Pigs and Princeton; graduate students

[From Bill Powers (950513.0730 MDT)]

Martin Taylor (950512 14:00)--

Hans Blom said:

    If the control system had full information about all the
    fluctuations of all the disturbances, it could control perfectly.
    Indeed, "control" would not be necessary.

and I said

This is the same myth that Martin Taylor keeps proposing.I call it a
myth not because it isn't true in principle, but because to achieve
this "perfect control" by precalculation, you would need an infinitely
large and fast computer with unlimited memory ...

and you said

     I'm sorry you continue to call this a myth, since I thought we were
     in complete agreement that both what you say AND what Hans says are
     correct.

I don't see how what Hans said and what I said can possibly be
compatible. What Hans and you are saying is "If pigs had PhD's they
could, in principle, teach at Princeton." There are not and will never
be infinitely fast and capacious computers of any sort, organic or
inorganic. To speak of what they "would" be able to do is empty
nonsense, the modern equivalent of debating how many angels could dance
on the head of a pin (since angels supposedly have supernatural powers,
nothing is beyond them...).

Saying that disturbances could in principle be represented in a world-
model does not mean that if disturbances ARE resisted, they MUST HAVE
been correctly represented in a world-model. What you and Hans seem to
be overlooking is that we can and have set up experimental situations in
which a nonrepeating arbitrary disturbance is applied to a controlled
variable in a way that makes the disturbing variable unperceivable by
the person doing the experiment. After the person has learned to control
skilfully with that pattern of disturbance (which IS a matter of
adaptation), we find that if the pattern changes, the person does not
need any period of practice to restore control; control simply continues
without interruption. In fact, the kinds of disturbances we use can only
be seen as patterns that are in a continual state of unpredictable
change. Learning to control is not a matter of memorizing patterns of
disturbance. It is a matter of creating a specific physical
organization.

Even more to the point, we can set up models in which there is no
provision at all for dealing with any particular pattern of
disturbances; the model begins controlling immediately. Within the
envelope of maximum frequency and maximum amplitude that the control
system can deal with, control will occur for any pattern of disturbance,
regular or irregular, predictable (in principle) or unpredictable,
varying or constant. And we can explain in detail how this comes about.

The only counter to this argument is to start broadening what is mean by
"knowledge" and "prediction." But that ploy doesn't work; the point is
not what you call the processes, but what actual physical processes have
been put into the model. If the model contains no provision for
adaptively changing its parameters, and if it contains no mechanism for
extracting information of particular kinds, then it is a different kind
of model from the adaptive/predictive/information-based model, and it
deals with disturbances in a different way. To insist otherwise is only
to show your intention to keep explaining things in a particular way no
matter what anyone says.

There is one last point. A control system can become organized on the
basis of reference-signal changes alone, without ever experiencing a
disturbance from the outside world. My Artificial Cerebellum model works
that way, learning to control the position of a load that is a mass on a
damped spring (even if mass, spring constant, and damping change),
simply by moving the mass arbitrarily around. And once it is organized,
it will counteract the first external disturbance that ever acts on the
controlled variable. It needs no experience at all with external
disturbances in order to counteract their effects on the controlled
variable.

···

---------------------------------------------------------------------
isaac kurtzer (950509.2100)--

     throughout the course of this forum, and before, there have been
     concerns by those who feel that the study of living controls
     systems deserves considerable merit that this has not come to
     academic fruition... where the hell are the graduate students.

Good question. We have only recently heard about Len Lansky's honors
course on PCT, but how many others are teaching courses like that,
either undergraduate or graduate? I know that Clark McPhail teaches his
sociology course basically as a course in PCT (or so I assume from what
he has written), and that Dick Robertson fought at NIU until he was able
to make PCT an alternative course to standard psychology -- as long as
he was teaching. And Tom Bourbon did that as long as he could. But is
there anyone else systematically teaching PCT as the primary subject, as
a scientific course on human or animal behavior?

Not to make any extravagant claims, but didn't Christ say to his
followers that they should abandon all that is dear to them and follow
him? I feel similarly about PCT. The fainthearted are not going to put
PCT onto the map of science. We need people who are willing to stand up
in their departments and say "I am scheduling a three quarter course in
PCT, and plan to follow it up by designing a curriculum leading to a PhD
in Perceptual Control Theory. Put me in the course announcements.
Goddammit!"

When you give serious consideration to this, doesn't it tell you
something about academia and science? Paddling against the mainstream is
a scarey prospect when you really think about doing it. It isn't enough
to have a new approach; what you need is clout, other people who will
stand up with you and face down the nay-sayers. To get a new scientific
idea into circulation, you need more than a good idea: you need
political savvy and power. Or you need courage.
----------------------------------------------------------------------
Best to all,

Bill P.

[Martin Taylor 950515 12:10]

Bill Powers (950513.0730 MDT)

Hans Blom said:

    If the control system had full information about all the
    fluctuations of all the disturbances, it could control perfectly.
    Indeed, "control" would not be necessary.

and I said

This is the same myth that Martin Taylor keeps proposing.I call it a
myth not because it isn't true in principle, but because to achieve
this "perfect control" by precalculation, you would need an infinitely
large and fast computer with unlimited memory ...

and you said

    I'm sorry you continue to call this a myth, since I thought we were
    in complete agreement that both what you say AND what Hans says are
    correct.

I don't see how what Hans said and what I said can possibly be
compatible.

It's very disconcerting to have a relatively careful discussion, to come to
an agreement, and then within a very few days find that the agreement has
been completely forgotten. I really don't want to start all over again.
Just, please, if you don't agree with yourself of a few days ago, tell us
what has changed in your mind. Please?

There are not and will never
be infinitely fast and capacious computers of any sort, organic or
inorganic.

That's a small part of what we agreed. So why bring it up now as a
criticism of a "myth?"

To speak of what they "would" be able to do is empty
nonsense, the modern equivalent of debating how many angels could dance
on the head of a pin (since angels supposedly have supernatural powers,
nothing is beyond them...).

Unfair to appropriate my own metaphor, which I used in exactly the same way,
as a criticism.

···

---------------------

Saying that disturbances could in principle be represented in a world-
model does not mean that if disturbances ARE resisted, they MUST HAVE
been correctly represented in a world-model.

I have not said that, if you mean by "world-model" an explicit representation
of the world. But I have said, and I interpreted Hans as saying, that
the properties of a control system constitute a model of the world in
which it functions. I think you agree that a control system with
inadequate bandwidth or strength to counter the disturbances applied to it
will be an unsuccessful controller. In my (our?) terms, it is a control
system with an inadequate world-model, that model being implicit in the
structure of the control system.

The only counter to this argument is to start broadening what is mean by
"knowledge" and "prediction." But that ploy doesn't work; the point is
not what you call the processes, but what actual physical processes have
been put into the model. If the model contains no provision for
adaptively changing its parameters, and if it contains no mechanism for
extracting information of particular kinds, then it is a different kind
of model from the adaptive/predictive/information-based model, and it
deals with disturbances in a different way. [False ad hominem statement
omitted].

If the model doesn't adapt, then it was designed. The designer used
"knowledge" and "prediction" to set the unchanging parameters of the
control system. The information used is, in part, about the kinds of
disturbance to be countered, and, in part, about the kind of world to
which the control system's output will be applied. And the predictions
had better be right, or the control system might well be overwhelmed
by the first little disturbance to come along.

You can't avoid the issue by that ploy.

=============

There is one last point. A control system can become organized on the
basis of reference-signal changes alone, without ever experiencing a
disturbance from the outside world.

I'm not sure where this comment comes from. It's a given that an adaptive
control system will apply its own variation to the world (i.e. signals
generated independently of the outer world--reference signal variation
does just fine). Disturbances from the outer world do fine, too, so long
as the control system is actually applying output to the part of the
environment perceived by the PIF of the control system.

In my old discussions of the "kinds of perception," what you are talking
about was labelled "exploration." Exploration is conducted so that:

once it is organized,
it will counteract the first external disturbance that ever acts on the
controlled variable.

There is a separate point, on which I am explicitly NOT commenting at this
time, and that is the question of whether temporal regularities in the
disturbance can usefully be modelled (either explicitly or in the control
system's structure). Variation in the reference, or other internally
generated variation, will be of no value here.

Martin