Evolving references

[From Shannon Williams (960103)]

Bill Powers (960101.1115)--

Suppose we're controlling the perceptual signal y at a value y0. The
error, y0 - y, is amplified and send to two output functions, which
affect the values x1 and x2 in appropriate directions. The result will
not be to bring x1 and x2 to any particular values, but to bring y to
the value y0.

I am excited! If X1 and X2 are considered potential perceptions for
another loop, then this is how you build an evolving reference.

See it this way:

ยทยทยท

****************
1) At first the infant has some disturbance (hunger, cold, ?)
2) If a perception fixes the disturbance then he tries to maintain that
   perception. This happens to decrease the likely hood of him
   encountering the disturbance again. In other words, a perception that
   proves successful becomes a reference.
3) The reason that #2 works is because perception generates (causes)
   output. In other words, there is a direct relationship between
   what you perceive and how you behave.

In other words:
***************
p = the (learned) perception/interpretation of an input.

1) y1 = e0 = y0 - p0(y)
2) y2 = e1 = y1 - p1(y)
3) y3 = e2 = y2 - p2(y)
etc.

or:
y0 = p0(y) + p1(y) + p2(y) + p3(y) + ... + pi(y) + ei

Also lets say you have another reference:
*****************************************
1) r1 = E0 = r0 - F0(y)
2) r2 = E1 = r1 - F1(y)
3) r3 = E2 = r2 - F2(y)

if it happens that E1 = e1 then r0 = F0(y) + F1(y) + p2(y) + ...

In other words an interpretaion used to fix y0 can be used to fix r0.

So the scenario is:
*******************
0) you start with an error signal e.
1) you have a perception p.
2) you have an output to the world O.

3) diferent p's generate (cause) different Os.

4) When an error signal occurs p is (internally) changed and changed and
changed until the error signal goes away. (O is linked automaticaly to p)
In other words: An interpretaion of the input from the world is found which
causes the error to go away. This interpretation DIRECTLY generates your
behavior.

5) If the interpretaion/perception, p, is successful again and again, then
p becomes a reference itself. So that if p is not maintained, then an
error is generated.

This is kinda what I see:

        > > > > >
       \|/ input \|/ from \|/ the \|/ world \|/
-------------------------------------------------------------------------
      ------------------ ------------------ ------------------
      > perception #1 | | perception #2 | | perception #3 |
      ------------------ ------------------ ------------------
        > > > > > >
       \|/ \|/ \|/ \|/ \|/ \|/
        r1----> C r2 ----> C r3 ----> C
                > > >
                e1 e2 e3
               \|/ \|/ \|/
------------------------------------------------------------------------
      ------------------ ------------------ ------------------
      > perception #4 | | perception #5 | | perception #6 |
      ------------------ ------------------ ------------------
        > > > > > >
       \|/ \|/ \|/ \|/ \|/ \|/
        r4----> C r5 ----> C r6 ----> C
                > > >
                e4 e5 e6
               \|/ \|/ \|/
-------------------------------------------------------------------------

1) When each perception above is realized, an output is automatically
generated. In other words, each perception is an input to some or many
loops. One of those loops eventually causes behavior.

2) All of the interpretations above can share hardware.

3) I imagine the references above to have evolved in the brain, but once
they have evolved they seem as concrete as the original references.

4) The reference could work like this: There is something that remembers
when it is not stimulated (or is stimulated?). As it gets to be stimulated
less and less, then there comes a point when it generates a signal if it
is stimulated. In other words, when a neural firing pattern becomes
recognizeable then when that pattern is not generated an error occurs.

Going back to the project:
**************************

Suppose that x1 is the magnitude of one formant and x2 is the magnitude
of another. We can assume that the same phoneme will be heard if some
function of x1 and x2 remains constant.

I think that how we perceive x1 and x2 depends on how we have learned our
interpretations. (Our interpretations are just one segment of the loop.)

By changing the reference level for the perceptual signal of the
phoneme-control system, a higher system can change the phoneme that is
heard.

I do not think that this is right. I think that our perceptions create our
references.

I think it's best to start by defining what
has to evolve; then we can see what aspects of the system need to be
modified during evolution.

OK.

-Shannon

(bob hintz - 2300 - 1/3/96)

I would like to try again in expressing the difference between control
exercised by information processes and control exercised by energy
processes. In a tracking exercise, if I am working the handle, I am
providing the energy that moves the handle (or mouse) in response to my
error signals. However, if you are working the handle, but cannot see
the screen, while I can see the screen, but can only tell you how to
change the handle, then I provide information about the error and you
provide the energy to correct it.

Babies cannot correct error by energy processes. They cannot acquire
food without assistance. They cannot cover themselves when cold. They
cannot change themselves or even move away from their own waste when
their skin becomes irritated. The only thing they can do in an effort
to control their own error signals is attempt to provide information to
a caretaker. Crying typically is treated as an indication of error from
the baby's point of view by the caretaker, but it initially provides no
information about the source of that error. A caretaker typically does
the reverse of "the test", in that s/he seeks to counter what might be a
disturbance. When the baby stops crying, one assumes that a disturbance
has been controlled. This presumeably helps the baby to distinguish
different sources of error. As the baby learns to distinguish hunger,
cold and skin irritation, it is possible to alter behavior for the
purpose of informing the caretaker about the internal experience of
error so that the caretake can counter the disturbance.

This would be very similar to the problem you and I would have if I can
see the screen and you are working the handle. I might try telling you
what to do, ie., push it forward or pull it back, and see what happens.
I might try explaining the relationship between what I believe happens
when you move the handle. I might try simply reporting error and let
you figure out how to move the handle. It is this kind of situation
that allows us to observe how language (information behavior) can
actually be used to organize activity (energy behavior) that makes a
difference when energy variables are being disturbed and it takes two or
more persons to control the perception.

take care, bob

[From Rick Marken (960103.1930)]

Shannon Williams (960103) --

1) At first the infant has some disturbance (hunger, cold, ?)

We have to be careful about terminology; a disturbance is a physical
variable that affects a controlled perception. Hunger and cold are
perceptions -- perceptual _variables_ really (eg. hunger ranges
from mild to famished). Glucose level is probably a disturbance
to the perception of hunger; the temperature of the air is one disturbance
to the perception of coldness.

2) If a perception fixes the disturbance then he tries to maintain that
perception.

Actions "fix" (control) perceptions. Perceptions don't fix anything;
they're just there -- happening; sometimes the perceptions that are
happening are the one's we want; sometimes they're not (because we
haven't got them under control).

In HPCT, the action that controls a perception can be the setting of
a reference for a lower level perception. But you can't control a
perception (compensating for any disturbances) by generating the same
action each time; what must be learned (in order to control) is a
_relationship_ between error and action. For example, the child might
find that a certain amount of crying leads to food. But the same amount
of crying won't always produce the food; the child has to learn a
relationship between error and crying; the relationship is probably:
increased error leads to increased crying. Once the child learns this
relationship it will continue to cry until the food is forthcoming -- or
learn a new way to get fed (control its perception).

This happens to decrease the likely hood of him encountering the
disturbance again.

Learning can't be based on a decrease in the likelihood of encountering
disturbance; disturbances (like shit) happen; what you learn is how to
compensate for any disturbance to a controlled perception; you learn to
control.

In other words, a perception that proves successful becomes a
reference.

I think you can learn to set references for perceptions which, when
produced (the reference level of the perception is produced, that is)
result in control of other perceptions. For example, I have learned
that by setting a reference for perceiving a clockwise twisting
force, I can perceive a door opening. What I have learned is to set a
reference for a perception of twist (clockwise rather than
counterclockwise) in order to control a perception of "openness".

The reason that #2 works is because perception generates (causes)
output.

If by "works" you mean "control" then the only "causal" relationship
that works is one between error (difference between perception and
reference) and output.

This is kinda what I see:

When each perception above is realized, an output is automatically
generated.

What you are seeing is, I'm afraid, not a perceptual control system.

Bill said:

By changing the reference level for the perceptual signal of the
phoneme-control system, a higher system can change the phoneme
that is heard.

Shannon replies:

I do not think that this is right. I think that our perceptions create our
references.

I don't see how this could be true -- even for a learning system. In PCT
terms, you are saying that the results you get (your perceptions)
determine the results you want (references for these perception). I
think we can show very easily that perceptions don't create references;
perceptions are controlled relative to references.

If you have a PC you really should try the demo1 program to get a good
sense of how control actually works. Once you understand how control
works, you can get into trying to understand how organisms might
_learn_ to control.

Best

Rick