reply to Bill Leach

[HANS_BLOM2]

[Hans Blom 950511b]

(Bill Leach 950511.01:17 U.S. Eastern Time Zone)

In absolutely _NO_ real world application can a purely model based
system achieve control (that is work inspite of the presence of
disturbance) -- at the _very_ least ALL mechanical output must be
either negative feedback control or always operated against fixed
physical limits.

Why must that necessarily be so? I can imagine -- a thought
experiment, I admit it -- that what you call "disturbances" are, to a
very large extent, predictable (through "extrapolation" of earlier
perceptions) and that the "unmodeled dynamics" could therefore be
insignificant -- and zero on average. Because I can imagine this, it
is a belief, which might turn out to be true or false. At the least,
it would require an extremely complex model. It would also require
that we do not attempt to control too many dimensions of the world.

What data support your absolute claim? Is it more than a belief?

In contrast, I claim that I can design a small artificial world and a
model-based controller that survives successfully in it. Make the
artificial world slightly more complex, and the controller might
survive marginally. Make the artificial world even more complex, and
the controller might not survive any more. Artificial Life
simulations show the same. What makes you so "absolutely" certain?

Such system can and indeed have proven to be very useful and
productive but they are "control systems" only in the very loosest
use of the term and are not related to living system control.

Again, how do you know? Experiments with food finding by bees make me
believe that they have something like a mental map. If they are
forced to find food in a very unlikely location, say on a small
artificial island in the middle of a large lake where they normally
do not go, they reliably go back there to look again. I cannot
explain such a finding without postulating some kind of internal
model.

In the first place, as I envision possible "model control", the
model would not do any "calculations". The "model" is probably
created "as a matter of course" just from the fact that the organism
IS normally controlling perception.

Certainly. A "model" is a concept in our human minds, an abstraction.
You will not find a "model" in any mind, wherever you look. It is not
an anatomical but a psychological (or physiological) notion. A model
does not "do" calculations. It arises as a result of "calculations"
through more basic mechanisms.

Neither "is" an organism a control system. But one can LOOK AT an
organism in its world AS a control system.

This is a matter of looking at the thing from a different level, just
like a table is both a collection of wooden boards and a collection
of atoms. Sorry to annoy you with some basic ontology ;).

What do you mean by "FORWARD kinematics and dynamics representation
..."?

I mean that the "world" in the demo contains an equation

xt := ct + at * xt + bt * u + ft

and that the world-model contains an equation

x := c + a * x + b * u + f

where every variable directly stands for something in the real world,
and not through some kind of reverse transformation.

                        It is _FLATLY_ impossible to CONTROL for

something that is not sensed but is subject to disturbance.

If the disturbance is unpredictable, sure. But look like it from a
different perspective: you might not NEED to control for something
that you perceive. Bill added a sine- wave as a disturbance to my
model-based controller that was not resisted. We observe a
sinewave-like variation of light intensity through the 24 hours of
the day that we do not attempt to control away. So is the sinewave
something to be controlled away? From Bill's perspective, yes. MY
INTERPRETAT-ION OF EXACTLY THE SAME THING COULD BE THAT THE SINEWAVE
IS OBSERVED BUT, SINCE IT IS NOT MODELED, IS IRRELEVANT AND NEED NOT
BE RESISTED. Does this interpretation strike you as less applicable?

The idea that we might really use a model based system to _attempt_
to achieve (or maintain) a perception in absence of the perceptual
input, even for short periods of time does not in any way suggest
that living being are not fundamentally and basically negative
feedback control systems. Model based "control" isn't... and if
Hiesenburg (sp?), Bohr and the likes of QED are even close in
principle then IT CAN NOT control.

Heisenberg. But I cannot make sense of what you try to tell me here.

The problem however is that model based systems have no way of
knowing when they fail since they do not perceive the controlled
variable -- sorta like the psychotic living in non-existant castles
or the acid head "flying like a bird" from atop a twenty story
building.

As my demo shows (at higher system noise levels), it is not necessary
to have a correct model (i.e. model parameter = world parameter) for
control to be (almost as) good. Goes right against the grain, doesn't
it?

Greetings,

Hans

[Bill_Leach1]

<[Bill Leach 950512.00:12 U.S. Eastern Time Zone]

[Hans Blom 950511b]

... disturbances are, to a very large extent, predictable ...

In the engineered control system field this is largly true because the
environment is strictly limited. When some disturbance occurs that the
control system can not handle then generally "something is considered to
be broke".

My statement was probably too broad but as far as I am concerned, you
answer your own question when you say "It would also require that we do
not attempt to control too many dimensions of the world."

Like Bruce, I believe that study of "world model control" methods IS
useful and can probably even lead to good information for HPCT.

bees

I am NOT in disagreement with this. I have "maps" of some sort in my own
mind that I am quite conscious of not to mention that there are likely
many thousands of others that I am not aware of in existance.

However unless your bees frequently fail to return to their food (or hive
for that matter) then they are using some sort of perceptual control --
their model for instance is not likely an inertial guidance system and
therefore can not compensate for drift do to wind changes unless there is
a perception to be controlled that would be affected by the wind.

Neither "is" an organism a control system. But one can LOOK AT an
organism in its world AS a control system.

I am afraid that you probably find yourself at severe odds with most of
us here -- the belief (strongly supported by sound evidence and refuted
by none currently known) is that living systems ARE closed loop negative
feedback control systems. Not that they _may_ be viewed as such but
rather that they _ARE_ such.

That "world models" exist and that such may be used at various times for
various reasons while debatable in detail is probably otherwise accepted.

This is a matter of looking at the thing from a different level, just
like a table is both a collection of wooden boards and a collection
of atoms. Sorry to annoy you with some basic ontology ;).

No annoyance but what you are saying is similar to the idea that
behaviour analysis can be done using "lineal cause and effect" analysis
because lineal cause and effect is ALWAYS valid. Well, yes it is always
valid in that physical systems obey (without exception that we know of) a
set of relatively simple physics principles. Feedback Control Systems
are no exception but it is equally true that application of lineal cause
and effect analysis to such a system provides only an "instantaneous snap
shot" of the system which actually obscures any chance for an
understanding of what is "going on".

So yes, you can say that a disturbance "stimulates" a control system to
respond and a correctly timed lineal analysis will show this to be
"true" but such an analysis entirely misses what is actually going on and
that of course is the phenomenon of control. Again, many forms analysis
may be correct in that they produce scientifically valid results but not
necessarily valid for all purposes (that is it is not so difficult to
attempt to extend knowledge obtained through "valid scientific method" to
domains where assumptions have made the extension untrue (much less for
those situation where then initial "knowledge" is not obtained through
scientific method in the first place).

It is _FLATLY_ impossible to CONTROL for something that is not sensed
but is subject to disturbance.

If the disturbance is unpredictable, sure. But look like it from a
different perspective: you might not NEED to control for something
that you perceive. Bill added a sine- wave as a disturbance to my ...

We must have really been talking past one another on this one. I don't
know of anyone that questions the idea that there are perceptions that
are not controlled. That a perception exists or even when a control
system exists for a perception there in no necessity that such
perceptions much be controlled.

I agree that disturbances may not matter in several ways. In the first
place disturbances to uncontrolled environmental variables are from the
organisms point of view not disturbances at all but only observed
behaviour of the variable (that is perception of course, not the variable
itself).

In the second place even disturbance to a CEV are not resisted unless the
CEV is perceived to vary enough from the reference to cause a change in
error signal. For example when crossing the Golden Gate Bridge in high
wind, I do not correct for position in the lane as the bridge swings side
to side unless the net motion is sufficient to cause the car to leave the
lane. There is more to that example than I am talking about here of
course.

MY INTERPRETATION OF EXACTLY THE SAME THING COULD BE THAT THE SINEWAVE
IS OBSERVED BUT, SINCE IT IS NOT MODELED, IS IRRELEVANT AND NEED NOT BE
RESISTED. Does this interpretation strike you as less applicable?

This is a major source of our mutual problem. If the perception that is
to be controlled is still within its desired value limits in the presence
of the unresisted disturbance then OK, no problem.

However, I have a lot of trouble thinking that a living system can rely
on blind stupid luck and remain living. I can model (mentally anyway) a
means for me to cross between the roofs of two twenty story buildings on
a 2X4 but let me assure you that I would have absolutely NO intention of
actually trying to cross with the use of control in the normal PCT sense.

Heisenberg. But I cannot make sense of what you try to tell me here.

All I was getting at is that we have sufficient reason to believe that NO
amount of knowledge about the physical world will ever allow for exact
prediction. Thus, a model based system will always have error if run for
a finite period of time. It is true that the error may be acceptable or
even negligable for a particular application but there is no way to be
certain such will always be the case.

As my demo shows (at higher system noise levels), it is not necessary
to have a correct model (i.e. model parameter = world parameter) for
control to be (almost as) good. Goes right against the grain, doesn't it?

Not against mine, I have worked with "blind" controllers that were very
reliable and quite accurate (primarily because of the limited universe in
which they operated). For engineered systems, if you can prevent the
disturbances in the first place, it usually is best to do so and
eliminate the need for feedback control.

-bill