position and handle
velocity. Of course, f itself is a model, and I suppose you could then argue
that behaviorists actually propose generative models if they do any more than
curve-fitting (per the above, the Newtons of their day!?!?). Bill argues below
that behaviorists do NOT propose generative models, but I suppose that he is
claiming that all they ever do is curve-fit. And here I am trying to claim
that at least some behaviorists go beyond curve-fitting (like Newton did) and
yet use only observables in their "models" -- which aren't the sorts of models
I've always been thinking underlying generative models to be. I think curve-
fitting doesn't count as making generative models, but I'm not convinced that
hypothesizing functions containing only observables does count as making
underlying generative models.

>The PCTer's
>problem with the typical S-R model is that it says that behavior (R) is
>the last step in a causal chain that starts in the environment, at the
>sensory surface or in the brain (the latter being the reason we know that
>cognitive "models" of behavior are really S-R).

As I said above, I'd like some documentation on this claim. Maybe you can
direct me to some typical behaviorists to ask whether they think there can
never be ... S-R-S-R-S... situations?

>Bill Powers (930109.1530)

>Greg, I've just realized something that may clear up this whole
>argument. You've been proposing that in the pursuit tracking
>experiment, the observed relationship between handle, cursor, and
>target is H = k*int(T-C). I've been accepting that as true, but
>it isn't true. That is not an observed relationship; it's a
>HYPOTHETICAL relationship. It is, in fact, a generative model of
>the control system that would explain the observable behavior.

But is it solely at the level of the observed phenomena, or at the level of
underlying phenomena? I'd been thinking about this before you brought it so
sharply in focus. A "better" PCT model to predict cursor movement more
accurately would include some extra terms or equations containing hypothetical
internal-to-the-organism variables which would predict the occurrences of
"noise." On the face of it, such a model would (1) be generative, (2) be at a
lower level than the observables, and (3) include reference to hypothetical
variables inside the organism. But it isn't so clear to me that the standard
PCT tracking model used to date is either (2) or (3). And so the differences
between behaviorist "models" and PCT models aren't clear to me. But I'm still
mulling it all over.

I still think that attempting to make a PCT model to predict cursor movement
more accurately would be a worthy endeavor. Perhaps if you or Rick don't want
to try, you can send me your tracking program and I can attempt it.

As ever,

Greg