Constancy, Information Segregation

[From Rick Marken (960723.1300)]

Hans Blom (960722b)--

If I know that some variable _is going to be constant_ for a while, I
can use that knowledge in (predictive) control.

What do you mean by _know_? Is it the way Santa _knows_ if you've been
bad or good (that is, perfect _knowing_)? Or the way you _know_ that it
won't rain today (that is, imperfect knowing)? I don't think you can ever
_know_ perfectly that a variable is going to be constant for a while unless
you are god. And let me tell you, I knew god, god was a good friend of
mine and you, sir, are not god;-)

We mortals come closest to _knowing_ that a variable will be constant for a
while (the way god knows that a variable will be constant for a while) when
we have that variable under control. And even then, we can't be completely
sure what the variable will do in the next instant.

In the design of a control system, it is important to know the environment
in which it will operate, just like an organism can only survive in its
ecological niche.

This is true, but not for the reasons you think it's true. We don't need to
be able to predict the behavior of environmental variables (disturbances) in
order to design systems that control; we just need to know the amplitude and
frequency range of these disturbances. Organismic control systems evolved to
control in the context of a certain range of disturbances; when a disturbance
happens to exceed this range (you get caught in a tornado) you can no longer
control variables influenced by that disturbance.

Martin Taylor (960723 1430) --

Yes, I do claim that the control system uses information about disturbing
_influences_. What, from prior interactions with Rick, I understood the
statement to mean, was that I claimed the control system needs to know what
the disturbing influence _is_. I don't. It doesn't.

In the expression

p = a.1*o + b.1*d1+b.2*d2+...b.n*dn

I would say that the disturbing _influence_ on the perception (p) is:

b.1*d1+b.2*d2+...b.n*d.n.

The disturbing _variables_ are

d1, d2...dn.

Does that seem right?

To restate: Yes, the control system "needs" information about the disturbing
influences, and it cannot work otherwise.

OK. So the control system needs information about b.1*d1+b.2*d2+...b.n*dn
but not about d1, d2...dn. Is that correct?

No, this does not imply that the control system segregates this information

I don't understand this. What is "segregation of information"? If I need
information about

b.1*d1+b.2*d2+...b.n*dn

then I need information about that value. Are you saying that I get
information about

a.1*o + b.1*d1+b.2*d2+...b.n* dn

but that I don't segregate the information about

b.1*d1+b.2*d2+ ...b.n* dn

out? Please explain.

Best

Rick

<[Bill Leach (690728.2035 EDT)]

[From Rick Marken (960723.1300)]

Hans Blom (960722b)--
If I know that some variable _is going to be constant_ for a while, I
can use that knowledge in (predictive) control.

Rick:
What do you mean by _know_? Is it the way Santa _knows_ if you've been
bad or good (that is, perfect _knowing_)? Or the way you _know_ that it
won't rain today (that is, imperfect knowing)? I don't think you can ever
_know_ perfectly that a variable is going to be constant for a while unless
you are god. And let me tell you, I knew god, god was a good friend of
mine and you, sir, are not god;-)

Rick, adaptive (model based) control does work. Systems that "build" models
(really more like alter existing designed models though some of the
alterations can result in models that the original designers would not
recognize) of their "world" do successfully control and (as a rule--at least
eventually) control better than "traditional" controllers were able to
control (where such systems are typically used).

The real issue here is not the apparent "mumbo jumbo" that Hans sometimes
seems to be using to express how these "predictions" work but rather does
any of the model based theory apply to PCT and if so how?

I think that the answer to the first part of that question is yes (and Bill
Powers has also stated that he believed that model based control concepts
could be a part of how living systems control).

The big issue (with Hans and maybe others) exists, I think, in the answer to
the second part of the question. I think it is safe to say that at the
present state of PCT experimental work that the answer is... we don't have a
clue and clues, even if we had them are useless at this point.

This is exactly the same sort of problem with the issue of "identifying" the
"hard wired" portions of the system. I don't think anyone argues that there
are such "predesigned" portions but it appears that many fail to see the
dangers associated with "rushing to identify" them.

I am not really even trying to claim that Hans' "top down" approach is
necessarily as wrong as is the attempt to "cast in concrete" what prebuilt
systems exist. I just think that it is impossible to accomplish the sort of
design that Hans proposes at our present level of knowledge (humanity's
knowledge not just PCT's). Additionally, it seems to me that there is
greater danger that the resulting design could deviate in some significant
way from the living system and by the time the error was discovered there
would be no practical way to determine what design feature was a fault.

Our "bottom up" approach requires a "powerful resistance" to either jumping
to conclusions about what "pre-existing" hardware may be present or adding
complexity to the layered system of simple control loops prior to failure of
such simple loops to match experimental evidence (and I think that this last
part may well be the single most important concept).

Even now, I think that we sometimes forget that while our simple model works
exceedingly well in the sort of experiments that we have so far conducted,
our assumption that our implementation of many "parallel" control loops that
exist in the living system with a single control loop for each level may
hide dynamics that could be significant in future work. That is, it seems
reasonable to me that in the future, if we forget this basic assumption
(replacing parallel control loops with a single "averaging" control loop) we
might begin looking for solutions that are far more complex conceptually
than exists in the "real" system.

If one can't tell already, even though I enjoy the discussions about model
based control and the like, I am absolutely convinced that our first duty to
the theory is to prove through exhaustive effort that some combination of
the simple structure can not explain observed behaviour before "leaping" to
more complex solutions. To me, just the single lesson that an array of
analog control loops with partially in phase interactions need NOT perform
ANY calculations in the classic sense NOR even have ANY "cross connections"
or "cross awareness" to achieve the amazing range of control that our
experiments have already demonstrated, tells us that we have a far too much
to learn about combinations of simple control loops for us to be putting
together any combinations of "complex loop" systems much less any "grand
model theory models".

Martin Taylor (960723 1430) --

Yes, I do claim that the control system uses information about disturbing
_influences_. What, from prior interactions with Rick, I understood the
statement to mean, was that I claimed the control system needs to know what
the disturbing influence _is_. I don't. It doesn't.

Rick, you are confusing the OBSERVERS formula (and Martin's discussion of same).
We talk about the effect of "o" and "d" on the CEV and therefore upon "p".
Information Theory IS valid for such discussions and that is what I perceive
Martin to be saying (yet again).

I think that we really beat this one up a lot but unless my memory is
hopelessly failed:

1. Martin agreed that the control system itself "is not aware of any
   information in the perception" that is under control.

2. IF the observer could access the perceptual signal AND has some other
   relevent knowledge such as the actual system output value (and he was
   considerably most specific than I am being here now), then IT could
   extract information concerning the disturbance from the perception.

Qualitatively this is not a bit different than our own use of a disturbance
value and transform in the PCT experimental work. In either case, the
analyist knows things about the disturbance and the perception that the
control system (living or model) does not and indeed can not (if it is
closed loop negative feedback control).

Martin:

To restate: Yes, the control system "needs" information about the disturbing
influences, and it cannot work otherwise.

Rick:

OK. So the control system needs information about b.1*d1+b.2*d2+...b.n*dn
but not about d1, d2...dn. Is that correct?

Martin:

No, this does not imply that the control system segregates this information

Rick:

I don't understand this. What is "segregation of information"? If I need
information about ...

What Martin is saying is that THERE IS NO DIFFERENCE IN SOURCE DETECTED by
the control system itself such as is implied by OUR designations "bn" and
"dn" even though both Information Theory and Control Theory assert that
perception is made up from both. And more specifically his assertion that
the control system "needs information about" the disturbance is absolutely
correct as long as one does NOT try infer that "needs information about"
means more than is implied.
IT does NOT imply that someone looking just at the perceptual signal, with
_NO_ other information and knowledge about the system, can determine
ANYTHING about the disturbance(s).

Just as we can take a data set containing all of the parameters and values
for an experimental run except the value of the disturbance and then
reproduce the disturbance curve (as can be done for any single missing term
in the equation set), IT can be used for the same purpose. The difference
being that IT is essentially the theory that addresses the nature of the
signals that we deal with in Control Theory.

In its purest sense, Control Theory recognizes the very operative nature of
the effect of the disturbance upon the perception but does not attempt
normally to analyze the effects of that disturbance seperately from the
effects of the system's own output. Usually the analysis centers upon the
force generated by the control system vs the force generated by the
disturbance AT THE CEV (or the observers view of the net effect upon the CEV
anyway).

Everytime we start talking about the effect of the disturbance and that such
an effect is "cancled" by the operation of the control loop we are
immediately talking about something that touches upon the IT issue that is
so dear to Martin's heart. The fact that from a practical matter (to us) the
magnitude of the disturbance effect "seen" in the error or perception is
related to the receprical of the loop gain and thus small enough to ignore
for our purposes... Or the fact that without full knowledge of all other
system parameter values we can not distinguish between variations due to
disturbance vs other variations (changes in system gain, changes in PIF
transfer functions, transient response errors, etc.) does not mean that
either the IT assertion that "information about the disturbance is present
in the perception" or the Control Theory assertion that the perception will
be "offset" by the disturbance by the amount necessary to produce enough
error signal value for the output function to produce its' counter force
(and yes I know that I am ignoring "reset" but this statement is already too
complex).

In other words, even though you are probably right, that Martin's short
statements on this subject are probably often confusing to many people, his
statements are otherwise correct already!

bill leach
b.leach@worldnet.att.net
ars KB7LX