sequence control and curvaceous errors

CSGnet Archive CSG2002
1

[From Bruce Nevin (2002.05.07 21:10 EDT)]
We’re asking why no action in the face of disturbance to a controlled
variable. Is the gain lowered? Is there a universal function in all
output functions that replicates the physical properties of stretched
muscle fibers? What else might result in a cessation of action to control
a disturbed CV? If I’m controlling in imagination a disturbance in the
environment does not affect control. What else?
The latch-trip mechanism proposed in B:CP (diagrams on pp. 142, 144) is
described in terms of event recognition, e.g. /p/ followed by /i/
followed by /n/ results in recognition of the word-event /pin/. The input
function for /i/ in this sequence requires a signal from the /p/
recognizer in order to pass a signal along to the input of the /n/
recognizer. (A summing element would give a signal for /pin/ with less
than full strength if some part of the sequence were missing. Subjective
experience suggests that missing elements are imagined.)
Presumably a like mechanism controls production of an event or sequence.
Sequences that are not short and stereotyped, however well learned they
may be, are not events at level 5 in the (current) standard proposal
about the perceptual hierarchy. They are thought to have category and
relationship perceptions in their inputs, so they are above the
demarcation between analog perceptions and digital (binary, yes/no)
perceptions. (But note in passing that the elements in an event-sequence
are categorial – either the element is perceived and the latch tripped,
or not – and perhaps it is on/off character of the latch mechanism
itself that imposes a categorization.) Such sequences may be thought of
as the simplest form of programs. Among other differences, sequences are
easily interrupted by other sequences and then resumed, but events that
are interrupted are typically started over (performance) or the missing
parts are imagined to complete it (recognition).
Interruption and resumption of performing a sequence seems to involve
some kind of addressing to return to the incomplete sequence. To return
to the right place, something must maintain the perceptual signal that is
the third. In the proposed latch-trip mechanism, it is the
“recirculation loop” that does this. In the word recognition
example in B:CP there is a loop back from event completion to inhibit the
signal. There would also have to be an inhibitory input going from the
tripped /pin/ recognizer to the recognizer for /pit/, /pil/, /pip/ to
prevent false recognition of an ensuing /t/ etc. that belongs to the next
word. In the sequence control system, absent any inhibitory signal, the
recirculation loop for the second step keeps alive a perception that step
2 has been completed and – just how is not quite clear – a perception
that step 3 has not been completed.
What is not clear is how, having successfully controlled step 2, we now
start controlling step 3. Either the reference for step 3 is set,
and it was not previously, or the gain for controlling step 3 is raised,
or both.

At this point, the sequence is interrupted. We have to go to the
bathroom. The telephone rings. There is a power outage. We want some ice
cream. For whatever reason, we put the sequence on a memory shelf, we
remember that step 2 is completed, and we remember that we must control
step 3 to continue. As described above, we are controlling a variable
(completion of step 3) in the face of disturbance (step 3 is not even
started), and we are not taking any action to resist the
disturbance.

There is some evidence that we are controlling resumption of the
interrupted sequence, and that whatever we are controlling during the
interruption is the means for controlling resumption of the interrupted
sequence. I might try to complete the telephone interruption as quickly
as I can so I can resume what I was doing (impatience). I might take my
time over the ice cream because step 3 is difficult or distasteful or in
some other way occasions conflict, and indeed that might be why the
desire for ice cream came up (procrastination).

The speed at which I control during the interruption (which is a
controllable variable, not merely gain, cp. the enjoyment of ice cream)
is a means of controlling the length of the interruption and the
quickness with which I return to controlling step 3 of the sequence.

Of course this is true of any of the steps in the sequence: the faster I
perform step 2, the more quickly I can start step 3.

Bill Curry (2002.05.07.1030 EDT) suggests that imagination is involved in
the horse bolting for the barn to get fed. The ride is part of the
sequence go out - ride - come back - get fed. ‘Imagination’ colloquially
includes more than is modelled by the imagination loop in PCT. The
controlling of something more quickly (of slowly) as means of starting to
control the next element of a sequence more quickly (or slowly) could be
thought of as due to ‘imagining’ the next element and either avoiding it
or (like the horse) hastening toward it.

    /Bruce

Nevin

···

sent from the output of the second step (say) to the perceptual input of

2

[From Rick Marken (2002.05.08.0800)]

Bruce Nevin (2002.05.07 21:10 EDT)--

We're asking why no action in the face of disturbance to a controlled
variable.

I think we're asking whether an increase in error ever leads to no
change or even a _decrease_ in action. So the first order of business is
to demonstrate this phenomenon (which I would call "control range
limitation" rather than "giving up") experimentally. Ee have had
several observations described that _seem to_ involve control range
limitation: the knife in the water, the tethering experiment, the
barn-sour horse, etc . Bill Powers (2002.05.06.1621 MDT) described what
is needed now:

All that's needed is to apply a disturbance, as usual. If the effect

of a

disturbance acting _toward_ a goal-state is to _increase_ the effort
toward that state, we are looking at an example of the UEC

Once we've got this we can start looking at (and testing) alternative
explanations.

Best regards

Rick

···

--
Richard S. Marken, Ph.D.
The RAND Corporation
PO Box 2138
1700 Main Street
Santa Monica, CA 90407-2138
Tel: 310-393-0411 x7971
Fax: 310-451-7018
E-mail: rmarken@rand.org

3

[From Bill Powers (2002.05.08.0807 MDT)]

Bruce Nevin (2002.05.07 21:10 EDT)--

We're asking why no action in the face of disturbance to a controlled
variable.

Let's convert to the more general question: why does an increase in error
cause a decrease in effort, and the opposite (first, of course, asking if
this is ever observed).

Is the gain lowered? Is there a universal function in all output functions
that replicates the physical properties of stretched muscle fibers?

Well, that wouldn't be the reason for its existence, if it exists. But
before we ask about the U in UEC, let's ask whether this form of
error-to-output curve (or its effect) is _ever_ observed. If it's not ever
observed we can drop the subject. If it is at least sometimes observed, we
can ask how often and under what circumstances.

What else might result in a cessation of action to control a disturbed
CV? If I'm controlling in imagination a disturbance in the environment
does not affect control. What else?

There are too many other things that can cause action to cease. I think it
would be better to focus on a reversal of the normal error-to-output
relationship,. rather than extremes that involve either-or conditions. My
inclination is to look at the simplest questions first.

>Sequences that are not short and stereotyped, however well learned they
may be, are not events at >level 5 in the (current) standard proposal about
the perceptual hierarchy. They are thought to have >category and
relationship perceptions in their inputs, so they are above the demarcation
between >analog perceptions and digital (binary, yes/no) perceptions. (But
note in passing that the elements >in an event-sequence are categorial --
either the element is perceived and the latch tripped, or not -- >and
perhaps it is on/off character of the latch mechanism itself that imposes a
categorization.)

I'm not very happy about the state of Level 5. It's too much like the
sequence level. My reasons for wanting events to be at that level are
getting pretty ragged, if not fading from memory entirely. I would support
a principled effort to justify doing away with this level altogether and
letting its supposed functions be handled by the current sequence level.
Why shouldn't the next level up from transitions be relationships, still an
analog type of variable? Then the process of categorization would provide a
clean demarkation between continuous and discrete control processes. For
some reason that sounds nice to me.

Interruption and resumption of performing a sequence seems to involve some
kind of addressing to return to the incomplete sequence. To return to the
right place, something must maintain the perceptual signal that is sent
from the output of the second step (say) to the perceptual input of the
third. In the proposed latch-trip mechanism, it is the "recirculation
loop" that does this. In the word recognition example in B:CP there is a
loop back from event completion to inhibit the signal. There would also
have to be an inhibitory input going from the tripped /pin/ recognizer to
the recognizer for /pit/, /pil/, /pip/ to prevent false recognition of an
ensuing /t/ etc. that belongs to the next word.

Actually, when the final input occurs, it resets the first element, and
loss of that signal disables the second element, and so on to the last
element. So the final /t/ would turn off the recognition signal and all
signals including itself in the chain. Don't know whether that's good or
bad, but that's what this circuit would do.

In the sequence control system, absent any inhibitory signal, the
recirculation loop for the second step keeps alive a perception that step
2 has been completed and -- just how is not quite clear -- a perception
that step 3 has not been completed.

It enables the third-stage latch so it can be fired by the third input. No
element can latch or stay latched unless the previous latch is on. But
there is no specific perception that the perception has +not+ been
completed. There is simply no final recognition signal. Perhaps whatever is
waiting for that signal might conclude that it has not yet occurred.

What is not clear is how, having successfully controlled step 2, we now
start controlling step 3. Either the reference for step 3 is set, and it
was not previously, or the gain for controlling step 3 is raised, or both.

The steps are not individually controlled. This was a model of a perceptual
input function, not a whole control system. This PIF receives connections
representing three specific phonemes (from a lower phoneme-recognizer
level) and puts out a perceptual signal (briefly) if and only if signals
occur in the three input lines in the right sequence. The circuit would
have to be a little more complicated to make it reset if any input occurred
in the _wrong_ sequence -- though I'm not sure that's what should happen.

At this point, the sequence is interrupted. We have to go to the bathroom.
The telephone rings. There is a power outage. We want some ice cream. For
whatever reason, we put the sequence on a memory shelf, we remember that
step 2 is completed, and we remember that we must control step 3 to
continue. As described above, we are controlling a variable (completion of
step 3) in the face of disturbance (step 3 is not even started), and we
are not taking any action to resist the disturbance.

My tendency, and what I think I observe in others, is to start interrupted
sequences over again from the beginning. The output function for a sequence
control system has to be some kind of pattern generator. It may be possible
to stop the pattern generator and later start it again from the next
point. I know we can do this with walking-pattern generators. There are
many points during a full stride where you can pause, in balance, and then
go on. You can do the pattern fast or slow, or even pause it and then run
it in reverse from where it was paused. Vocal pattern generators probably
can't do that, though maybe we could learn to do it at least between words
if not within words.

I'm not sure what we're doing here. Does this have something to do with the
proposed UEC?

Best,

Bill P.

4

[From Bruce Nevin (2002.05.17 00:26 EDT)]

Bill Powers (2002.05.08.0807 MDT)–

Let’s convert to the more general question:
why does an increase in error cause a decrease in effort, and the
opposite (first, of course, asking if

this is ever observed).

That is the first question. If this is a universal property of
output functions (as has been proposed), or in some other way is built
into every control loop universally, it should be very easy to find
examples. Conversely, there should be no counterexamples.
I’m standing up now … a stack of books starts to tumble, but I catch
them and lift them, bending a bit to do so, but without stopping the
process of standing up. It takes increased effort to overcome that
disturbance to my control of erect posture on my feet, and indeed despite
a transient increase in error my body smoothly produces the required
effort automatically, that is, without my particular awareness of it.
In fact, any of our stock examples of control as resistance to
disturbance will do as a counterexample. The wind blows the car, error to
my control of alignment with the lane increases, but I increase effort on
the steering wheel so as to counter the disturbance. I point at a bird,
someone jostles my shoulder, error in my control of fingertip-on-bird
increases, but I increase various muscular efforts precisely to counter
that disturbance.
Are there any positive examples? Consider Bill Curry’s horse bolting for
the barn to be fed. I contend that the horse is controlling a sequence
perception. The sooner this part of the sequence (getting back to the
barn) is completed, the sooner he can start the next part of the sequence
(eating).
I assume that you have started adding the UEC to the crowd program, the
e. coli demo, the baseball demo, arm, little man, and so on to see its
effects. If the UEC is universal then surely it is testable in these
models. Can you report what you have found?
To an external observer, it certainly appears that the when an e. coli
agent starts tumbling it must be because the error in its control of
nutrients has increased and so it has ‘given up’ eating. To an external
observer, when a crowd agent turns around and goes the other way, or
stops before reaching the goal just because some other agents intervene,
an obvious interpretation is that it has given up.
But we know that a crowd agent stops not because some other agent hinders
it but because it is concurrently controlling another variable, and it
halts at an equilibrium of an internal conflict (“go closer to
target” vs. “go no closer to other agents”). The
other-avoidance control system performs the function of the UEC with
respect to the goal-seeking control system. The UEC has no opportunity to
come into play.
You would have to look at situations where the agent was prevented from
controlling a specified variable at all. To invoke the UEC, the
preventing would have to be done by disturbances outside the agent rather
than by internal conflict. Suppose we close a cul de sac after an agent
has entered it. In the crowd program the agent would just keep moving
around inside the box, it wouldn’t give up. What would the crowd agent
need in order to model the behavior of a human or an animal in that
situation?
Giving up seems to me to be
(a) In the face of irresistable disturbance, stop controlling X by means
that aren’t working and start looking for alternative means, or for
something that must be done first in order to then resume controlling X.
Either way, this is control of a sequence. (First control A, then B, then
X. Or, first find effective means of controlling X, then control X.)
(b) After a protracted period of failure to identify and control
prerequisite antecedent steps or alternative means of controlling X, wait
for something to change that might be exploited as alternative means of
controlling X or that might be exploited as preliminary steps prior to
controlling X (e.g. escape). This ‘waiting’ might take the form of simply
controlling other variables normally while seeming to have ‘given up’
controlling X.
In sequence control it can appear that the agent has stopped controlling
X, but that is not the case. This is shown by the fact that as soon as
the conditions for controlling X arise, the agent starts actively
controlling X. Just like the bully idling in the schoolyard waiting for
little Albert to appear. At each stage, control of X doesn’t stop,
i.e. if the disturbance goes away control actions resume.

To model that requires considerably more complexity than the agents have
in the crowd program.

The UEC is a programmatically simple mechanism – sounds like a kind of
circuit breaker in the output function, in one version anyway –
postulated for muscle fibers (Isaac has pointed out the artificiality of
that example), and generalized to every output function. I can see why it
would be seductive for a modeler. One simple device that accounts for so
much. Some simple bit of code that you reuse everywhere in the model.
That strength is also its weakness. Why not just postulate divine
intervention?

I’m reminded of Mencken’s aphorism, “For every complex problem,
there is a solution that is simple, neat, and wrong.”

It also underscores an essential hazard of modelling, or the method of
specimens. Just because a computer model behaves like the modelled
organism does not guarantee that the model provides an accurate portrayal
of what is inside the black box of the organism. More than one mechanism
may generate that behavior. Where is the corresponding mechanism in the
physiology of the organism? The muscle fiber mechanism, even if it is a
valid explanation as far as it goes, is found functioning in that way
only in muscle fibers. It tells us nothing about an output function
implemented with neurons and neurochemicals higher in the hierarchy.
Where are the data that are being modeled? Supposing you find examples of
an inverse relationship of error and effort, is the curve the same in all
of them? If it is universal, why are there any exceptions – how do you
determine when the UEC works and when it … ummm … gives up? Where
there are alternative explanations (and I hope I have shown why I believe
there is always an alternative explanation), what determines which is
true?

If there are other explanations for at least some cases (e.g. internal
conflict as the explanation of the crowd agent appearing to ‘give up’;
sequence control in the “barn-sour” horse) why not see how much
can be accounted for with existing properties of the model without this
new and extremely powerful hypothesis?

Not too many years ago you were making much the same argument to me and
to others who were suggesting various changes to aspects of the theory.
“We’ve barely scratched the surface, let’s see how far we can get
with the conceptual apparatus that is already established, without adding
a bunch of new bells and whistles.” I’ve been taken aback by your
infatuation with this UEC notion. I guess that its appeal is as a tidy
bit of code that you can reuse everywhere. But there is no particular
reason to suppose that efficiencies of programming languages for
simulations on digital computers are mirrored in the evolution (and
ontogeny) of massively parallel analog systems in living organisms.

So,

o Are there any data?

o If so, are the data really parallel at different levels of the
hierarchy? Or even from example to example at the same level? Is it
really one thing that we are talking about?

o For each apparent example, have we really exhausted alternative
explanations employing existing properties of the theory without the
UEC?

    /Bruce

Nevin

···

At 02:44 PM 5/8/2002 -0600, Bill Powers wrote: