back to basics

HANS_BLOM1 · March 16, 1993, 12:20pm

[Hans Blom, 930316]

(Rick Marken (930311.1500))

Could you draw me a
diagram of the Q control system? I really want to see how the Q
control system works so I can simulate it on my computer.

Regrettably, diagrams are not a very convenient language to describe
optimal control systems; mathematical formula's are the standard tools that
are employed. Those, however, can certainly be translated into computer
simulations. A classic is still Feldbaum's "Optimal Control Systems", but
any more modern textbook on optimal control systems will provide you with
the formulae. Translation into a computer simulation is straightforward.

Adaptive control theory, a simplification of optimal control theory, is
easier to put into a diagram:

···

---------------------
         ------->| system to |-------
         > > be controlled | |
         > --------------------- | outside world
         > > -----------------
         > --------------------- | inside controller
actions |------>| model of system |<-----| perceptions
  u(k) | | to be controlled | | x(k)
         > --------------------- |
         > > > >
         > --------------------- |
         >-------| control system |<-----|
                 > >
                 ---------------------

Much simplified (no-noise case): The control system issues actions u(k) and
observes responses x(k). The model is any well-chosen formula such as

x(k) = a + b*x(k-1)+c*x(k-2)+... + p*u(k-1)+q*u(k-2)+...

whose parameters a, b, ... p, q, ... must be tuned such that the model's
response will approximate the system's response as closely as possible.
This tuning is achieved through a correlation of actions and perceptions,
both of which are provided to the model. This part is called "system iden-
tification" or "parameter estimation"; an on-line technique for parameter
estimation is, for example, the Extended Kalman Filter. Then, given the
resulting 'knowledge' of the system to be controlled in the form of the
estimated parameters, the model formula can be used either to predict
future x'es given u's or future u's given x's.

                 --------------------- system response
         ------->| system to |----------->
         > > be controlled | |
         > --------------------- --- +
         > ---------------| | comparator
         > > tuning --- -
         > --------------------- |
actions |------>| model of system |----------->
  u(k) | | to be controlled | model response
         > ---------------------

The above diagram may also help. The system's response is compared to the
model's response and the difference is used to tune the parameters of the
model. Thus, 'reorganisation' is considered to be a 'purposeful' feedback
process guided by the error between expectations and what really goes on,
rather than the 'random' process that Bill Powers proposes. System identi-
fication in an adaptive control system is therefore a (learning) feedback
loop within a (controlling) feedback loop.

The above scheme may explain what perceptual 'information' is used for: to
build/tune the model, i.e. to learn. Initially, a baby's model is immature;
the baby generates more or less random movements. But through observation
of their perceptual effects, slowly the model is built up, and it is fine-
tuned forever later in life. An important consequence is that parts of the
model may become so perfect (because some aspects of life are so predict-
able) that little or no information (perception) is required while yet
(almost) perfect actions are generated.

Back to basics. I do not understand the following section of yours
(930313.1000):

What PCT says (in no uncertain terms) is that this assumption [that
sensory input is an INDEPENDENT VARIABLE] is FALSE -- sensory input
is NOT an independent variable in a living control system (ie. all
living organisms) -- it is a DEPENDENT variable (the independent
variable being the reference signal inside the organism). I think
it is clear that, if PCT is right about this then the WHOLE edifice
of behavioral science comes crashing down -- somwthing up with which
most behavioral scientists will unquestionably not put.

There is obviously not much room for a compromise resolution to this
argument. PCT is just inherently revolutionary ...

This must have to do what you earlier called "the central insight of PCT".
This position is not revolutionary, however. Bishop Berkeley already
exposed the view that sensory input is a dependent variable in his "A
treatise concerning the principles of human understanding", published in
1710. His philosophical position is called a strict empiricism: "esse est
percipi", to be is to be observed. His position leads to propositions like
"the tree that I see over there exists because I see it". Other philoso-
phers have never found philosophical fault in this position: it is
logically consistent and therefore both respectable and tenable. However,
strict empiricists are not liked very much. People seem to resent being
considered a figment of the other's perception.

That goes for me, too. I respect your point of view, but if you really
believe that the universe up there at night has no existence independent
from your perception, then there is very little that I can meaningfully
discuss with you: there would be just too little common language/under-
standing.

However, in a later post (930314.1800) you react to Bill Powers
(930314.0900):

So how DOES information get into the organism from the
environment? Clearly, only through uncontrolled perceptions.

Brilliant; could this be one sideways contribution of IT to PCT?

Your post makes me realize that I should say that only CONTROLLED
sensory input variables are dependent variables; uncontrolled
sensory variables (like the moving target in a pursuit tracking
task) are still not independent variables; maybe a good name for
them is INFORMATIONAl VARIABLES; these are perceptions that are "just
there".

Now suddenly the view that there are also uncontrolled perceptions seems
acceptable to you, even brilliant. How should I read this? To me it seems
as if you are logically inconsistent here. Is that intentionally so or do
you retract your earlier opinion at this point? It is important for me to
know this, i.e. to understand you, i.e. to establish a correct model of you
in my mind. Am I trying to have a scientific discussion with someone who
has mutually inconsistent views, or are you in a learning process, where
you change your views when error (what could that be? disagreement with
Bill Powers?) develops? What IS the central insight of PCT for you?

Best,

Hans

Marken_Richard_S1 · March 16, 1993, 7:01pm

[From Rick Marken (930316.1100)]

Hans Blom (930316) --

I'll let the good cop handle most of this; I do have a
job, you know. But this stuff is just to juicy to lay
off of completely.

I said:

Could you draw me a
diagram of the Q control system? I really want to see how the Q
control system works so I can simulate it on my computer

I made this request after Hans said that perception is not
always what is controlled by a control system. So Hans
has complied with some diagrams. Unfortunately, Q never appears
in these diagrams -- and Q was the non-perceptual variable that
he said was controlled. All that Hans sent were diagrams of an
optimal control system which controls a perception that represents
the difference between the actual and modelled effects of a system
on a controlled variable. Nu? Looks like control of perception to me.

The above scheme may explain what perceptual 'information' is used for: to
build/tune the model, i.e. to learn.

Oy vay. Another control theorist who thinks that there is perceptual
"information" that can be used to guide ("build/tune") behavior. I
won't even ask ("where is it?", that is).

Hans says:

Back to basics. I do not understand the following section of yours
(930313.1000):

What PCT says (in no uncertain terms) is that this assumption [that
sensory input is an INDEPENDENT VARIABLE] is FALSE -- sensory input
is NOT an independent variable in a living control system (ie. all
living organisms) -- it is a DEPENDENT variable (the independent
variable being the reference signal inside the organism). I think
it is clear that, if PCT is right about this then the WHOLE edifice
of behavioral science comes crashing down -- somwthing up with which
most behavioral scientists will unquestionably not put.

There is obviously not much room for a compromise resolution to this
argument. PCT is just inherently revolutionary ...

This must have to do what you earlier called "the central insight of PCT".

You got it.

This position is not revolutionary, however. Bishop Berkeley already
exposed the view that sensory input is a dependent variable in his "A
treatise concerning the principles of human understanding", published in
1710.

Really? I thought Berkeley was saying that perception was all there is;
that we have no direct access to "boss reality" (which I happen to agree
with, by the way). But this has nothing to do with the idea that
sensory input is a dependent variable.

In conventional behavioral science (and in your statement about
information in perception) perception is treated as an
independent variable (independent of the actions of the system)
and behavior is the dependent variable so:

b = f(p)

behavior, b, is a function of input perception, p.

PCT says NO; p is a dependent variable -- viz.

p = f(r)

where r is a reference input, usually inside the organism itself.
This formulation is sometimes verbally accepted but, as is obvious
from the ongoing conversation about information in perception, not
thoroughly understood.

That goes for me, too. I respect your point of view, but if you really
believe that the universe up there at night has no existence independent
from your perception, then there is very little that I can meaningfully
discuss with you: there would be just too little common language/under-
standing.

This is all completely irrelevant; nothing I ever said could possibly
be construed as advocacy of solipcism (though you have apparently
managed to so construe it -- so it is possible); and solipcism has
nothing to do with PCT. Just for the record, I believe very firmly that
there is a "boss reality" determining my perceptions and constraining
the effects I can have on them. Our current best notion of the idea
of this boss reality is embodied in our models of physics and
chemistry.

No stone kicking necessary, see.

Now suddenly the view that there are also uncontrolled perceptions seems
acceptable to you, even brilliant. How should I read this?

I really don't know Hans; when did I ever give the impression that
that I didn't believe that there were uncontrolled perceptions?

Best

Rick