hPCT Learning - Trying Again

Dear Bill,

I’ve reduced my idea to a flow chart, attached. I hope
this is adequate.

Flow charting it actually helped me get a better grip on the
concept, and to see some further ramifications of it, such as its application
in Imagination Mode.

I would still like to hear from other members of the group.

Ted

Flow Chart2.doc (58.5 KB)

···

From: Control Systems
Group Network (CSGnet) [mailto:CSGNET@LISTSERV.ILLINOIS.EDU] On Behalf Of Bill
Powers
Sent: Tuesday, August 17, 2010 9:55 AM
To: CSGNET@LISTSERV.ILLINOIS.EDU
Subject: Re: hPCT Learning

[From Bill Powers (2010.17.0914 MDT)]

At 06:24 PM 8/16/2010 -0500, Ted Cloak wrote:

I’ve attached some thoughts about Hierarchical PCT learning.

Don’t get mad, Ted, but I’ve instituted a new policy. I’m thinking of working
up a boilerplate paragraph to include in all replies to proposals about new
models or rearrangements of old models, something like this:
*Dear Sir or Madam;
Thank you very much for showing me your model. I would like understand how it
works. Can you explain to me what makes this model behave in the way you
describe, or send me a working computer program demonstrating it, with source
code? I have a Windows computer, but can read Pascal or C source code if your
program doesn’t run under Windows.
If you would like me to convert your description into a working model, I will
need more details about what specific function is to be implemented by each
component. I will also need specifications for the passive components affected
by the active ones. My fee for consulting and programming is $300 per hour or
fraction of an hour, with an invoice to cover the first 16 hours. Work will
commence on receipt of payment. If I am unable to produce a working model based
on your specifications I will bill you only for the additional time needed to
reach that conclusion.
Yours truly, etc.*Of course if you’ve already worked out all the specifications and produced
a working simulation (or hired someone to produce it), you won’t need my help,
and I’d be happy to evaluate the simulation and make any suggestions that come
to mind in the normal academic way of free exchange of information. But I’ve
decided that I can’t spend the time I have left trying to understand how a
model works that has never been demonstrated to behave as its author claims it
will behave. That’s really the responsibility of the originator of the model.

Come on, Ted, isn’t that a reasonable request?

Best,

Bill P.

[From Bill Powers (2010.08.21.1050 MDT)]

TC: I've reduced my idea to a flow chart, attached. I hope this is adequate.

Flow charting it actually helped me get a better grip on the concept, and to see some further ramifications of it, such as its application in Imagination Mode.

BP: It's good that you label your diagram a "flow chart," because that's what it is: things are done or accomplished according to which box you're looking at, but the system that is doing or accomplishing those things is not shown.

The key to recognizing a flow chart while you're looking at the box labeled "B: reference signal in register r undergoes Radiant Variation" is to ask what is happening right now in the box labeled "START: A addresses B (Register r)" If that question makes no sense,` you're probably looking at a flow chart or at least not at a system diagram.

In a system diagram of a control system, you can ask what the value of the reference signal is at the same time the input function is generating a perceptual signal with a present value that depends on a number of input quantities, at the same time that an error signal generated by a comparator receiving the perceptual and reference signals is driving an output function to produce a certain value of the output quantity, at the same time that the output quantity and a disturbance are combining their effects to create a particular value of the input quantities. All those things are being done at the same time by the various functions that make up the system, receiving and emitting signals or other physical effects continuously and without waiting for other functions to act.

The relationship between the two kinds of diagrams can be seen by asking what it is that accomplishes each of the processes described in the boxes, and what is required for that to be done. For example, in the flow chart there is a diamond-shaped box containing the question "B addresses C?" This decision requires a system that at least can sense B and C, and perceive the relationship between them called "addressing." And this system must generate one output when the questioned relationship exists, and a different output when it doesn't exist. The outputs must have physical effects on something else -- here either another system of the same kind, or a box labeled "pause" which is accomplished, presumably, by some sort of timer or delaying device.

Of course the key is seen, as in the famous TOTE units of Miller, Galanter, and Pribram, in the little box labeled END ("Exit" in the TOTE unit). What happens to all the other boxes when we reach this final box? They all disappear -- in fact, each one disappears as soon as we're finished with it and no other arrows lead back into it. This shows that the boxes do not represent physical entities as they do in a system diagram.

I hope now it makes sense when I say that a flow chart is a description of the effects generated by an undefined system somewhere in the background, the same system that is achieving the results described in each box. So the flow chart describes a behavior that requires explanation -- it is not an explanation of anything in itself, but a (proposed) description of observed phenomena that need an explanation.

Best,

Bill P.

[From Bill Powers (2010.08.21.1050 MDT)]

...

I hope now it makes sense when I say that a flow chart is a
description of the effects generated by an undefined system somewhere
in the background, the same system that is achieving the results
described in each box. So the flow chart describes a behavior that
requires explanation -- it is not an explanation of anything in
itself, but a (proposed) description of observed phenomena that need
an explanation.

[From Ted Cloak (2010.08.21.1600 MDT)]

Fine. But using your definitions, shouldn't we discuss whether the proposed
description is adequate, or at least plausible, before we try constructing a
system diagram to explain it?

Best
Ted

[From Bill Powers (2010.08.21.1605 MDT)]

Fine. But using your
definitions, shouldn’t we discuss whether the proposed

description is adequate, or at least plausible, before we try
constructing a

system diagram to explain it?

Yes. The diagram to the right is a system diagram; the flow chart on the
left pertains to it. But does the system diagram contain everything
needed to make the flow chart work? For example, where in the system
diagram is the section described by the two choice-diamonds? It seems
that there is nothing in the system diagram to go with the operations
specified in the flow chart. What is it that determines whether B
addresses C or A addresses B? What part of the system copies a reference
signal in register r to the comparator?
Also, the description seems to treat the reference signal as if it’s the
same thing as the memory address of the reference signal. A
“register” would seem to refer to an address; a signal in PCT
is one-dimensional variable and has no parameters except its magnitude.
So it’s not clear what the blind variation is causing to vary.
The main conclusion about learning isn’t supported by either the system
diagram or the flow chart. How does the system determine
“success?” When a stored memory is tried, how is it determined
that it gives a better or worse result? What result is being observed as
a criterion to judge success, and what is doing the observing and
judging?
And perhaps the most difficult: how does a system B “know” what
perception to control? This implies, as I said a day or two ago, a
complete control system that is able to pick different perceptions to
control. How would such a control system know there is more than one
perception it could control, when its input function defines the
perception and reports only the state of that one perception? A single
perceptual input function in current PCT can detect the presence of only
one kind of perception, and the perceptual signal it generates simply
reports how much of it there is. Your picture of perceptions and
reference signals seems critically different from mine. I won’t say mine
is necessarily right, but I don’t understand how yours would
work.

We’re on the fringes of an area where the details of PCT are vague. All
we can do, when there are differences like these, is to try to state our
propositions as clearly and unambiguously as possible, so we might see
ways of experimentally choosing between different possibilities. I don’t
think your proposed organization has been stated clearly enough – either
that, or I’ve lost another few dozen neurons.

Best,

Bill P.

Dear Bill,

I’ve reduced my idea to a flow chart, attached. I hope this is adequate.

Flow charting it actually helped me get a better grip on the concept, and to see some further ramifications of it, such as its application in Imagination Mode.

I would still like to hear from other members of the group.

Ted

···

From: Control Systems Group Network (CSGnet)
[mailto:CSGNET@LISTSERV.ILLINOIS.EDU] On Behalf Of Bill Powers
Sent: Tuesday, August 17, 2010 9:55 AM
To: CSGNET@LISTSERV.ILLINOIS.EDU
Subject: Re: hPCT Learning

[From Bill Powers (2010.17.0914 MDT)]

At 06:24 PM 8/16/2010 -0500, Ted Cloak wrote:

I’ve attached some thoughts about Hierarchical PCT learning.

Don’t get mad, Ted, but I’ve instituted a new policy. I’m thinking of working up a boilerplate paragraph to include in all replies to proposals about new models or rearrangements of old models, something like this:

*Dear Sir or Madam;

Thank you very much for showing me your model. I would like understand how it works. Can you explain to me what makes this model behave in the way you describe, or send me
a working computer program demonstrating it, with source code? I have a Windows computer, but can read Pascal or C source code if your program doesn’t run under Windows.

If you would like me to convert your description into a working model, I will need more details about what specific function is to be implemented by each component. I will also need specifications for the passive components affected by the active ones. My fee for consulting and programming is $300 per hour or fraction of an hour, with an invoice to cover the first 16 hours. Work will commence on receipt of payment. If I am unable to produce a working model based on your specifications I will bill you only for the additional time needed to reach that conclusion.

Yours truly, etc.

  • Of course if you’ve already worked out all the specifications and produced a working simulation (or hired someone to produce it), you won’t need my help, and I’d be happy to
    evaluate the simulation and make any suggestions that come to mind in the normal academic way of free exchange of information. But I’ve decided that I can’t spend the time I have left trying to understand how a model works that has never been demonstrated to behave as its author claims it will behave. That’s really the responsibility of the originator of the model.

Come on, Ted, isn’t that a reasonable request?

Best,

Bill P.

(Gavin Ritz 2010.08.22.12.35NZT)

Ted
I wanted to mention this too. I have a family member that I have analysed and watched for 20 years. This family member has serious debilitating social skills. So I attempted to use the TEST ie finding the PCV’s.

I first just tried just writing down the actions of this person without any
qualitative connotations, then with qualitative notions. I got over thirty actions of this person. However it became increasingly obvious the Test would not work because of something very interesting. One the language we use, two almost all language is fused with qualities and semantics. As soon as we go above a certain level in HPCT things go haywire.

I of course do know what this person is “controlling for” but I wanted to a bottom approach using HPCT. It could be done only top down as I knew exactly what perception this person is controlling. Only then did the PCV’s make any sense. So what I’m saying is there is nothing in HPCT that could lead me to the controlled
perceptual signals and associated input structures, reference signals and errors.

Kind regards
Gavin

[From Ted Cloak (2010.08.22.1630 MDT)]

I’m working on a system diagram. Meanwhile, there
are some things we need to chew on.

[From Bill Powers (2010.08.21.1605 MDT)]

BP: And perhaps the most difficult: how does a system B
“know” what perception to control? This implies, as I said a day or
two ago, a complete control system that is able to pick different perceptions
to control. How would such a control system know there is more than one
perception it could control, when its input function defines the
perception and reports only the state of that one perception? A single
perceptual input function in current PCT can detect the presence of only one
kind of perception, and the perceptual signal it generates simply reports how
much of it there is.

TC: This is a problem of terminology only. For reasons I
can no longer access, I have come to use the expression “This control
system controls perception P” to mean that P is the perception it is
attempting to obtain and maintain; i.e., its perceptual goal; not that P is
whatever perception is being handled by its Input Function.

BP: …a signal in PCT is one-dimensional variable and
has no parameters except its magnitude…

BP: Your picture of perceptions and reference signals
seems critically different from mine. I won’t say mine is necessarily right,
but I don’t understand how yours would work.

TC: I’m not sure how different. Is your picture
that every signal depicted in Figure 15.2 on page 218 of B:CP is a
one-dimensional variable with no parameters except its magnitude? That is
attractive because of its simplicity; single-parameter control systems, converging
from lower levels to higher levels, add up to the incredibly complicated
perceptions that every animal tries to control; and, going down, the incredibly
fine-tuned movements through which we attempt to control them.

But then what’s the use of each CS having a Memory,
if all that a Memory can contain is a bunch of different numerical values, of
which it can release, as a reference signal, only one at a time? Why won’t
Fig. 15.1, p. 209, suffice?

My assumption has been, apparently, that address signals
and error signals can be one-dimensional, but that perceptual signals,
including reference signals, may be multi-dimensional, at levels above, say, 2,
at any rate. Memory stores one (or maybe more) of these as a reference standard
for each of the higher level CSes that address it. The Input Function cobbles together
(infers) a similar multi-dimensional signal from the signals coming in from
below, and the Comparator does a pattern-recognition routine and determines
whether the reference signal and the inference signal are dissimilar enough to
require it to send an error signal to the Output Function.

If we were simulating this on a digital computer, the Memory
would be made up of addressable registers, each containing a string of 1s and
0s. When a register is addressed, its contents are copied to a register in the
Comparator, where it is compared cell by cell with a register into which the
Input Function’s inference has been copied. How this would be
accomplished in meatware I have no real clue, but I’m sure a billion
years of mutation and natural selection can have found a way.

I’ll try to deal with the rest of your helpful
message in my system diagram, coming up. I’ll need help in expressing
some things in it, I fear.

Best

Ted

[From Bill Powers (2010.08.23.0440 MDT)]

Ted Cloak (2010.08.22.1630 MDT)

TC: ... For reasons I can no longer access, I have come to use the expression "This control system controls perception P" to mean that P is the perception it is attempting to obtain and maintain; i.e., its perceptual goal; not that P is whatever perception is being handled by its Input Function.

BP: As I have imagined it, a perceptual input function constructs a perception by applying a particular computing process to a collection of perceptual signals from lower orders of perception. P[n] = Fi(P1[n-1] ... Pm[n-1]). The computing process determines the nature of the perception being generated, and the value of the function, a single magnitude, depends on the values of the contributing lower-order perceptions. This is quite analogous to the way the "area" of a rectangle is defined as a function of two linear measures, so that A = X * Y. The value of A can be indicated by a single signal, variable only in magnitude. The same is true of X and Y. But A is of a different logical type from X and Y, the difference being due to the intervening computation of multiplication.

TC: I'm not sure how different. Is your picture that every signal depicted in Figure 15.2 on page 218 of B:CP is a one-dimensional variable with no parameters except its magnitude?

BP: Yes.

TC: That is attractive because of its simplicity; single-parameter control systems, converging from lower levels to higher levels, add up to the incredibly complicated perceptions that every animal tries to control; and, going down, the incredibly fine-tuned movements through which we attempt to control them.

BP: Exactly as I imagine it.

TC: But then what's the use of each CS having a Memory, if all that a Memory can contain is a bunch of different numerical values, of which it can release, as a reference signal, only one at a time? Why won't Fig. 15.1, p. 209, suffice?

BP: It will probably suffice for lower-level systems, perhaps even up to the level of relationships. But remember that all neural signals are alike, being just magnitude indicators which can be represented by scalar numbers. The meaning of a signal is given by the computations that derive the numbers from lower-order signals. A complex memory is simply a set of numbers indicating the magnitudes that some set of signals had at a given time. When those numbers are played back in the way a digital recording of music is played back, they create higher-order perceptions according to the way higher-order input functions transform sets of incoming numbers to an outgoing number.

TC: My assumption has been, apparently, that address signals and error signals can be one-dimensional, but that perceptual signals, including reference signals, may be multi-dimensional, at levels above, say, 2, at any rate.

BP: That's not far from my idea given just above. Perceptual input functions are many-to-one functions. The signals themselves are one-dimensional, but each one is a function of many variables. The area of a rectangle is representable as a single number, but it is a function of two numbers, X and Y.

Don't press too hard here. As I keep noting in various posts, there is a huge problem here that I can't solve yet and may have to leave for future scientists to resolve (I hope they thank me for leaving them something to do). Why do all these numbers look to consciousness like the world we experience? We do not experience just one perceptual signal at a time. We experience a collection of them all in parallel. Though each of them is just a number, when we experience many of them at the same time, they look like a three-dimensional world in living color and surround sound with all the modalities of perception and all the levels of abstraction there are, all in one big pattern. Paul Churchland proposed a network theory of meaning, and that looks pretty much like what we have. No number in isolation means anything. Its meaning comes from its place in the pattern made by all the numbers.

In other words, it's magic.

TC: Memory stores one (or maybe more) of these as a reference standard for each of the higher level CSes that address it. The Input Function cobbles together (infers) a similar multi-dimensional signal from the signals coming in from below, and the Comparator does a pattern-recognition routine and determines whether the reference signal and the inference signal are dissimilar enough to require it to send an error signal to the Output Function.

BP: That's an attempt to combine the two main models of perception: the pattern model, in which one complex signal can represent many different perceptions depending on what "code" is being transmitted, and the "pandemonium" model, in which a zillion little demons, each recognizing just one variable, yell with a loudness proportional to how much of their variable is present in the multiple input signals that they all see. I use the pandemonium model (Selfridge). See p. 38 in the first edition of B:CP.

http://webspace.ship.edu/cgboer/pandemonium.html

TC: If we were simulating this on a digital computer, the Memory would be made up of addressable registers, each containing a string of 1s and 0s. When a register is addressed, its contents are copied to a register in the Comparator, where it is compared cell by cell with a register into which the Input Function's inference has been copied. How this would be accomplished in meatware I have no real clue, but I'm sure a billion years of mutation and natural selection can have found a way.

BP: Please, that's just another way of saying "I don't know." And it's the pattern model, not just scalar variables. I can't say you're wrong, but my model is far simpler and so far has worked pretty well. I sometimes wish I had never brought up the memory model, though if I hadn't, there would have been a chorus of questions: "what about memory?" This is all very conjectural and you may have noticed that none of my demos has any memory function in it.

We progress slowly. Keep up the good work.

Best,

Bill P.

[From Ted Cloak (2010.08.23.1200 MDT)]

[From Bill Powers (2010.08.23.0440 MDT)]

Ted Cloak (2010.08.22.1630 MDT)

TC: ... For reasons I can no longer access, I have come to use the
expression "This control system controls perception P" to mean that
P is the perception it is attempting to obtain and maintain; i.e.,
its perceptual goal; not that P is whatever perception is being
handled by its Input Function.

TC: OK, I propose that we use "This CS controls _for_ perception P" and
"This CS controls _input_ perception P", respectively, in non-technical
discourse. Does that work for you?

BP: As I have imagined it, a perceptual input function constructs a
perception by applying a particular computing process to a collection
of perceptual signals from lower orders of perception. P[n] =
Fi(P1[n-1] ... Pm[n-1]). The computing process determines the nature
of the perception being generated, and the value of the function, a
single magnitude, depends on the values of the contributing
lower-order perceptions. This is quite analogous to the way the
"area" of a rectangle is defined as a function of two linear
measures, so that A = X * Y. The value of A can be indicated by a
single signal, variable only in magnitude. The same is true of X and
Y. But A is of a different logical type from X and Y, the difference
being due to the intervening computation of multiplication.

TC: I'm not sure how different. Is your picture that every signal
depicted in Figure 15.2 on page 218 of B:CP is a one-dimensional
variable with no parameters except its magnitude?

BP: Yes.

TC: That is attractive because of its simplicity; single-parameter
control systems, converging from lower levels to higher levels, add
up to the incredibly complicated perceptions that every animal tries
to control; and, going down, the incredibly fine-tuned movements
through which we attempt to control them.

BP: Exactly as I imagine it.

TC: But then what's the use of each CS having a Memory, if all that
a Memory can contain is a bunch of different numerical values, of
which it can release, as a reference signal, only one at a time? Why
won't Fig. 15.1, p. 209, suffice?

BP: It will probably suffice for lower-level systems, perhaps even up
to the level of relationships. But remember that all neural signals
are alike, being just magnitude indicators which can be represented
by scalar numbers. The meaning of a signal is given by the
computations that derive the numbers from lower-order signals. A
complex memory is simply a set of numbers indicating the magnitudes
that some set of signals had at a given time.

BP: When those numbers are
played back in the way a digital recording of music is played back,
they create higher-order perceptions according to the way
higher-order input functions transform sets of incoming numbers to an
outgoing number.

TC: Aha!

TC: My assumption has been, apparently, that address signals and
error signals can be one-dimensional, but that perceptual signals,
including reference signals, may be multi-dimensional, at levels
above, say, 2, at any rate.

BP: That's not far from my idea given just above. Perceptual input
functions are many-to-one functions. The signals themselves are
one-dimensional, but each one is a function of many variables. The
area of a rectangle is representable as a single number, but it is a
function of two numbers, X and Y.

Don't press too hard here. As I keep noting in various posts, there
is a huge problem here that I can't solve yet and may have to leave
for future scientists to resolve (I hope they thank me for leaving
them something to do). Why do all these numbers look to consciousness
like the world we experience? We do not experience just one
perceptual signal at a time. We experience a collection of them all
in parallel. Though each of them is just a number, when we experience
many of them at the same time, they look like a three-dimensional
world in living color and surround sound with all the modalities of
perception and all the levels of abstraction there are, all in one
big pattern. Paul Churchland proposed a network theory of meaning,
and that looks pretty much like what we have. No number in isolation
means anything. Its meaning comes from its place in the pattern made
by all the numbers.

In other words, it's magic.

TC: Memory stores one (or maybe more) of these as a reference
standard for each of the higher level CSes that address it. The
Input Function cobbles together (infers) a similar multi-dimensional
signal from the signals coming in from below, and the Comparator
does a pattern-recognition routine and determines whether the
reference signal and the inference signal are dissimilar enough to
require it to send an error signal to the Output Function.

BP: That's an attempt to combine the two main models of perception:
the pattern model, in which one complex signal can represent many
different perceptions depending on what "code" is being transmitted,
and the "pandemonium" model, in which a zillion little demons, each
recognizing just one variable, yell with a loudness proportional to
how much of their variable is present in the multiple input signals
that they all see. I use the pandemonium model (Selfridge). See p. 38
in the first edition of B:CP.

Pandemonium

TC: That is outstanding.

TC: It's just occurred to me that when the contents of one register, e.g. in
Memory, is copied into another register, e.g., in the Comparator, it is
perforce done sequentially, one bit at a time. Maybe the distinction between
pattern and pandemonium doesn't make a difference serious enough to impede
our forward progress.

TC: If we were simulating this on a digital computer, the Memory
would be made up of addressable registers, each containing a string
of 1s and 0s. When a register is addressed, its contents are copied
to a register in the Comparator, where it is compared cell by cell
with a register into which the Input Function's inference has been
copied. How this would be accomplished in meatware I have no real
clue, but I'm sure a billion years of mutation and natural selection
can have found a way.

BP: Please, that's just another way of saying "I don't know."

TC: Isn't that just another way of saying "I have no real clue"? And don't
we all assume that Nature did it?

And
it's the pattern model, not just scalar variables. I can't say you're
wrong, but my model is far simpler and so far has worked pretty well.
I sometimes wish I had never brought up the memory model, though if I
hadn't, there would have been a chorus of questions: "what about
memory?" This is all very conjectural and you may have noticed that
none of my demos has any memory function in it.

TC: I think the reason we need a Memory function is not that it's
traditional, as you imply here, but that it keeps us modular. I think it's
better if the address signal is just "Do Something", and the addressed CS
has to know, or figure out, what to do -- rather than the address signal
telling the addressed CS what perception to control for. Does that make
sense to you?

We progress slowly. Keep up the good work.

TC: I'm on it!

Best
Ted

[From Bill Powers (2010.08.25.0943 MDT)]
Ted Cloak (2010.08.23.1200 MDT)–
Thanks for the diagram. It’s really on the same track I have been on –
what you call Radiant Variation is the same thing I call E. coli
reorganization. It’s a form of random walk optimization, I have
discovered – though not quite the same. A bunch of physicists got into
the act and called this sort of optimization “annealing,”
because something about the math resembled the math of annealing, I
guess. Physicists have this superstition that if two processes are
described by the same equation, they must have something in
common.
However, check out my thinking about this:
The error signal is the basis for closing one of the two relays and for
choosing another reference signal somewhat different from the previous
one. The error signal is based on the difference between the reference
signal and the perceptual signal.
As you have arranged the components of this reorganizing system, the
reference signal will be adjusted until the error approaches zero. This
means that the reference signal is being changed until it matches the
perceptual signal (which will also be changing). I don’t think that is
what we want.
The demonstrations in LCS3, particularly 7-2 and 8-1, use the error
signal as the criterion for reorganization as yours does, but what is
reorganized is the output weightings that connect the output function to
the reference inputs of a set of lower-order systems. The result
is to change how much of each lower-order perception reaches the
higher-order system’s perceptual input function. That results in making
the higher-order system’s perceptual signal a better and better
match to the varying higher-order reference signal as control
improves.

Memory can be introduced (though I haven’t done so) just by turning the
“weights” assigned to the signals going down to the lower-order
system into memory addresses. You almost did this – you said the higher
system above the one shown stops sending address signals when it is
getting the perception it wants, but this arrangement adapts the
reference signal to the perceptual signal instead of the other way
around. The reorganizing process should have been part of the output
function of the higher system, using the higher error signal as the basis
for the radiant variation.

With those changes you would have recreated the same process of
reorganization that I use.

Best,

Bill P.

···

[From Bill Powers
(2010.08.23.0440 MDT)]

Ted Cloak (2010.08.22.1630 MDT)

TC: … For reasons I can no longer access, I have come to use
the

expression “This control system controls perception P”
to mean that

P is the perception it is attempting to obtain and maintain;
i.e.,

its perceptual goal; not that P is whatever perception is
being

handled by its Input Function.

TC: OK, I propose that we use “This CS controls for perception
P” and

“This CS controls input perception P”, respectively, in
non-technical

discourse. Does that work for you?

BP: As I have imagined it, a perceptual input function constructs
a

perception by applying a particular computing process to a
collection

of perceptual signals from lower orders of perception. P[n] =

Fi(P1[n-1] … Pm[n-1]). The computing process determines the
nature

of the perception being generated, and the value of the function,
a

single magnitude, depends on the values of the contributing

lower-order perceptions. This is quite analogous to the way the

“area” of a rectangle is defined as a function of two
linear

measures, so that A = X * Y. The value of A can be indicated by
a

single signal, variable only in magnitude. The same is true of X
and

Y. But A is of a different logical type from X and Y, the
difference

being due to the intervening computation of multiplication.

TC: I’m not sure how different. Is your picture that every
signal

depicted in Figure 15.2 on page 218 of B:CP is a
one-dimensional

variable with no parameters except its magnitude?

BP: Yes.

TC: That is attractive because of its simplicity;
single-parameter

control systems, converging from lower levels to higher levels,
add

up to the incredibly complicated perceptions that every animal
tries

to control; and, going down, the incredibly fine-tuned
movements

through which we attempt to control them.

BP: Exactly as I imagine it.

TC: But then what’s the use of each CS having a Memory, if all
that

a Memory can contain is a bunch of different numerical values,
of

which it can release, as a reference signal, only one at a time?
Why

won’t Fig. 15.1, p. 209, suffice?

BP: It will probably suffice for lower-level systems, perhaps even
up

to the level of relationships. But remember that all neural
signals

are alike, being just magnitude indicators which can be
represented

by scalar numbers. The meaning of a signal is given by the

computations that derive the numbers from lower-order signals. A

complex memory is simply a set of numbers indicating the
magnitudes

that some set of signals had at a given time.

BP: When those numbers are

played back in the way a digital recording of music is played
back,

they create higher-order perceptions according to the way

higher-order input functions transform sets of incoming numbers to
an

outgoing number.

TC: Aha!

TC: My assumption has been, apparently, that address signals
and

error signals can be one-dimensional, but that perceptual
signals,

including reference signals, may be multi-dimensional, at
levels

above, say, 2, at any rate.

BP: That’s not far from my idea given just above. Perceptual
input

functions are many-to-one functions. The signals themselves are

one-dimensional, but each one is a function of many variables.
The

area of a rectangle is representable as a single number, but it is
a

function of two numbers, X and Y.

Don’t press too hard here. As I keep noting in various posts,
there

is a huge problem here that I can’t solve yet and may have to
leave

for future scientists to resolve (I hope they thank me for
leaving

them something to do). Why do all these numbers look to
consciousness

like the world we experience? We do not experience just one

perceptual signal at a time. We experience a collection of them
all

in parallel. Though each of them is just a number, when we
experience

many of them at the same time, they look like a
three-dimensional

world in living color and surround sound with all the modalities
of

perception and all the levels of abstraction there are, all in
one

big pattern. Paul Churchland proposed a network theory of
meaning,

and that looks pretty much like what we have. No number in
isolation

means anything. Its meaning comes from its place in the pattern
made

by all the numbers.

In other words, it’s magic.

TC: Memory stores one (or maybe more) of these as a
reference

standard for each of the higher level CSes that address it.
The

Input Function cobbles together (infers) a similar
multi-dimensional

signal from the signals coming in from below, and the
Comparator

does a pattern-recognition routine and determines whether
the

reference signal and the inference signal are dissimilar enough
to

require it to send an error signal to the Output Function.

BP: That’s an attempt to combine the two main models of
perception:

the pattern model, in which one complex signal can represent
many

different perceptions depending on what “code” is being
transmitted,

and the “pandemonium” model, in which a zillion little
demons, each

recognizing just one variable, yell with a loudness proportional
to

how much of their variable is present in the multiple input
signals

that they all see. I use the pandemonium model (Selfridge). See p.
38

in the first edition of B:CP.


http://webspace.ship.edu/cgboer/pandemonium.html

TC: That is outstanding.

TC: It’s just occurred to me that when the contents of one register, e.g.
in

Memory, is copied into another register, e.g., in the Comparator, it
is

perforce done sequentially, one bit at a time. Maybe the distinction
between

pattern and pandemonium doesn’t make a difference serious enough to
impede

our forward progress.

TC: If we were simulating this on a digital computer, the
Memory

would be made up of addressable registers, each containing a
string

of 1s and 0s. When a register is addressed, its contents are
copied

to a register in the Comparator, where it is compared cell by
cell

with a register into which the Input Function’s inference has
been

copied. How this would be accomplished in meatware I have no
real

clue, but I’m sure a billion years of mutation and natural
selection

can have found a way.

BP: Please, that’s just another way of saying “I don’t
know.”

TC: Isn’t that just another way of saying “I have no real
clue”? And don’t

we all assume that Nature did it?

And

it’s the pattern model, not just scalar variables. I can’t say
you’re

wrong, but my model is far simpler and so far has worked pretty
well.

I sometimes wish I had never brought up the memory model, though if
I

hadn’t, there would have been a chorus of questions: "what
about

memory?" This is all very conjectural and you may have noticed
that

none of my demos has any memory function in it.

TC: I think the reason we need a Memory function is not that it’s

traditional, as you imply here, but that it keeps us modular. I think
it’s

better if the address signal is just “Do Something”, and the
addressed CS

has to know, or figure out, what to do – rather than the address
signal

telling the addressed CS what perception to control for. Does that
make

sense to you?

We progress slowly. Keep up the good work.

TC: I’m on it!

Best

Ted

[From Ted Cloak (2010.08.25.1130
MTD)]

The error signal is
the basis for closing one of the two relays and for choosing another reference
signal somewhat different from the previous one.

Please look again. The error signal is the basis for opening
the NC (Normally Closed) relay, thus temporarily preventing the RVG
from changing the reference signal in Register r until or unless it “fails”;
i.e., shows that it can’t enable the input signal to the
higher-level CS to adequately approximate its reference signal.

Before I continue to reply, does this make a difference in your
comment?

Ted

[From Bill Powers (2010.08.25.1220 MDT)]

Ted Cloak (2010.08.25.1130 MTD) –

The error signal is the basis for closing one of the two relays and
for choosing another reference signal somewhat different from the
previous one.

Please look again. The error signal is the basis for opening the
NC (Normally Closed) relay, thus temporarily preventing the RVG
from changing the reference signal in Register r until or unless it
“fails”; i.e., shows that it can’t enable the input signal to the
higher-level CS to adequately approximate its reference
signal.

Before I continue to reply, does this make a difference in your
comment?

Yes and no. Your analysis is incomplete, because it only tangentially
mentions the error signal in the upper system not shown in the diagram.
It is that error signal, you are now saying, that sets the RVG into
action by closing the NO relay and starting random changes in
what we can now call the lower-order reference signal (it would not be
the address that closes the NO relay). The basis for starting the
changes is therefore the higher-order error signal, so in fact the NC
relay is superfluous. If there is reorganization due to the error signal
shown in the diagram, it will occur in the connection between that system
and lower-order systems off the bottom of the diagram. That will change
the way the “Address signals to lower-level control systems”
affect the perceptual signals entering the input function that is
shown.

In Manchester, Yu Li is developing a program to explore multi-level,
multi-system reorganization, for a graduate student working on his
doctoral degree. I think he will end up with something close to what you
are describing, if you accept my small modifications.

Best,

Bill P.

from Ted Cloak (2010.08.26.1700 MTD)

Let's adopt a convention about which Control System (CS) is which:
    CSA is the Parent CS, which addresses CSB.
    CSB is the Subject CS, the one shown in the System Diagram.
    CSC is the Child CS, which is (may be) addressed by CSB.

I've modified the diagram to use that convention, but not otherwise.

Bill Powers (2010.08.25.1220 MDT)

BP: ... Your analysis is incomplete, because it only tangentially mentions

the error signal in the upper system not shown in the diagram. It is that
error signal, you are now saying, that sets the RVG into action by _closing_
the NO relay and starting random changes in what we can now call the
lower-order reference signal (it would not be the address that closes the NO
relay).

TC: No, the address signal from CSA is, indeed, the trigger and sustainer
for the RVG's action, by closing the NO relay and keeping it closed. In the
first place, why should CSB be privy to what's going on in CSA's inner
workings? In the second place, for all we know, CSA's Output Function may
have elected not to address CSB despite receiving the error signal from
CSA's Comparator; it might address some other CS at CSB's level instead, in
which case we wouldn't want the reference signal currently in Register r of
CSB's Memory to be changed.

BP: The basis for starting the changes is therefore the higher-order error

signal, so in fact the NC relay is superfluous. If there is reorganization
due to the error signal shown in the diagram, it will occur in the
connection between that system and lower-order systems off the bottom of the
diagram. That will change the way the "Address signals to lower-level
control systems" affect the perceptual signals entering the input function
that is shown.

TC: The mechanism proposed in my diagram is not about how CSB learns to
control the perceptions coming into its Input Function. It is, rather, about
how CSB learns how to control the perceptions coming into CSA's Input
Function so that CSA will stop addressing CSB. More technically, it is about
how CSB continually adjusts the reference standard in its memory until, as
it attempts to control its perception to that standard, the perceptual
signals coming into _CSA's_ Input Function approximate CSA's reference
standard, so CSA ceases to send address signals to CSB. If and when that
happens, CSB's Memory stops sending reference signals to its Comparator and
its RVG shuts down, leaving in Register r a reference signal perhaps
somewhat more likely to succeed when CSB is next addressed by CSA. Thus CSB
learns how to satisfy CSA's input needs.

TC: The NC relay is superfluous only if the frequency with which the RVG
changes the contents of Register r is slower than the time it takes for the
entire hierarchy from CSB on down to operate and return a perceptual signal
to CSA. I assume that the RVG would go a lot faster than that, since much of
the time the new reference signal would approximate the perceptual signal
from the Input Function closely enough that no error signal would be going
out.

BP: In Manchester, Yu Li is developing a program to explore multi-level,

multi-system reorganization, for a graduate student working on his doctoral
degree. I think he will end up with something close to what you are
describing, if you accept my small modifications.

TC: Fascinating. Recursion will no doubt be required. I hope Yu Li is
reading this and will chime in.

TC: I wish other CSGNeters would chime in, too.

System Diagram second cut.pdf (47.4 KB)

[From Rick Marken (2010.08.26.1900)]

Ted Cloak (2010.08.26.1700 MTD)--

TC: I wish other CSGNeters would chime in, too.

Hi Ted. My chime is a question: What specific behavioral phenomenon is
this model designed to handle? More specifically, what kind of data
does the model account for?

I have used a very simply version of the basic reorganization model to
account to the behavior in a control task where the consequences of
action are random. You can see the data and behavior of the model at:

http://www.mindreadings.com/ControlDemo/Select.html

The "Subject" button collects data from you; the "Control" button runs
an E. coli reorganization model and the "Reinforcement" button runs a
reinforcement model. The "Control" model generally does better than
the "Reinforcement" model, in the sense that it, like the "Subject" is
able to get the dot to the target consistently.

I'd like to know specifically what kind of behavior is it that your
model is designed to handle. When I know that I'll be more interested
in trying to understand how your model works.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Erling Jorgensen (2000.08.27 1035 EDT)]

Ted Cloak (2010.08.26.1700 MTD)

TC: I wish other CSGNeters would chime in, too.

EJ: Allow me to briefly chime, to clarify a background issue as I, too, try
to understand your model better.

TC: The mechanism proposed in my diagram is not about how CSB learns to

control the perceptions coming into its Input Function. It is, rather, about
how CSB learns how to control the perceptions coming into CSA's Input
Function so that CSA will stop addressing CSB.

EJ: If CSB is guided by the CSA address to providing a perception that
the CSA Input Function needs, I'm not sure that CSA addressing stops once
the CSA perception is brought to sufficient control. To stop addressing
CSB would be to lose the contribution of the CSB perception, which would
seem to make control worse for CSA.

EJ: Once CSA control is adequately achieved, you would want to stop
_changing_ the address signal to CSB, essentially saying "That's the one;
you got it!" I think that's what might be achieved by having the CSA
_error signal_ decide how much drift in the address signaling is desired,
as Bill was proposing.

EJ: I appreciate your efforts to try to bring memory & references-as-
address-signals more explicitly into the PCT model.

All the best,
Erling

[From Bill Powers (2010.08.27.0835 MDT)]

Ted Cloak (2010.08.26.1700 MTD) --

TC: No, the address signal from CSA is, indeed, the trigger and sustainer
for the RVG's action, by closing the NO relay and keeping it closed.

BP: OK, so we have an address signal that not only can specify an address, but can cause a relay to close. What if it specifies an address of 00000000? In binary, that's as good an address as any other. Does that make the relay close?

There's something very odd going on in your modeling process that's going to take me a while to figure out, if I can do it at all.

TC: In the first place, why should CSB be privy to what's going on in CSA's inner workings?

BP: But it's not. It has no perception of signals coming into it from above, or of anything not represented in its own perceptual signal, which ooriginates in lower systems. What it "knows" is only what it perceives, and that is represented by the perceptual signal.

TC: In the second place, for all we know, CSA's Output Function may
have elected not to address CSB despite receiving the error signal from
CSA's Comparator; it might address some other CS at CSB's level instead, in
which case we wouldn't want the reference signal currently in Register r of
CSB's Memory to be changed.

BP: Well, either that reference signal would be zero, or some other system or set of systems at level A would be determining it. Reference signals don't just sit where you put them like writing on a blackboard. They're continuous variables reflecting the current outputs of higher systems that are contributing changing amounts to them.

I think you're confusing the address signal, which selects a storage location, with the value of the recorded perceptual signal that is stored in that location. The address signal just says "play back whatever is in this location." It can't also specify what particular waveforms are to be stored in that location.

TC: The mechanism proposed in my diagram is not about how CSB learns to
control the perceptions coming into its Input Function. It is, rather, about
how CSB learns how to control the perceptions coming into CSA's Input
Function so that CSA will stop addressing CSB.

But how does CSB even know that CSA exists? And why should a lower system be able to tell a higher one to stop addressing it? In PCT, it doesn't. It has control only over its own perceptions as they relate to the reference signal it receives (which it also does not perceive since it comes into the comparator, a simple subtractor, and not into its input function). CSB can't decide what to send on up to CSA and what not to send, or even know what effects it is having on CSA.

Let me put that another way. Of course it CAN do such things, but in order to do them it will need abilities that have to be put into the model. What will you have to put into CSB to enable it to do the things you describe? What you're describing goes far beyond anything in a PCT model, and I have no idea how we would go about simulating it to show that it would actually behave in the way you describe.

TC: More technically, it is about
how CSB continually adjusts the reference standard in its memory until, as
it attempts to control its perception to that standard, the perceptual
signals coming into _CSA's_ Input Function approximate CSA's reference
standard, so CSA ceases to send address signals to CSB.

BP: Now you have CSB adjusting its own reference standards. What part of CSB does that, and how does it act so the adjustment is what is needed to accomplish the result you describe? How does CSB know when CSA's perception is approaching its reference standards? How can CSA act if CSB is adjusting its own reference standards instead of CSA doing so? You've taken away CSA's only means of controlling its own perceptions, which is to adjust the reference standards for CSA.

TC: The NC relay is superfluous only if the frequency with which the RVG
changes the contents of Register r is slower than the time it takes for the
entire hierarchy from CSB on down to operate and return a perceptual signal
to CSA. I assume that the RVG would go a lot faster than that, since much of
the time the new reference signal would approximate the perceptual signal
from the Input Function closely enough that no error signal would be going
out.

You're making claims about how this new organization would behave, so now the burden is on you to show that it would actually behave this way. I'm kind of relieved to turn that over to you, since I have no idea how such things could be done, or whether they should be done. How do you know that the organization you describe would, if simulated as accurately as possible, generate the results you describe just above?

>BP: In Manchester, Yu Li is developing a program to explore multi-level,
multi-system reorganization, for a graduate student working on his doctoral
degree. I think he will end up with something close to what you are
describing, if you accept my small modifications.

TC: Fascinating. Recursion will no doubt be required. I hope Yu Li is
reading this and will chime in.

Recursion? Do you mean a function that calls itself as a subroutine, such as
sin(sin(sin ...(x))? Feedback loops are not recursions. There are no recursions in my PCT models.

Best,

Bill P.

from Ted Cloak (2010.08.27.1638 MTD)

I'll try to pull the three most recent replies into on response, to try to
keep the thread tight. First Rick, then Erling, then Bill.
Thanks to all of you for keeping up with this.

Rick Marken (2010.08.26.1900)

Ted Cloak (2010.08.26.1700 MTD)--

RM:Hi Ted. My chime is a question: What specific behavioral phenomenon is
this model designed to handle? More specifically, what kind of data
does the model account for?

TC:The "data", if I may call it that, is my observations of a kitten growing
into an adolescent cat. I've watched Sidi (short for Obsidian) practice
running, jumping, stalking, pouncing -- hour after hour, day after day. Each
day (well, week) I observed her getting more and more expert at those tasks.
I'm trying to explain this natural (as opposed to laboratory) learning which
is, of course, common to all vertebrates and probably all animal species.

TC:I do think we need to understand better how a control system (CS) learns
how to control its output to obtain/maintain the input demanded by its
reference signal, and I've got a suggestion for that, below. But being in a
hierarchy, a CS also needs to learn how to help the CS which addresses it
obtain/maintain the input demanded by that CS's reference signal. Otherwise
you don't have a hierarchy, actually, you just have a stack of CSes each
doing its own thing when addressed by the CS above it.

RM:I have used a very simply version of the basic reorganization model to
account to the behavior in a control task where the consequences of
action are random. You can see the data and behavior of the model at:

Selection of Consequences

The "Subject" button collects data from you; the "Control" button runs
an E. coli reorganization model and the "Reinforcement" button runs a
reinforcement model. The "Control" model generally does better than
the "Reinforcement" model, in the sense that it, like the "Subject" is
able to get the dot to the target consistently.

TC:That's an excellent demo of the superiority of PCT over "learning theory"
(as we used to call it). I've admired it before. But it doesn't show how a
CS works, nor need it. And it doesn't include learning, unless you count the
learning of the person working the space bar. E. coli, apparently, doesn't
learn.

RM:I'd like to know specifically what kind of behavior is it that your
model is designed to handle. When I know that I'll be more interested
in trying to understand how your model works.

TC: I hope I've piqued your interest.

···

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

Erling Jorgensen (2000.08.27 1035 EDT)

Ted Cloak (2010.08.26.1700 MTD)

TC: I wish other CSGNeters would chime in, too.

EJ: Allow me to briefly chime, to clarify a background issue as I, too, try

to understand your model better.

TC: The mechanism proposed in my diagram is not about how CSB learns to

control the perceptions coming into its Input Function. It is, rather,

about

how CSB learns how to control the perceptions coming into CSA's Input
Function so that CSA will stop addressing CSB.

EJ: If CSB is guided by the CSA address to providing a perception that
the CSA Input Function needs, I'm not sure that CSA addressing stops once
the CSA perception is brought to sufficient control.

TC:Sure it stops, because the error signal from CSA's Comparator to CSA's
Output Function stops, because CSA's perception approximates CSA's reference
signal, which is implied by your expression "brought to sufficient control".
Or do I misunderstand your question?

To stop addressing
CSB would be to lose the contribution of the CSB perception, which would
seem to make control worse for CSA.

TC:Excellent point. I started with a simple binary, on-off, idea of signals.
With that, once CSA's perception is brought to sufficient control, it stays
that way. That's not realistic, of course; indeed it's not PCT, and it's not
how Sidi's nervous system works, in all probability.

EJ: Once CSA control is adequately achieved, you would want to stop
_changing_ the address signal to CSB, essentially saying "That's the one;
you got it!" I think that's what might be achieved by having the CSA
_error signal_ decide how much drift in the address signaling is desired,
as Bill was proposing.

TC:Yes, having the error signal work through the address signal to manage
the RVG, perhaps by varying the frequency of changes in the reference
signal, might do the trick.

EJ: I appreciate your efforts to try to bring memory & references-as-
address-signals more explicitly into the PCT model.

TC:You're welcome. But references-as-address-signals? That's not my
intention.

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

Bill Powers (2010.08.27.0835 MDT)

Ted Cloak (2010.08.26.1700 MTD) --

TC: No, the address signal from CSA is, indeed, the trigger and sustainer
for the RVG's action, by closing the NO relay and keeping it closed.

BP: OK, so we have an address signal that not only can specify an
address, but can cause a relay to close. What if it specifies an
address of 00000000? In binary, that's as good an address as any
other. Does that make the relay close?

TC:Probably not. I now realize that the NO relay has to be a lot more
sophisticated, modifying the frequency RVG's actions according to the
intensity, or some other feature, of the incoming address signal. See above
discussion with Erling.

BP:There's something very odd going on in your modeling process that's
going to take me a while to figure out, if I can do it at all.

TC: In the first place, why should CSB be privy to what's going on
in CSA's inner workings?

BP: But it's not. It has no perception of signals coming into it from
above, or of anything not represented in its own perceptual signal,
which originates in lower systems. What it "knows" is only what it
perceives, and that is represented by the perceptual signal.

TC:Okay, you interpret "be privy to" as "have perception of". Let's try
this: I think it's better to have all communication between CSes in a
hierarchy be through as few interfaces as possible. Having a direct line
from the error signal of one CS to the Memory of another violates that
principle. Of course, who knows how evolution will have rigged such a thing?

TC: In the second place, for all we know, CSA's Output Function may
have elected not to address CSB despite receiving the error signal from
CSA's Comparator; it might address some other CS at CSB's level instead,

in

which case we wouldn't want the reference signal currently in Register r

of

CSB's Memory to be changed.

BP: Well, either that reference signal would be zero, or some other
system or set of systems at level A would be determining it.
Reference signals don't just sit where you put them like writing on a
blackboard. They're continuous variables reflecting the current
outputs of higher systems that are contributing changing amounts to them.

TC:I agree, providing you mean "address signal" rather than "reference
signal". You refer next to "the recorded perceptual signal that is stored in
that location". In my usage, a recorded perceptual signal that is stored in
Memory is a reference signal stored in a register.

BP:I think you're confusing the address signal, which selects a storage
location, with the value of the recorded perceptual signal that is
stored in that location.

TC: No, I'm not. Please look at my diagram.

BP:The address signal just says "play back
whatever is in this location." It can't also specify what particular
waveforms are to be stored in that location.

TC:No, but it can regulate random change in [one of] those waveforms via
regulating the RVG.

TC: The mechanism proposed in my diagram is not about how CSB learns to
control the perceptions coming into its Input Function. It is, rather,

about

how CSB learns how to control the perceptions coming into CSA's Input
Function so that CSA will stop addressing CSB.

BP:But how does CSB even know that CSA exists?

TC:It doesn't. It only "knows" (my usage, not yours, hence quotes) that its
Register r and its RVG are receiving a signal of a certain value.

BP:And why should a lower
system be able to tell a higher one to stop addressing it? In PCT, it
doesn't. It has control only over its own perceptions as they relate
to the reference signal it receives (which it also does not perceive
since it comes into the comparator, a simple subtractor, and not into
its input function). CSB can't decide what to send on up to CSA and
what not to send, or even know what effects it is having on CSA.

TC:I defy you to find in my diagram any indication that CSB directly
influences CSA, let alone controls CSA. The only way CSB influences CSA's
input perception (or anything else to do with CSA) is through the long, long
loop down the hierarchy, through the environment, and back up through a
series of Input Functions. My model is intended to show how CSB learns how
to do that better and faster.

Let me put that another way. Of course it CAN do such things, but in
order to do them it will need abilities that have to be put into the
model. What will you have to put into CSB to enable it to do the
things you describe? What you're describing goes far beyond anything
in a PCT model, and I have no idea how we would go about simulating
it to show that it would actually behave in the way you describe.

TC:I hope I answered that above.

TC: More technically, it is about
how CSB continually adjusts the reference standard in its memory until, as
it attempts to control its perception to that standard, the perceptual
signals coming into _CSA's_ Input Function approximate CSA's reference
standard, so CSA ceases to send address signals to CSB.

BP: Now you have CSB adjusting its own reference standards. What part
of CSB does that, and how does it act so the adjustment is what is
needed to accomplish the result you describe?

TC:That is done by the RVG.

BP:How does CSB know when
CSA's perception is approaching its reference standards?

TC:Now using "know" in my loose sense, it "knows" because the address signal
eases or stops. That completes the long, long loop.

BP:How can CSA
act if CSB is adjusting its own reference standards instead of CSA
doing so? You've taken away CSA's only means of controlling its own
perceptions, which is to adjust the reference standards for CSA.

TC:CSA adjusts CSB's reference standards by the address signal it sends,
which varies the behavior of CSB's RVG.

TC: The NC relay is superfluous only if the frequency with which the RVG
changes the contents of Register r is slower than the time it takes for

the

entire hierarchy from CSB on down to operate and return a perceptual

signal

to CSA. I assume that the RVG would go a lot faster than that, since much

of

the time the new reference signal would approximate the perceptual signal
from the Input Function closely enough that no error signal would be going
out.

BP:You're making claims about how this new organization would behave, so
now the burden is on you to show that it would actually behave this
way. I'm kind of relieved to turn that over to you, since I have no
idea how such things could be done, or whether they should be done.
How do you know that the organization you describe would, if
simulated as accurately as possible, generate the results you
describe just above?

TC:I'd appreciate it if you'd stick with this until you understand what
claims I am actually making. I admit, however, that I don't know that this
organizational scheme would really work, when instantiated in the head of a
kitten with millions of neurons in hundreds of thousands of control system
hierarchies all going at once. Would it cause a gradual convergence on more
and more expert behavior? I don't know, but I do know that there is such a
convergence, and I'm as certain as I can be that it is based on a process of
blind variation and selective retention.

>BP: In Manchester, Yu Li is developing a program to explore multi-level,
multi-system reorganization, for a graduate student working on his

doctoral

degree. I think he will end up with something close to what you are
describing, if you accept my small modifications.

TC: Fascinating. Recursion will no doubt be required. I hope Yu Li is
reading this and will chime in.

BP:Recursion? Do you mean a function that calls itself as a subroutine,

such as

sin(sin(sin ...(x))? Feedback loops are not recursions. There are no
recursions in my PCT models.

TC:OK, I was being a trifle facetious there. I didn't mean that hPCT works
by recursion, that would be absurd. I meant that a computer program
simulating hPCT might utilize recursion to avoid writing the same code over
and over for each level simulated. Perhaps iteration would do the trick.

TC:Next question: How does a CS learn to get the perceptions it needs to
approximate its reference signals in its Comparator? Does it matter, in a
hierarchy, whether it does get that approximation, unless it's the top CS in
the hierarchy? Would an RPG in the Output Function which variates the
outgoing address signals do the job?

Best to all
Ted

[From Bill Powers (2010.08.28.0740 MDT)]

Ted Cloak (2010.08.27.1638 MTD) --

TC:The "data", if I may call it that, is my observations of a kitten growing
into an adolescent cat. I've watched Sidi (short for Obsidian) practice
running, jumping, stalking, pouncing -- hour after hour, day after day. Each
day (well, week) I observed her getting more and more expert at those tasks.
I'm trying to explain this natural (as opposed to laboratory) learning which
is, of course, common to all vertebrates and probably all animal species.

BP: So am I. Reorganization is my proposed method of learning, which I think may be the natural way common to all animals that learn.

However, I'm finding it difficult to communicate just how reorganization works, so this conversation is turning out to be useful. Evidently you as well as others already have a concept of learning in mind, and you're assuming that the reorganization approach works the same way./

···

TC:I do think we need to understand better how a control system (CS) learns
how to control its output to obtain/maintain the input demanded by its
reference signal, and I've got a suggestion for that, below. But being in a
hierarchy, a CS also needs to learn how to help the CS which addresses it
obtain/maintain the input demanded by that CS's reference signal. Otherwise
you don't have a hierarchy, actually, you just have a stack of CSes each
doing its own thing when addressed by the CS above it.
>
>RM:I have used a very simply version of the basic reorganization model to
>account to the behavior in a control task where the consequences of
>action are random. You can see the data and behavior of the model at:
>
> Selection of Consequences
>
>The "Subject" button collects data from you; the "Control" button runs
>an E. coli reorganization model and the "Reinforcement" button runs a
>reinforcement model. The "Control" model generally does better than
>the "Reinforcement" model, in the sense that it, like the "Subject" is
>able to get the dot to the target consistently.

TC:That's an excellent demo of the superiority of PCT over "learning theory"
(as we used to call it). I've admired it before. But it doesn't show how a
CS works, nor need it. And it doesn't include learning, unless you count the
learning of the person working the space bar. E. coli, apparently, doesn't
learn.
>
>RM:I'd like to know specifically what kind of behavior is it that your
>model is designed to handle. When I know that I'll be more interested
>in trying to understand how your model works.
>
TC: I hope I've piqued your interest.

=================================================
>
>Erling Jorgensen (2000.08.27 1035 EDT)
>
>>Ted Cloak (2010.08.26.1700 MTD)
>
>>TC: I wish other CSGNeters would chime in, too.
>
>EJ: Allow me to briefly chime, to clarify a background issue as I, too, try

>to understand your model better.
>
>>TC: The mechanism proposed in my diagram is not about how CSB learns to
>control the perceptions coming into its Input Function. It is, rather,
about
>how CSB learns how to control the perceptions coming into CSA's Input
>Function so that CSA will stop addressing CSB.
>
>EJ: If CSB is guided by the CSA address to providing a perception that
>the CSA Input Function needs, I'm not sure that CSA addressing stops once
>the CSA perception is brought to sufficient control.

TC:Sure it stops, because the error signal from CSA's Comparator to CSA's
Output Function stops, because CSA's perception approximates CSA's reference
signal, which is implied by your expression "brought to sufficient control".
Or do I misunderstand your question?

>To stop addressing
>CSB would be to lose the contribution of the CSB perception, which would
>seem to make control worse for CSA.

TC:Excellent point. I started with a simple binary, on-off, idea of signals.
With that, once CSA's perception is brought to sufficient control, it stays
that way. That's not realistic, of course; indeed it's not PCT, and it's not
how Sidi's nervous system works, in all probability.
>
>EJ: Once CSA control is adequately achieved, you would want to stop
>_changing_ the address signal to CSB, essentially saying "That's the one;
>you got it!" I think that's what might be achieved by having the CSA
>_error signal_ decide how much drift in the address signaling is desired,
>as Bill was proposing.

TC:Yes, having the error signal work through the address signal to manage
the RVG, perhaps by varying the frequency of changes in the reference
signal, might do the trick.
>
>EJ: I appreciate your efforts to try to bring memory & references-as-
>address-signals more explicitly into the PCT model.
>
TC:You're welcome. But references-as-address-signals? That's not my
intention.

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

>Bill Powers (2010.08.27.0835 MDT)
>
>Ted Cloak (2010.08.26.1700 MTD) --
>
>>TC: No, the address signal from CSA is, indeed, the trigger and sustainer
>>for the RVG's action, by closing the NO relay and keeping it closed.
>
>BP: OK, so we have an address signal that not only can specify an
>address, but can cause a relay to close. What if it specifies an
>address of 00000000? In binary, that's as good an address as any
>other. Does that make the relay close?

TC:Probably not. I now realize that the NO relay has to be a lot more
sophisticated, modifying the frequency RVG's actions according to the
intensity, or some other feature, of the incoming address signal. See above
discussion with Erling.
>
>BP:There's something very odd going on in your modeling process that's
>going to take me a while to figure out, if I can do it at all.
>
>>TC: In the first place, why should CSB be privy to what's going on
>>in CSA's inner workings?
>
>BP: But it's not. It has no perception of signals coming into it from
>above, or of anything not represented in its own perceptual signal,
>which originates in lower systems. What it "knows" is only what it
>perceives, and that is represented by the perceptual signal.

TC:Okay, you interpret "be privy to" as "have perception of". Let's try
this: I think it's better to have all communication between CSes in a
hierarchy be through as few interfaces as possible. Having a direct line
from the error signal of one CS to the Memory of another violates that
principle. Of course, who knows how evolution will have rigged such a thing?
>
>>TC: In the second place, for all we know, CSA's Output Function may
>>have elected not to address CSB despite receiving the error signal from
>>CSA's Comparator; it might address some other CS at CSB's level instead,
in
>>which case we wouldn't want the reference signal currently in Register r
of
>>CSB's Memory to be changed.
>
>BP: Well, either that reference signal would be zero, or some other
>system or set of systems at level A would be determining it.
>Reference signals don't just sit where you put them like writing on a
>blackboard. They're continuous variables reflecting the current
>outputs of higher systems that are contributing changing amounts to them.

TC:I agree, providing you mean "address signal" rather than "reference
signal". You refer next to "the recorded perceptual signal that is stored in
that location". In my usage, a recorded perceptual signal that is stored in
Memory is a reference signal stored in a register.
>
>BP:I think you're confusing the address signal, which selects a storage
>location, with the value of the recorded perceptual signal that is
>stored in that location.

TC: No, I'm not. Please look at my diagram.

>BP:The address signal just says "play back
>whatever is in this location." It can't also specify what particular
>waveforms are to be stored in that location.

TC:No, but it can regulate random change in [one of] those waveforms via
regulating the RVG.
>
>>TC: The mechanism proposed in my diagram is not about how CSB learns to
>>control the perceptions coming into its Input Function. It is, rather,
about
>>how CSB learns how to control the perceptions coming into CSA's Input
>>Function so that CSA will stop addressing CSB.
>
>BP:But how does CSB even know that CSA exists?

TC:It doesn't. It only "knows" (my usage, not yours, hence quotes) that its
Register r and its RVG are receiving a signal of a certain value.

>BP:And why should a lower
>system be able to tell a higher one to stop addressing it? In PCT, it
>doesn't. It has control only over its own perceptions as they relate
>to the reference signal it receives (which it also does not perceive
>since it comes into the comparator, a simple subtractor, and not into
>its input function). CSB can't decide what to send on up to CSA and
>what not to send, or even know what effects it is having on CSA.

TC:I defy you to find in my diagram any indication that CSB directly
influences CSA, let alone controls CSA. The only way CSB influences CSA's
input perception (or anything else to do with CSA) is through the long, long
loop down the hierarchy, through the environment, and back up through a
series of Input Functions. My model is intended to show how CSB learns how
to do that better and faster.
>
>Let me put that another way. Of course it CAN do such things, but in
>order to do them it will need abilities that have to be put into the
>model. What will you have to put into CSB to enable it to do the
>things you describe? What you're describing goes far beyond anything
>in a PCT model, and I have no idea how we would go about simulating
>it to show that it would actually behave in the way you describe.

TC:I hope I answered that above.
>
>>TC: More technically, it is about
>>how CSB continually adjusts the reference standard in its memory until, as
>>it attempts to control its perception to that standard, the perceptual
>>signals coming into _CSA's_ Input Function approximate CSA's reference
>>standard, so CSA ceases to send address signals to CSB.
>
>BP: Now you have CSB adjusting its own reference standards. What part
>of CSB does that, and how does it act so the adjustment is what is
>needed to accomplish the result you describe?

TC:That is done by the RVG.

>BP:How does CSB know when
>CSA's perception is approaching its reference standards?

TC:Now using "know" in my loose sense, it "knows" because the address signal
eases or stops. That completes the long, long loop.

>BP:How can CSA
>act if CSB is adjusting its own reference standards instead of CSA
>doing so? You've taken away CSA's only means of controlling its own
>perceptions, which is to adjust the reference standards for CSA.

TC:CSA adjusts CSB's reference standards by the address signal it sends,
which varies the behavior of CSB's RVG.
>
>>TC: The NC relay is superfluous only if the frequency with which the RVG
>>changes the contents of Register r is slower than the time it takes for
the
>>entire hierarchy from CSB on down to operate and return a perceptual
signal
>>to CSA. I assume that the RVG would go a lot faster than that, since much
of
>>the time the new reference signal would approximate the perceptual signal
>>from the Input Function closely enough that no error signal would be going
>>out.
>
>BP:You're making claims about how this new organization would behave, so
>now the burden is on you to show that it would actually behave this
>way. I'm kind of relieved to turn that over to you, since I have no
>idea how such things could be done, or whether they should be done.
>How do you know that the organization you describe would, if
>simulated as accurately as possible, generate the results you
>describe just above?

TC:I'd appreciate it if you'd stick with this until you understand what
claims I am actually making. I admit, however, that I don't know that this
organizational scheme would really work, when instantiated in the head of a
kitten with millions of neurons in hundreds of thousands of control system
hierarchies all going at once. Would it cause a gradual convergence on more
and more expert behavior? I don't know, but I do know that there is such a
convergence, and I'm as certain as I can be that it is based on a process of
blind variation and selective retention.
>
>> >BP: In Manchester, Yu Li is developing a program to explore multi-level,
>>multi-system reorganization, for a graduate student working on his
doctoral
>>degree. I think he will end up with something close to what you are
>>describing, if you accept my small modifications.
>>
>>TC: Fascinating. Recursion will no doubt be required. I hope Yu Li is
>>reading this and will chime in.
>
>BP:Recursion? Do you mean a function that calls itself as a subroutine,
such as
>sin(sin(sin ...(x))? Feedback loops are not recursions. There are no
>recursions in my PCT models.
>
TC:OK, I was being a trifle facetious there. I didn't mean that hPCT works
by recursion, that would be absurd. I meant that a computer program
simulating hPCT might utilize recursion to avoid writing the same code over
and over for each level simulated. Perhaps iteration would do the trick.

TC:Next question: How does a CS learn to get the perceptions it needs to
approximate its reference signals in its Comparator? Does it matter, in a
hierarchy, whether it does get that approximation, unless it's the top CS in
the hierarchy? Would an RPG in the Output Function which variates the
outgoing address signals do the job?

Best to all
Ted

[From Bill Powers (2010.08.28.0740 MDT)]

Ted Cloak (2010.08.27.1638 MTD) –

TC:The “data”, if I
may call it that, is my observations of a kitten growing

into an adolescent cat. I’ve watched Sidi (short for Obsidian)
practice

running, jumping, stalking, pouncing – hour after hour, day after day.
Each

day (well, week) I observed her getting more and more expert at those
tasks.

I’m trying to explain this natural (as opposed to laboratory) learning
which

is, of course, common to all vertebrates and probably all animal
species.

BP: So am I. Reorganization is my proposed basic method of learning,
which I think may be the natural way common to all animals that
learn.
However, I’m finding it difficult to communicate just how PCT
reorganization works, so this conversation is turning out to be useful.
If I keep working on making it clearer, perhaps eventually it will become
comprehensible. Evidently you as well as others already have a concept of
learning in mind, and you’re assuming that the reorganization approach
works the same way. But it doesn’t.
The main fact is that PCT reorganization does not work by trying out
different reference signals or different behaviors. The best way to see
it in action (other than watching cats grow up) is to look at the demos
in LCS3.
Demo 7-2 shows a simple system with just three controllers and three
environmental variables. It starts out with all three control systems
sensing variables constructed in three different ways from all three
environmental variables. The input weights of the three systems, nine in
all, are set at random at the start of a run and remain the same from
then on; adding input reorganization is a project for the future. The
output functions of the three systems are each connected to all three
environmental variables through adjustable weights, which are initially
set to zero. In this demo, the three reference signals vary in a
repeating pattern of magnitudes that traces out a Lissajjous pattern in
three dimensions. This pattern of reference signals *never changes.*But control is so poor at first that the controlled perceptions
change very differently from the way the reference signals
change.
The only variables being altered by reorganization are the three output
weights in each of the three systems. For each system, there is (in
addition to the three weights) a set of three auxiliary numbers that are
set at random to values between -1 and 1. These “speed” numbers
are added (multiplied by a very small constant like 1/100,000) to the
output weights on every iteration of the program. The output weights
therefore begin to change at a constant speed, the speeds being
proportional to the “speed” numbers. This corresponds to the
“swimming” phase of E. coli’s behavior.
The three control systems start life in a very crude form. The loop gain
is zero because all the output weights are zero, Soon after the start,
all the output weights are nonzero, so each of the three control systems
is trying to affect all three environmental variables. The weights,
however are very small so the amount of action generated is small at
first, and the weights are not adjusted appropriately, so the control is
very bad. The control systems interfere with each other because each one
affects not only the variable it is supposed to be controlling, but
variables in the other two system as well, and it may not be affecting
its own variable enough or in the right direction.
This beginning situation is similar to what is found by neurologists
looking at motor systems in neonates. There is a crude general
input-output arrangement connecting senses to muscles, so the control
systems are sort of sketched in, thanks to evolution. But there are far
more connections from sensory to motor nuclei than are needed, and most
of them are not the right connections. As maturation and practice
proceed, the number of connections is gradually reduced, a process they
call “pruning.” In the end, the wrong connections and
superfluous connections are pruned away (I would say, the weights are
reduced close to zero), leaving only the right connections. And what
doesn’t show in the crude observations possible in living brains is that
the remaining weights continue to change so the control systems become
more stable and more skillful at keeping errors very small.
So at first, we have the weights changing at some rate, a different rate
for each of the three output weights. If the running average of squared
error of a control system is decreasing as a result, because control is
getting better even though still not very good, that change continues.
Eventually, however, the three weights will be as close to the right
amounts as possible, and then the changes will start making the error
larger. As soon as that happens, there is a tumble. A reorganizing
control system sitting off to one side senses the increase in absolute
error, and produces an output that changes the auxiliary
“speed” variables to new values between -1 and 1 – at
random
. This starts the output weights changing, iteration by
iteration of the program, in different proportions, so if we plotted the
weights in three dimensions, the three-dimensional direction in which the
resultant is moving would be different. That change of direction is a
tumble.

There is just as much chance that this random change will leave the error
still increasing or increasing faster as making it start to decrease, in
which case another tumble will occur immediately. If the tumbles come
close together, the weights will not change by very much. Eventually, a
tumble will set the weights to changing in a direction that makes the
error decrease, and the tumbles will cease. The weights will go on
changing in the new proportions as long as the error keeps getting
smaller.

Clearly, this principle should make all the weights approach the values
they must have for the error in each control system to get as small as it
can get. I have found that multiplying the effects of the speed variable
by a number proportional to the absolute amount of error produces more
efficient convergence, making the speed of change approach zero as the
error approaches zero.

It must be understood that all during this process, the reference signals
for the three systems in the demo are varying in a fixed pattern that
never changes. In other demos, the change in reference signal is made
random, and random disturbances are added, and the reorganizing process
still converges to the best control possible to achieve by slowly
adjusting output weights. The actual behavior patterns, in that case, may
never repeat, and the reference values may never repeat their patterns,
either, yet control will continue to improve over time. I have shown all
these effects in various demonstrations, though only the simplest of them
are in LCS3.

Demo 8-1 shows this same principle with random disturbances and with the
reference signals varying in a fixed pattern, for an arm with 14 degrees
of freedom. You can see the movements becoming more regular as time goes
on, starting with clumsy flailings and ending with a smooth regular Tai
Chi exercise pattern. The Tai Chi reference pattern remains exactly the
same from beginning to end of the demo, but the ability to control the
arm to match it while resisting the effects of disturbances continually
improves.

TC:I do think we need to
understand better how a control system (CS) learns

how to control its output to obtain/maintain the input demanded by
its

reference signal, and I’ve got a suggestion for that,
below.

BP: I hope you can see now that reorganization theory provides the
explanation you’re looking for, and the demos do exactly what you
describe.

TC: But being in a

hierarchy, a CS also needs to learn how to help the CS which addresses
it

obtain/maintain the input demanded by that CS’s reference
signal.

BP: I disagree. All that a control system in the hierarchy has to learn
to do is adjust its outputs to lower systems in a way that keeps its own
perceptions matching whatever reference signal it is given, as quickly
and accurately as possible. When that is achieved, a still-higher control
system can use the lower one as its means of controlling its own,
higher-order, perception: adjusting the reference signal in the lower
system will quickly and accurately make the perceptual signal in that
lower system follow the changing values of the reference signal, and thus
provide an input (among many) to the higher perceptual input function
that is needed to produce the desired amount of the higher-order
perception.

I’ll leave it at that for now.

Best,

Bill P.

[Martin Taylor 2010.08.28.13.23]

> From Ted Cloak (2010.08.26.1700 MTD)

Let's adopt a convention about which Control System (CS) is which:
     CSA is the Parent CS, which addresses CSB.
     CSB is the Subject CS, the one shown in the System Diagram.
     CSC is the Child CS, which is (may be) addressed by CSB.

I've modified the diagram to use that convention, but not otherwise.

TC: I wish other CSGNeters would chime in, too.

As you wish...

In your diagram, you show the CSB reference signal coming from CSA and the CSB perceptual signal going back to the same CSA, as though the hierarchy were a ladder, even though you show many outputs to CSCs and inputs from several CSCs at lower levels. While it is true that sometimes the perceptual signal controlled by a control unit will be returned to a higher level unit that contributes to its reference input, that is not always the case.

Consider the classic example of keeping your car in its lane. The inputs to the control unit that senses where the car is in the lane are all visual, but the output that matters provides a reference value to the unit that controls steering wheel angle. The perception of steering wheel angle does not contribute in any way to the perception of where the car is in its lane. Changing the location of the car in its lane is a pure side-effect of changing the steering wheel angle, since only if the car is moving and there happens to be a functional steering linkage between steering wheel angle and road wheel angle does altering the steering wheel angle influence where the car is in its lane, even though control of the steering wheel angle may be perfect while the car is stationary and the steering linkage is cut.

As another example, consider the standard pursuit tracking experiment. The controlled perception is the visual perception of the separation between target and cursor. The output from the unit that controls this perception goes to some set of systems that control joystick movement (either position or velocity -- it doesn't matter which to make the point). Neither the position nor the velocity of the joystick enter into the perception of the separation of the cursor and the target. The effect of the joystick on the cursor is a pure side-effect of control of the joystick position (or velocity).

One could easily describe hundreds of such cases, in which the output of a CSA provides a reference value to some CSB, but the perception controlled by that CSB does not contribute to the perception controlled by the CSA. In those cases, CSB control has as a side-effect the ability of CSA to control. I have no idea whether this is more often true than not, but does the fact that it is sometimes true affect the design of your model?

As a gratuitous aside, I recommend you study carefully Bill's description of e-coli reorganization. I know you were trying to provide a workable model for Bill's insight that the reference value may well not be the signal output by CSA, but instead may be the output of a content-addressable memory addressed by the CSA output, while Bill's e-coli reorganization demos (so far as I know) do not include this memory function. Would your scheme work within e-coli reorganization as Bill described it? For example, if the initial contents of the content-addressable memory were random, with more addresses than were likely to be needed, and the contents were updated with higher probability when reorganization is proceeding slowly, while unused addresses were removed over time from the list (the neonate pruning noted by Bill), would that work?

Just a few questions. Maybe they are off-base, but I don't know the answers to them. And, as an aside to Bill, if I ever get around to modelling them, it won't be soon. I do have someone interested in doing some PCT programming with me, but not until after the New Year.

Martin