Cued chains of responses

[From Bruce Gregory (991110.1234 EST)]

Bruce Abbott (991110.1030 EST)

Thee:

Sorry, your mistake. My statement was "the system will begin to reorganize
to as to CONTROL FOR the perception of the reinforcer." As I recall,
control involves VARYING ACTIONS so as to produce a prespecified RESULT.
Please explain how this statement can be reduced to your "mindlessly
repeating the same actions." Failing that, how about a gentlemanly apology
for your mindless S-R kneejerk response? (:->

Me:

Since I drive essentially the same route to work each morning, there is no
need for me to pay the "cognitive overhead" of controlling for a sequence of
perceptions. If I drove through the toll gate at Auburn and immediately
encountered the toll gate at Weston (instead of driving for thirty miles) I
would be cued by the Weston tolls to pass through them and continue,
skipping nary a beat. In fact, I guess I wouldn't even notice that something
was amiss, since I would be responding appropriately to the cues I
encounter.

Yes, I'm not disputing that. I'm suggesting that under certain conditions
the organism may default to the simpler process of using the cues as they
arise consequent on the control processes that produce them as the souces

of

the next reference values.

How does the system "recover" from its default state?

Thee:

I assume that the cued behavioral chain was set into motion by a
higher-level system for the purpose of producing a specified outcome. When
the expected outcome does not materialize, the lower-level process is
interrupted and the problem dealt with.

Me: Oh? How does this higher level system know when to start worrying
because the expected outcome has not materialized?

Thee:

I should also explain what I had in mind by my use of the term "default." I
certainly do not mean that the default state of the entire organism is an
S-R mode of operation. We were discussing how a system might produce a
sequence without exercising control over the sequence. I suggested that
with sufficient practice under conditions where the cued sequence reliably
leads to the intended result, execution of the sequence might be turned over
to this simpler, unsupervised process. That's basically what the issue
comes down to -- supervised (controlled or closed-loop) vs. unsupervised
(uncontrolled or open-loop) execution of movement to the next element in the
chain.

Indeed it is the issue.

That is not sufficient. If a sequence has been progressing without error,
how do you know that the first error is not about to occur?

Why would I have to? Does your sequence-controller know when an error is
"about to occur"? No. So what's the point?

It's clear that you don't get the point. How does the system recover from
the "default" state? How does it know that it _needs_ to recover from this
state? How does it switch from an S-R system to a control system?

Thee;

I understand _your_ point, but Bill's statement, quoted above, doesn't seem
to be making any point at all. No system, control or otherwise, "knows
whether an error is or is not about to occur," unless it has available to it
some reliable predictor of such an event.

Me:

We differ then. I think that's exactly the point Bill is making.

Bruce Gregory

[From Bill Powers (991110.0923 MDT)]

Bruce Abbott (991109.2015 EST)]

If the "cue" indicates that one should turn at the next intersection, I
don't see any problem.

Bill:
How do you arrange for a street sign at an intersection to "indicate that
one should turn"? And how do you arrange for it to indicate this to one
driver, but not to another, and never to a dog? In other words, I am asking
you to explain your model -- how does this "indication" process work? Or
would you rather try to explain how the process you call "association" works?

If, as you say, this "indication" tells you to turn left at "the next
corner," how does the system for doing that (turning at the next corner)
work? Does the next corner say "Yoo, hoo, I am the next corner and you are
supposed to turn NOW?" Or is there some system that waits for the next
corner to be perceived, and executes a left turn when that perception
appears? And if that is the case, why couldn't it wait for the right street
sign to be perceived, and execute a left turn when it is perceived? In
short, why aren't you simply describing sequence control?

If one driver who wishes to turn at Lincoln street approaches the
intersection with Washington street, why does not the cue " WASHINGTON ST"
causes him to turn, while a second driver who wishes to turn at Washington
Street does turn? Is the cue "WASHINGTON ST" not identical in both cases?
Is it something about the cue, or something about the driver that causes a
turn?

An association inside the driver, you say. But later the same day, the same
driver approaches the same intersection with the same street, and sails
right on past it. What determines whether a given association will be in
effect? How can the fact that the driver has different destinations in mind
possibly determine which association will be active? Could it be that you
have to take into account what the driver WANTS?

Bruce:

Also, one can be maintaining some variables within a
particular range of values independently of the system orchestrating when
one should turn or go straight. It's a massively parallel system. So I can
have a system that keeps the car centered in its lane, and this system will
continue to do its job whether some other system orders a left turn or not.

Bill:
Not unless the other system can change the specification for which lane the
car is to be kept centered in. A left turn does not have the visual
appearance of driving down a straight road. Independent systems cannot
independently control variables that strongly affect each other. That is
why hierarchical control is required. If there is a steering control
system, then ALL steering must be done by adjusting the reference signal(s)
for that single or multiple control system. If you abandon that idea, then
you're stuck with Rodney Brooks' "subsumption" architecture, in which every
new system that needs to execute a turn would need its own turning-control
system, as well as the ability to inhibit all the other ones. This involves
not only massivly parallel systems, but massive duplication of functions.

I hate words like "orchestrating" and "indicating" qand "association."
These are words that sound vaguely as if they mean something, but on close
inspection reveal only that their user has no idea of what he is talking
about.

Bruce:

One can have a system that controls the centripital force experienced by
adjusting the speed of the auto according to the sharpness of the turning
angle, and so on. I don't see that these systems must be receiving their
references from the system that is setting references for turning here and
going straight there.

Bill:
If not, then there will instantly be a conflict when either system
institutes a turn. The force-control system will have its perception
disturbed by _any other_ system that actually causes a turn to occur. If
the higher system does not change the reference signal for experienced
force at the same time the other system is told to turn right or left, the
inertial control system will try to turn the steering wheel to keep the
experienced force from changing from its current reference level.

Bill:

Your way of conceiving this process is reflected in many other concepts of
behaviorism. For example, when a reinforcer occurs, it is said to increase
the probability that the behavior which led to the reinforcement will occur
again under the same circumstances.

Bruce:

Another way of stating this is that the system will begin to reorganize so
as to control for the perception of the reinforcer. The organism _wants_
the reinforcer and reorganizes so as to be able to produce it on demand.

Bill:
See the next paragraph of my post.

Bruce:

If what I stated above is correct, then your dichotomy is false. Control
gets established because the reinforcer is desired, and reorganization
occurs until that desire is fulfilled (if that is possible under the
circumstances). The pellet appears because the behavior produces it (the
behavior is the cause of the reinforcer), but the control system exists
because the rat wanted the pellet, and reorganized until its behavior
produced it.

Bill:
Let's not confuse how the control system came into being with how it works.
How it works is that there is a reference signal specifying a certain
quantity of reinforcer to be tasted, ingested, or whatever the immediate
controlled variable is. Once that control system has been acquired, it is
the animal's reference signal entering that system from higher systems that
determines the quantity of reinforcer that will be obtained, if possible.

What is viewed as the reinforcer itself may or may not be the input that is
critical to learning this control system. If the reinforcer is food or
water or some such necessity, we can suspect that it is at least closely
related to the intrinsic variable that is involved in reorganization. As
long as this intrinsic variable is far from its built-in reference level,
reorganization will continue. Reorganization will cease when the organism
is behaving so as to control some sense-based perception (not necessarily
directly related to the intrinsic variable) at the required reference
level. Of course doing this must also be maintaining the intrinsic variable
near its intrinsic reference level, for that, and not the sensory
experience, is the basis of reorganization.

Bill:

In a parallel way, we have a choice of seeing the cued behavior
differently. We can see a cue simply as a perceptual consequence of the
behavior that preceded it, rather than as a cause of the behavior that
follows it. Instead of seeing the chain as being grouped cue-behavior,
cue-behavior, we can see it as behavior-consequence, behavior-consequence,
with the consequences (which you call cues) being perceived. The organism,
under this view, selects the next consequence to be produced,

Bruce:

Yes, but this leaves unanswered why the organism makes any particular
selection. I would say that it does so because it has learned that a
particular selection makes further action possible, which eventually leads
to a desired perceptual state. If following those selections reliably leads
to the desired end-state, then with practice the "decisions" will become
automatic and stimulus-bound.

Bill:
I would say that the organism basically doesn't give a damn about action;
all it cares about are perceived results. So indeed, the organism makes a
selection of the next perception to be created as a means of achieving some
other perception. That is the essence of Hierarchical Perceptual Control
Theory.

As to the automatizing of the "decisions" and action, whether that is
feasible depends entirely on how rigidly the environment is protected
against unexpected disturbances, and how rigorously the organism can repeat
exactly the same behaviors so that exactly the same results will occur even
in a perfectly consistent environment. In the laboratory, you can protect
against most disturbances, and make sure the experimental conditions are
completely repeatable. However, even when stringent precautions are taken
you do not get the same behavior repeated over and over. In a simple
T-maze, in which the food is always at the same end of the crossbar and
the environment is absolutely the same on each trial (even to the rat's
being confined in a form-fitting box so it starts exactly the same way on
each trial), and even after extensive practice, the rat will choose the arm
leading to the cheese only about 95% of the time.

In a natural environment, repeating exactly the same behavior each time
food was sought would guarantee starvation.

Bruce:

I'm suggesting that under certain conditions
the organism may default to the simpler process of using the cues as they
arise consequent on the control processes that produce them as the souces of
the next reference values.

What on earth gives you the idea that the process you propose is "simpler"
than a control process? To produce a correct result in an open-loop way
requires complex computations and the sensing of all possible disturbances,
as well as the ability to compute the expected effects of the disturbances
on the outcome and to modify the output choice accordingly.
Bruce:

As for the HPCT sequence-controller having a mechanism to report the state
of the sequence, it seems to me that the signal which this mechanism
generates is too ambiguous for the purpose.

Bill:

If you want to propose a mechanism that is too ambiguous, you can. Why not
propose one that is adequate, once you know what is needed?

Bruce:

Pardon me? The mechanism I described is _your_ proposal, not mine. So, why
don't you ask yourself that question? (:->

Bill:
Are you implying that my proposal for a sequence-perceiving input function
in B:CP is the final design for a whole sequence-control system that I, or
anyone else, could come up with? If you or I were serious about coming up
with a workable model of this level of control, I assume that we would
specifically study the control of sequences by human beings, and then on
the basis of what we found, design a model accordingly. You're not going to
stick me with an imaginary straw-man model of sequence control that I never
proposed.

Bill:

Of course other information is needed, too, to allow the person to know
when one step has been finished and the next one can be initiated.

Bruce:

Like, for instance, the appearance of what I've been calling a "cue?

No. I don't know how you distinguish between a perceptual signal that is a
cue and another one that is not a cue. The word "cue" is loaded with
unspoken implications that give it some special role that other perceptions
do not have. I don't believe there are any such special perceptions. You
have to explain how the simple occurrance of a perception like the
lettering on a street sign can be turned into just the choice and timing of
turning behavior that is required on that day, in those weather conditions,
in that car, with that traffic present, in this driver's current state of
fatigue, and in relation to this driver's current intentions as to the
destination. You're simply sweeping all the complexities under the rug so
you can pretend they're not there, when you say "It's simple: the cue
triggers off the right behavior." Balderdash.

Bruce:

It begins with zero value and
increments by some amount as each perception in the sequence appears. A
moderate level of this signal at the end of the sequence may remain below
reference, but this by itself cannot inform the output function what

part of

the sequence was incorrect.

Bill:

Then that is not what this kind of perception is for.

Bruce:

It is what you proposed for sequence control in B:CP and it is the variable
Rick so confidently pointed to when asserting that sequence control is based
on a scalar variable whose intensity indicates the degree of error in the
sequence.

Bill:
God damn it, Bruce, when are you going to stop that ridiculous insistence
that I have proposed a model of sequence control and am stuck with it
forever? I described a _possibility_. I don't care what Rick proposed; if
you want to criticize it, aim your arguments at him. I haven't made any
serious proposals about sequence control because I don't know anything
about how real people actually control sequences. Neither do you.

Bill:

But some such
perception is needed, because people can judge, I claim, whether they are
near the beginning, middle, or end of a sequence in progress.

Bruce:

This only tells me that people can perceive the sequence in memory and
locate where they currently are in the remembered sequence.

Bill:
That's your theory, which requires imagining some information and proposing
a half-assed model. Now you have to explain how a person can perceive a
sequence in memory, perceive the current element, find that element in the
memory sequence, and judge where (what do you mean by "where"?) they are in
the remembered sequence.

Bruce:

I see no trouble designing a system
that will behave appropriately, once we know what that is. What I was
having problems with is the specific proposal made in B:CP, which seems
incapable of doing the job as I see it.

I don't know what you thought you saw in B:CP, but what was actually there
was a diagram showing one way in which sequence perception might occur, so
that one sequence could be perceived as different from another sequence. I
don't think I said anything about how the whole sequence control system
would work. I don't think I said that this was my final model of sequence
perception, or that there were no other ways to achieve the same thing.
You're having self-righteous fun with your straw man. I wish you'd quit.

···

-------------------------

Exactly so. I have suggested that this open-loop mode might develop under
conditions where such disturbances are unlikely.

Bill:
What conditions are those? What judges whether the disturbances are
"unlikely" (does that mean the same as "impossible?"), and what selects the
kind of system that will be used? What "develops" this mode?
-----------------------
Bill:

That is not sufficient. If a sequence has been progressing without error,
how do you know that the first error is not about to occur?

Bruce:

Why would I have to? Does your sequence-controller know when an error is
"about to occur"? No. So what's the point?

The point is that NO system can know whether a disturbance is about to
occur, so any practical design has to assume that a disturbance _will_
occur. The way you design a system so it can negate the effects of
disturbances is to design it as a control system. Then, if there are no
disturbances, it will produce the right outcome, and if there should happen
to be some disturbances, their effects will be minimized.

Bill:

And anyway, the
idea of a sequence error doesn't even apply in the cued-behavior model --
the only error that can occur is responding in the wrong way to whatever
cue occurs.

Bruce:

Right!

Then why are you talking about sequence at all? The cued-behavior model is
simply an S-R model of behavior. It can work only in an imaginary universe
and when carried out by an imaginary organism. The conditions required for
this model to work are never, ever, observed in the real world, not even in
the laboratory where an attempt is made to establish the required conditions.
-----------------------------
Bill:

The design issues we're talking about can't be settled in an ad-hoc way;
each involves considerable complexities so that a neural substrate suitable
for one kind of system would not be suitable for the other. Sequence
control versus cued chaining is a design difference that would result from
evolving along very different lines. The concept of an open-loop response
gets into prediction and modeling of the lower systems and the environment,
and all the complexities implied by that inverse-dynamics,
inverse-kinematics, world-modeling approach.

Bruce:

You keep slipping this in. It's not relevant. Only the initiation of the
next element in the sequence is carried out open loop in my proposal. We're
talking stimulus-reference, not stimulus-response, and the problems of
inverse-dynamics etc. do not enter in.

Bill:
No, but the problems of deciding which behavior to select do enter in, and
the problems of what to do when the expected sequence does not occur do
enter in, and the problems of organism-caused mistakes do enter in. You
proposal can work only under ideal conditions -- that is, imaginary
conditions.

Bill:

A control system organization
accomplishes the same overall effects, but in a far simpler and
structurally very different way.

Bruce:

If sequences _can_ be produced as my proposal suggests (which is not the
same as saying that they _are generally_ produced that way), it is actually
simpler than sequence-control, requiring less cognitive overhead.

Bill:
Your proposal does not involve producing any sequences. It merely says that
stimulus A causes the choice of response B, regardless of what happens
before A or after B. The _experimenter_ may arrange for the response to
produce another stimulus C on the list of known stimulus-response pairs, to
produce behavior D. But that sequence control is being done by the
experimenter. To explain it, we would need to analyze the experimenter, not
the experimental subject.
----------------------------
Bill:

I just don't think it's a tenable idea to
say that the organism adopts whichever strategy works for it. Once you've
committed to one organization, I think you've ruled out the other.

Well, that's one place where we strongly disagree.

Now, there's a convincing argument if I ever heard one!
----------------------------
Bruce:

Perhaps I did not make this sufficiently clear. In my proposal, the "cue"
serves as a _retrieval_ cue for associative memory; it retrieves the
parameters required to set up the next control process in the sequence
(e.g., the reference value).

Ah, I see. It does, somehow, whatever is required to make your idea work.
Asking what is required is not in the rules of the game the way you play
it. It is assumed that there is just one "next process in the sequence"
that will serve to achieve the higher-level goal under all circumstances.
And since you have said that, it must be true.
------------------------
Bill:

I was misled by your speaking as if you knew that reliable
sequence-production could occur open-loop. I now understand that you were
only conjecturing that it might be possible.

Yes.

-----------------
Bill:

doing something about it, and also showing that the elements of the
sequence are being generated independently.

Bruce:

Independently of what?

Bill:

Of the environment. When the person repeats part of the sequence after the
target movement has stopped, this shows that there is a sequence-generator
inside the person.

Bruce:

If the cues arise as a result of completion of each behavioral act, then the
sequence would continue under the open-loop proposal as well.

Bill:
Oh, that. What you're describing now is a closed loop system (an
oscillator) in which the _only_ source of the cue is the behavior itself.
The perception is produced by the action, and the action is the result of
comparing the perception with a reference perception (possibly zero). So
what's a "cue" that makes it different from a perception?
------------------------------
Bill:

We also know that from the fact that prior to the
target's stopping, the person is matching movements to the target movements
at a speed too great for simple tracking to work.

Bruce:

Please explain.

Bill:
I move my finger up, right, down, up, left, down and so on in a continuing
pattern of quick moves, while you track it with your finger. If I go very
slowly, you don't need to recognize the pattern; you track just as if my
finger movements were random. As I speed up, however, there will come a
speed at which your ability to correct errors in that way starts to
deteriorate. At a slightly higher speed, assuming you recognize the
pattern, you will begin generating that pattern yourself, and adjusting it
so the amplitude, phase, and perhaps a few points along the way match what
I am doing.

In the slower regime, if I suddenly stop my finger, your finger will stop,
too, after a very short delay. In the faster regime, if I suddenly stop my
finger, your finger will execute one more and perhaps two or three more
elements of the up, left, down, up, right, down sequence. This shows that
sequence control is slower than position control.
------------------------------
Bill:

Sorry, I have no idea what that means. How is a reference perception
different from a comparison perception?

Bruce:

If a sequence of perceptions is being read out of memory, this may _supply_
the references for the lower-order systems. Alternatively, the sequence of
memory-perceptions may be compared in real time to the arriving sequence of
perceptions; the output of the perceptual function would be a series of
same-different judgements. The reference of this system might be for
"same," in which case a "different" perception would generate an immediate
error when it occurred.

Bill:
Your model contains a sequence, but no sequence-perceiver. If you read out
of memory a set of elements one after another, you will get a set of
individual elements, one after another. What is needed is something that
will receive this set of elements as they occur and emit a signal if and
only if _one particular sequence is occurring_. That is what is required if
you are to explain how we can hear "Mary had a " in the right pitches, and
from that moment until the last stanza is sung perceive that _the same
sequence is in progress_. If, at any moment, you hear that something is
left out, changed, or inserted, that perceptual signal will change and will
cease, to some degree, to be that sequence.

My concept of sequence perception refers to that kind of phenomenon, which
differentiates between the _occurrance_ of a sequence (as an observer
perceives it) and the _perception_ of a sequence.
--------------------------------
Bruce:

Again, I don't agree that the architecture has to be either/or. I
understand that this violates one of the HPCT design principles and that it
should not be accepted without solid empirical evidence. But it is a
testable proposition, and seems consistent with some of my own inner

experience.

The architecture determines what computations can be carried out, and that
depends on the type of neurons that are present and how connections can
grow and change among them. For comparators to exist, for example, we must
have neurons that receive both excitory signals and inhibitory signals from
the afferent and efferent sides of the brain, respectively (although which
is excitatory and which is inhibitory can differ at different levels).
These connections are there and well-identified. In the spinal cord we even
have both arrangements, interconnected, set up between antagonistic muscle
control systems. In the brain stem, the comparators are in a layer of the
output nuclei, at least in the Red Nucleus; the sensory feedback signals
are uniformly the excitatory ones and the reference signals the inhibitory
ones.

The physical arrangements and propensities, therefore, appear to be set up
to favor the development of negative feedback control systems; this
continues at least to the inner layer of the cerebral cortex, known to be
involved in motion or transition perception.

Here is my final take on this matter, for now:

Suppose we have a control system with a reference signal of zero. How will
it behave when a sudden perceptual signal arises? The effect of the
perceptual signal will be to create an error signal, which passes through
the output function to become a change in lower-level reference signals. So
for a zero or constant reference signal, the control system will behave
exactly as if the change in perceptual signal were a "cue" and the change
in lower-level reference signal (or the effects thereof) were a "response"
to that "cue".

You will notice that _any_ control system at _any_ level will appear to be
a cued-response system if its reference signals remains constant and if its
perceptual signals change suddenly. So your proposal can be taken as a
subset of the control-system proposal when behavior is viewed under certain
conditions.

As to sequence control, sequence control as envisioned in HPCT requires not
just that sequences exist, but that a perceptual signal exist representing
the state of an ongoing sequence. A steady perceptual signal would indicate
that a particular sequence is in progress; that is required to explain how
we experience ongoing familiar sequences.

Your proposal does not address sequence control per se, but only the
appearance of producing particular reactions to particular stimuli. We can
explain those appearances in terms of disturbances of control systems with
constant reference signals. We do not need a different model for every
different situation; the whole point of modeling, from my point of view, is
to find a _single_ organization that will account for all the various
phenomena we observe.

Best,

Bill P.

[From Bruce Abbott (991110.1430 EST)]

Rick Marken (991009.2245) --

Bruce A? (more likely Bill P.)

It [sequence perception signal] begins with zero value and
increments by some amount as each perception in the sequence
appears. A moderate level of this signal at the end of the
sequence may remain below reference, but this by itself cannot
inform the output function what part of the sequence was incorrect.

More likely Bill P.? Pretty good, isn't it? Sorry to disappoint you, but I
_did_ write that.

Then that is not what this kind of perception is for.

Bruce Abbott (991109.2015 EST) --

It is what you proposed for sequence control in B:CP and it is
the variable Rick so confidently pointed to when asserting that
sequence control is based on a scalar variable whose intensity
indicates the degree of error in the sequence.

Bruce, you seem to think that Bill said something like "then
a sequence control system doesn't control a scalar perceptual
representation of the sequence".

What I _think_ is that the scalar perceptual representation of the sequence
provided by the mechanism described in B:CP be used to control (reliably
produce under conditions of non-overwhelming disturbance) the production of
a sequence.

What Bill actually said was
"then that is not what this kind of perception [scalar perceptual
representation of the sequence] is for".

I took Bill to mean what your elaboration suggests he meant. However, that
raises the question of what it _is_ for, and why it was described in B:CP in
the chapter on sequence control.

What Bill meant, I
believe, is "informing the output function what part of the
sequence was incorrect is not what this kind of perception
is for". It's for something else.

I notice that you don't offer any suggestion as to what that "something
else" might be. You don't know, do you?

It was presented in B:CP to show how a sequence-perceiver might be built
that would produce a scalar output whose value was proportional to the
degree to which the input perceptions occurred in the specified sequence.
It was supposed to provide an existence proof that something capable of
recognizing a sequence and generating an output perceptual signal
proportional to the "degree of sequence" could be constructed from neural
elements. What I've pointed out is that a sequence-control system designed
to produce a specified sequence by manipulating the references to
lower-order control systems, and which used the proposed sequence-perceiver
to generate its perceptual input function, would not be capable of
recognizing specifically where in the sequence an error occurred, or what
its nature is (e.g., wrong element, correct element out of sequence).

Determining what that
something else is is a job for a modeller who has to build
a model of a sequence control system that must produce different
outputs to correct errors that occur at different points in the
sequence. This problem is finessed in my sequence control
experiment since the same output (pressing) can be used to
correct a sequence error regardless of where in the sequence
the error occurs.

I'd say the problem is _avoided_ in your demo. What you have is pattern
matching: you keep changing the pattern until it matches the one specified.
(A sequence is just a pattern extended in time.) This is similar to aiming a
gun at a target and firing. The bullet's path is not controlled (any
disturbances acting on the bullet cannot be corrected once the bullet is in
flight), but across repetitions, you can correct the aim until the target is
hit. This sounds very much like Ashby's open-loop + error-based
parameter-adjustment mechanism.

I don't believe that Bill's comment calls into question the
PCT model of sequence control which I so "confidently pointed
to"; at least, it dosn't call into question the PCT notion that
what is controlled is a scalar perceptual representation of
a sequence.

In true control of the sequence, errors would be corrected _as they occur_
so that the intended sequence is produced each and every time, unless the
system is overwhelmed by disturbance. What is clear is that the sequence
perceiver sketched in B:CP cannot serve as the input function of a
traditional PCT-style control system (in which the error between perception
and reference gets amplified to drive the output that opposes the effect of
the disturbance) and do the job required of it for sequence control. The
error signal produced does not provide the information required by the
output function in order to do what is necessary to correct the error.

Besides, all that can really call the model into
question is evidence that the model does not behave like the real
system.

Not so! If analysis of the proposed model shows that it does not meet the
logical requirements for this type of control, then it can be rejected
immediately. Unless the model meets this test, it is not worth Testing.

As I noted, I have built a simple PCT model of sequence
control and found that it fits the behavior of subjects in
my sequence control experiment. That's why I "confidently
pointed to" the PCT model as a model of sequence control. Of
course, I would lose a great deal of confidence in the model
if someone produced evidence that the model does not fit the
data. But so far, I am the only one who has collected the
relevant data and, so far, what data I've collected is
perfectly compatible with PCT.

As you also noted, this model avoids some knotty problems in real sequence
control. Your "sequence control" model is really just a Darwinian selection
mechanism. The computer produces variants and the selection mechanism
preserves only the variant that meets the criterion of matching the sequence
you wish to see.

Again, I invite you to return to this discussion when you have
collected some relevant data or when you have a working model
that explains some existing (relevant) data.

I agree with the spirit of this suggestion, but its real purpose in being
raised here is to shut off further discussion of the issue, as if only
empirical work can provide any answers. Yet this is patently false. By
simply examining the logical requirements of the sequence-control task, I
have been able to show that the simple sequence perceiver proposed in B:CP
is not up to the task it was presumably designed for. Thus, it does appear
that there is something positive to be gained by such examinations. This
is, after all, how HPCT came to be in the first place.

Regards,

Bruce A.

[From Rick Marken (991110.1345)]

Bruce Abbott (991110.1430 EST)--

What I _think_ is that the scalar perceptual representation
of the sequence provided by the mechanism described in B:CP
be used to control (reliably produce under conditions of
non-overwhelming disturbance) the production of a sequence.

Yes. What you think is that perception is "used" to control
behavior. This is the cognitive version of the S-R model.
It is not the PCT model. In PCT the perceptual representation
of the sequence is _controlled_. Behavior is the control of
perception.

I notice that you don't offer any suggestion as to what that
"something else" might be. You don't know, do you?

That's right. I don't know. "Knowing" means being able to
build the model. I have some ideas, based on thinking about
control of variables that are defined over time (sequences)
in terms of control of variables that are defined over
space (like sensations). So a sequence control system that
must vary references for several lower level variables in
order to control the sequence might work something like the
sensation control systems in my spreadsheet hierarchy.

I'd say the problem is _avoided_ in your demo. What you
have is pattern matching:

What I have is control of input. The model I used to simulate
this process was _very_ simple; there was no perceptual function;
the perceptual signal was just a variable whose value was 1.0
when the sequence was in one state and 0.0 when it was in another.
The point was not to develop a detailed model but to show that
controlling a variable that is defined over time can still be
conceived as control of the _current state_ of a scalar input
variable; that is, it is possible to control a sequence by
controlling a scalar variable that does not vary sequentially.

What is clear is that the sequence perceiver sketched in
B:CP cannot serve as the input function of a traditional
PCT-style control system

Um. Excuse me but that is precisely what my little demo and
model show: the sequence perceiver sketched in B:CP _can_
serve as the input function of a traditional PCT-style control
system that controls a sequence.

Your "sequence control" model is really just a Darwinian
selection mechanism.

Where do you come up with this stuff? It's an input control
system; the input variable (a perceptual signal representing
the state of a sequence) is kept in a reference state,
protected from disturbance.

Me:

I invite you to return to this discussion when you have
collected some relevant data or when you have a working model
that explains some existing (relevant) data.

Bruce:

I agree with the spirit of this suggestion, but its real
purpose in being raised here is to shut off further discussion
of the issue

Actually, that's the spirit of the suggestion, too. Come back
when you have some relevant data or models. Do something
worthwhile for a change.

Best

Rick

···

--
Richard S. Marken Phone or Fax: 310 474-0313
Life Learning Associates mailto: rmarken@earthlink.net
http://home.earthlink.net/~rmarken

[From Bruce Abbott (991110.2110 EST)]

Bill Powers (991110.0923 MDT) --

Bruce Abbott (991109.2015 EST)

If the "cue" indicates that one should turn at the next intersection, I
don't see any problem.

Bill:

Is it _really_ necessary to label each paragraph with the name of the writer
(as above)? We got along fine without that device for years, now suddenly
it's become a fad. I think it lengthens posts unnecessarily.

How do you arrange for a street sign at an intersection to "indicate that
one should turn"? And how do you arrange for it to indicate this to one
driver, but not to another, and never to a dog? In other words, I am asking
you to explain your model -- how does this "indication" process work? Or
would you rather try to explain how the process you call "association" works?

Let me turn this around and ask how it works in sequence-control as you
envision it. How do the elements get stitched together so that the sequence
is followed, and what determines when one element in the control sequence
will be replaced by the next?

If, as you say, this "indication" tells you to turn left at "the next
corner," how does the system for doing that (turning at the next corner)
work? Does the next corner say "Yoo, hoo, I am the next corner and you are
supposed to turn NOW?" Or is there some system that waits for the next
corner to be perceived, and executes a left turn when that perception
appears? And if that is the case, why couldn't it wait for the right street
sign to be perceived, and execute a left turn when it is perceived? In
short, why aren't you simply describing sequence control?

I have nothing against the idea of sequence control. But this requires
advance knowledge of the sequence to be produced (to provide the means of
assessing that the proper sequence is in progress). Yet it is possible to
produce -- and reproduce -- a given sequence without that knowledge, if the
perceptual state achieved at the end of one control process serves as the
retrieval cue that allows one to recall what control process should come
next, and with what reference. If the sequence per se isn't being actively
monitored, the individual will switch control tasks and/or reference values
when the usual cue for such a switch appears, even if out of the usual
sequence, and will fail to take appropriate action if the cue is present but
not noticed by the actor. This could happen even if the person knows the
proper sequence, if he or she isn't paying conscious attention to the cues.

If one driver who wishes to turn at Lincoln street approaches the
intersection with Washington street, why does not the cue " WASHINGTON ST"
causes him to turn, while a second driver who wishes to turn at Washington
Street does turn? Is the cue "WASHINGTON ST" not identical in both cases?
Is it something about the cue, or something about the driver that causes a
turn?

Both. If the driver is to turn when cued by WASHINGTON ST, then the driver
will turn whether it is appropriate to do so or not (given where the driver
intends to go). He or she will not turn when the sign reads MONROE ST. The
reason why WASHINGTON ST will cause him to turn, and not MONROE ST., is that
the driver has followed this sequence in the past and has now associated
seeing WASHINGTON ST. with turning left onto WASHINGTON ST. as a result of
previously having done so on several occasions while making his/her way to a
particular destination.

An association inside the driver, you say. But later the same day, the same
driver approaches the same intersection with the same street, and sails
right on past it. What determines whether a given association will be in
effect? How can the fact that the driver has different destinations in mind
possibly determine which association will be active? Could it be that you
have to take into account what the driver WANTS?

Also, one can be maintaining some variables within a
particular range of values independently of the system orchestrating when
one should turn or go straight. It's a massively parallel system. So I can
have a system that keeps the car centered in its lane, and this system will
continue to do its job whether some other system orders a left turn or not.

Not unless the other system can change the specification for which lane the
car is to be kept centered in. A left turn does not have the visual
appearance of driving down a straight road. Independent systems cannot
independently control variables that strongly affect each other. That is
why hierarchical control is required. If there is a steering control
system, then ALL steering must be done by adjusting the reference signal(s)
for that single or multiple control system. If you abandon that idea, then
you're stuck with Rodney Brooks' "subsumption" architecture, in which every
new system that needs to execute a turn would need its own turning-control
system, as well as the ability to inhibit all the other ones. This involves
not only massivly parallel systems, but massive duplication of functions.

Perhaps you have overlooked the fact that in my proposal, I have not
abandoned hierarchical control. If hierarchical control accomplishes all
these appropriate adjustments according to conditions when sequence-control
is in progress, then it continues to do so under cued behavioral chaining.

I hate words like "orchestrating" and "indicating" qand "association."
These are words that sound vaguely as if they mean something, but on close
inspection reveal only that their user has no idea of what he is talking
about.

I don't think that you've conducted any such close inspection, so you are in
no position to draw the conclusion that these are hand-waving terms. I used
these terms in order to keep the level of detail appropriate to the level of
operations being discussed (I don't want to loose the forest for the trees).
"Orchestrate" is what the references sent to lower-level control systems do
in a well-designed hierarchy (they contribute in a unified way to the
higher-level performance, so the term "orchestrate" seems particularly
appropriate). "Indicating" is self-explanatory. Association is a
well-known process whose neurological mechanism is poorly understood.
Neural networks now exist which demonstrate the principle, although the
nervous system probably does it differently.

One can have a system that controls the centripital force experienced by
adjusting the speed of the auto according to the sharpness of the turning
angle, and so on. I don't see that these systems must be receiving their
references from the system that is setting references for turning here and
going straight there.

If not, then there will instantly be a conflict when either system
institutes a turn. The force-control system will have its perception
disturbed by _any other_ system that actually causes a turn to occur. If
the higher system does not change the reference signal for experienced
force at the same time the other system is told to turn right or left, the
inertial control system will try to turn the steering wheel to keep the
experienced force from changing from its current reference level.

Understood. Still, it isn't a problem specific to my proposal as it is also
a problem for the sequence-controller.

Your way of conceiving this process is reflected in many other concepts of
behaviorism. For example, when a reinforcer occurs, it is said to increase
the probability that the behavior which led to the reinforcement will occur
again under the same circumstances.

Another way of stating this is that the system will begin to reorganize so
as to control for the perception of the reinforcer. The organism _wants_
the reinforcer and reorganizes so as to be able to produce it on demand.

If what I stated above is correct, then your dichotomy is false. Control
gets established because the reinforcer is desired, and reorganization
occurs until that desire is fulfilled (if that is possible under the
circumstances). The pellet appears because the behavior produces it (the
behavior is the cause of the reinforcer), but the control system exists
because the rat wanted the pellet, and reorganized until its behavior
produced it.

Let's not confuse how the control system came into being with how it works.
How it works is that there is a reference signal specifying a certain
quantity of reinforcer to be tasted, ingested, or whatever the immediate
controlled variable is. Once that control system has been acquired, it is
the animal's reference signal entering that system from higher systems that
determines the quantity of reinforcer that will be obtained, if possible.

I'm not guilty of the confusion you accuse me of. I understand the
difference between how the system comes into being and how it works.

What is viewed as the reinforcer itself may or may not be the input that is
critical to learning this control system. If the reinforcer is food or
water or some such necessity, we can suspect that it is at least closely
related to the intrinsic variable that is involved in reorganization. As
long as this intrinsic variable is far from its built-in reference level,
reorganization will continue. Reorganization will cease when the organism
is behaving so as to control some sense-based perception (not necessarily
directly related to the intrinsic variable) at the required reference
level. Of course doing this must also be maintaining the intrinsic variable
near its intrinsic reference level, for that, and not the sensory
experience, is the basis of reorganization.

Alternatively, there are no "intrinsic variables" that drive reorganization
(although certainly there are crucial variables that must be kept within
limits if the organism is to survive). Instead, over the course of
evolution brains have evolved for which certain perceptual inputs are
desired, under certain specific conditions. I control for the warm, medium
rare steak in my mouth (and ultimately, in my belly) when I am hungry, and I
perceive hunger when I am in a particular physiological state. I eat the
steak because under those conditions eating the steak is pleasurable and
reduces what is sometimes an unpleasant state of hunger. I do not eat the
steak because during reorganization, eating steak eliminated the error in
some "intrinsic" variable. I eat the steak because I have learned that the
steak tastes great when I'm hungry and satisfies my hunger. It tastes great
partly because, in the evolutionary history of our species, eating things
that tasted like steak inceased the reproductive fitness of our ancestors.

In a parallel way, we have a choice of seeing the cued behavior
differently. We can see a cue simply as a perceptual consequence of the
behavior that preceded it, rather than as a cause of the behavior that
follows it. Instead of seeing the chain as being grouped cue-behavior,
cue-behavior, we can see it as behavior-consequence, behavior-consequence,
with the consequences (which you call cues) being perceived. The organism,
under this view, selects the next consequence to be produced,

Yes, but this leaves unanswered why the organism makes any particular
selection. I would say that it does so because it has learned that a
particular selection makes further action possible, which eventually leads
to a desired perceptual state. If following those selections reliably leads
to the desired end-state, then with practice the "decisions" will become
automatic and stimulus-bound.

I would say that the organism basically doesn't give a damn about action;
all it cares about are perceived results. So indeed, the organism makes a
selection of the next perception to be created as a means of achieving some
other perception. That is the essence of Hierarchical Perceptual Control
Theory.

I should have been more careful in my choice of words; I meant to convey
when I stated "makes further action possible" the idea that it makes
possible the opportunity to bring about the next required perceptual state.
As for the essence of HPCT, please note that completing a sequence of
control acts as a means of reaching the final state in the chain is not the
same "means" as specified in HPCT and in fact may require violating the HPCT
hierarchy: e.g., the hungry rat presses the lever (event control?) in order
to reduce an intensity signal (hunger) to some low (reference) value.

As to the automatizing of the "decisions" and action, whether that is
feasible depends entirely on how rigidly the environment is protected
against unexpected disturbances, and how rigorously the organism can repeat
exactly the same behaviors so that exactly the same results will occur even
in a perfectly consistent environment. In the laboratory, you can protect
against most disturbances, and make sure the experimental conditions are
completely repeatable. However, even when stringent precautions are taken
you do not get the same behavior repeated over and over. In a simple
T-maze, in which the food is always at the same end of the crossbar and
the environment is absolutely the same on each trial (even to the rat's
being confined in a form-fitting box so it starts exactly the same way on
each trial), and even after extensive practice, the rat will choose the arm
leading to the cheese only about 95% of the time.

The above falls under the category of true but nearly irrelevant. If the
control systems being activated in sequence are able to overcome the
environmental disturbances that may occur, then there is a high probability
that the sequence will reel off as my proposal suggests. You have fallen
back to attacking my proposal as if it were open-loop all the way down,
which decidedly it is not.

In a natural environment, repeating exactly the same behavior each time
food was sought would guarantee starvation.

What does that have to do with anything? In such an environment the rat
would never have learned to go always to the same place for its food. Also,
if the open-loop sequence execution failed to result in food, the
higher-level system that initiated the behavioral chain would end up with an
error and take action to correct it.

I'm suggesting that under certain conditions
the organism may default to the simpler process of using the cues as they
arise consequent on the control processes that produce them as the souces of
the next reference values.

What on earth gives you the idea that the process you propose is "simpler"
than a control process? To produce a correct result in an open-loop way
requires complex computations and the sensing of all possible disturbances,
as well as the ability to compute the expected effects of the disturbances
on the outcome and to modify the output choice accordingly.

You keep repeating this as if I had never explained it to you before. The
control processes being executed in series handle these difficulties, so
that making complex computations and sensing all disturbances is
unnecessary. Ignoring that fact doesn't make it go away.

As for the HPCT sequence-controller having a mechanism to report the state
of the sequence, it seems to me that the signal which this mechanism
generates is too ambiguous for the purpose.

If you want to propose a mechanism that is too ambiguous, you can. Why not
propose one that is adequate, once you know what is needed?

Pardon me? The mechanism I described is _your_ proposal, not mine. So, why
don't you ask yourself that question? (:->

Are you implying that my proposal for a sequence-perceiving input function
in B:CP is the final design for a whole sequence-control system that I, or
anyone else, could come up with? If you or I were serious about coming up
with a workable model of this level of control, I assume that we would
specifically study the control of sequences by human beings, and then on
the basis of what we found, design a model accordingly. You're not going to
stick me with an imaginary straw-man model of sequence control that I never
proposed.

A moment ago you were chastising me for not providing a mechanism that could
work. Now that you realize that I'm talking about _your_ mechanism, that's
no longer a reason for criticism. Come on, Bill, you can't have it both ways.

Now if you'll calm down, I'll explain that I agree that you or I probably
could come up with a workable mechanism. All I've said is that the
mechanism suggested in B:CP can't do the job, and that much you've already
agreed.

Of course other information is needed, too, to allow the person to know
when one step has been finished and the next one can be initiated.

Like, for instance, the appearance of what I've been calling a "cue?

No. I don't know how you distinguish between a perceptual signal that is a
cue and another one that is not a cue. The word "cue" is loaded with
unspoken implications that give it some special role that other perceptions
do not have. I don't believe there are any such special perceptions. You
have to explain how the simple occurrance of a perception like the
lettering on a street sign can be turned into just the choice and timing of
turning behavior that is required on that day, in those weather conditions,
in that car, with that traffic present, in this driver's current state of
fatigue, and in relation to this driver's current intentions as to the
destination. You're simply sweeping all the complexities under the rug so
you can pretend they're not there, when you say "It's simple: the cue
triggers off the right behavior." Balderdash.

Achieving a certain perceptual _state_ is necessary before the next element
in the sequence can be brought into play. That state is what I'm calling
the "cue." It has no special status as a perception except that its
appearance in perception tells the system that the next element in the
sequence can now be initiated. I'm suggesting that the mechanism whereby
this happens is associative, whereby the occurrence of a specific perceptual
state produces the retrieval of the reference value and other parameters of
the next control-system element in the sequence. Don't you believe that
certain perceptual inputs give rise to certain remembered perceptions? FBI
= Federal Bureau of ____________. Now why did "Investigation" leap to mind?
Or are you going to stonewall me and deny that it happened, or ever happens
to anyone?

It begins with zero value and
increments by some amount as each perception in the sequence appears. A
moderate level of this signal at the end of the sequence may remain below
reference, but this by itself cannot inform the output function what
part of the sequence was incorrect.

Then that is not what this kind of perception is for.

It is what you proposed for sequence control in B:CP and it is the variable
Rick so confidently pointed to when asserting that sequence control is based
on a scalar variable whose intensity indicates the degree of error in the
sequence.

God damn it, Bruce, when are you going to stop that ridiculous insistence
that I have proposed a model of sequence control and am stuck with it
forever? I described a _possibility_.

Hey, don't swear at me for repeating this before receiving your first
response, as if you explained it all to me and then I repeated the complaint
anyway. 'Taint fair.

I don't care what Rick proposed; if
you want to criticize it, aim your arguments at him. I haven't made any
serious proposals about sequence control because I don't know anything
about how real people actually control sequences. Neither do you.

Well, at last we agree on something.

But some such
perception is needed, because people can judge, I claim, whether they are
near the beginning, middle, or end of a sequence in progress.

This only tells me that people can perceive the sequence in memory and
locate where they currently are in the remembered sequence.

That's your theory, which requires imagining some information and proposing
a half-assed model. Now you have to explain how a person can perceive a
sequence in memory, perceive the current element, find that element in the
memory sequence, and judge where (what do you mean by "where"?) they are in
the remembered sequence.

I don't think I've proposed any half-assed models here. The explanations
you request fall out of the associative nature of memory. Each remembered
(or perceived) element of the series triggers the recall of the next,
because they have become associatively linked. One should have no trouble
comparing each element perceived to the element evoked in memory and
determining whether they are a match.

I see no trouble designing a system
that will behave appropriately, once we know what that is. What I was
having problems with is the specific proposal made in B:CP, which seems
incapable of doing the job as I see it.

I don't know what you thought you saw in B:CP, but what was actually there
was a diagram showing one way in which sequence perception might occur, so
that one sequence could be perceived as different from another sequence. I
don't think I said anything about how the whole sequence control system
would work.

I've ready stated in this post and in a previous one that the sequence
perceiver proposed in B:CP would not work for sequence control in which
corrections occurred as soon as errors were made. You did state that the
outputs of each element could be summed to provide a scalar signal whose
level would indicate the degree to which the input matched the required
sequence. The implication was that it was then a simple step to build a
sequence controller that controlled this perceptual signal by emitting
outputs in response to error that would keep the sequence-perception under
control.

I don't think I said that this was my final model of sequence
perception, or that there were no other ways to achieve the same thing.
You're having self-righteous fun with your straw man. I wish you'd quit.

I've already discussed the requirements for such a sequence controller, and
I don't see any serious problems in implementing one, other than the fact
that it lacks elegance. I never suggested that the model you suggested was
anything more than an attempt to establish an existance proof. But I am
having fun and think you're being a pooh for trying to spoil it. So there.

Exactly so. I have suggested that this open-loop mode might develop under
conditions where such disturbances are unlikely.

What conditions are those? What judges whether the disturbances are
"unlikely" (does that mean the same as "impossible?"), and what selects the
kind of system that will be used? What "develops" this mode?

Experience with the sequence. If A leads to B, which leads to C, on every
occasion, then the environment permits this mode of operation to develop.
As we're just discussing a preliminary idea, I don't have answers to all the
details.

That is not sufficient. If a sequence has been progressing without error,
how do you know that the first error is not about to occur?

Why would I have to? Does your sequence-controller know when an error is
"about to occur"? No. So what's the point?

The point is that NO system can know whether a disturbance is about to
occur, so any practical design has to assume that a disturbance _will_
occur. The way you design a system so it can negate the effects of
disturbances is to design it as a control system. Then, if there are no
disturbances, it will produce the right outcome, and if there should happen
to be some disturbances, their effects will be minimized.

But if in the experience of the supervisory system, disturbances never
happen, the chain may be followed automatically, by habit. No need to check
for problems if problems never occur. It is a somewhat risky strategy, but
one that can work when disturbances to the cue sequencing are rare.

And anyway, the
idea of a sequence error doesn't even apply in the cued-behavior model --
the only error that can occur is responding in the wrong way to whatever
cue occurs.

Right!

Then why are you talking about sequence at all? The cued-behavior model is
simply an S-R model of behavior. It can work only in an imaginary universe
and when carried out by an imaginary organism. The conditions required for
this model to work are never, ever, observed in the real world, not even in
the laboratory where an attempt is made to establish the required conditions.

Not when the cue sequence is rarely disrupted; all other disturbances are
countered by the component control systems. Bet you forgot.

The design issues we're talking about can't be settled in an ad-hoc way;
each involves considerable complexities so that a neural substrate suitable
for one kind of system would not be suitable for the other. Sequence
control versus cued chaining is a design difference that would result from
evolving along very different lines. The concept of an open-loop response
gets into prediction and modeling of the lower systems and the environment,
and all the complexities implied by that inverse-dynamics,
inverse-kinematics, world-modeling approach.

You keep slipping this in. It's not relevant. Only the initiation of the
next element in the sequence is carried out open loop in my proposal. We're
talking stimulus-reference, not stimulus-response, and the problems of
inverse-dynamics etc. do not enter in.

No, but the problems of deciding which behavior to select do enter in, and
the problems of what to do when the expected sequence does not occur do
enter in, and the problems of organism-caused mistakes do enter in. You
proposal can work only under ideal conditions -- that is, imaginary
conditions.

Again, for the reasons given, the supposed "ideal" conditions are probably
rather common as opposed to imaginary.

A control system organization
accomplishes the same overall effects, but in a far simpler and
structurally very different way.

If sequences _can_ be produced as my proposal suggests (which is not the
same as saying that they _are generally_ produced that way), it is actually
simpler than sequence-control, requiring less cognitive overhead.

Your proposal does not involve producing any sequences. It merely says that
stimulus A causes the choice of response B, regardless of what happens
before A or after B. The _experimenter_ may arrange for the response to
produce another stimulus C on the list of known stimulus-response pairs, to
produce behavior D. But that sequence control is being done by the
experimenter. To explain it, we would need to analyze the experimenter, not
the experimental subject.

I'm sure these cases happen in normal environments outside the lab, and
fairly commonly. My proposal will produce a given sequence so long as the
component control systems continue to produce the reference end-states
required to begin executing the next element in the chain. As I've said
before, it is not sequence _control_ but the proposed mechanism can
reasonably be expected to produce a given sequence so long as some not
unduly restrictive conditions are met.

I just don't think it's a tenable idea to
say that the organism adopts whichever strategy works for it. Once you've
committed to one organization, I think you've ruled out the other.

Well, that's one place where we strongly disagree.

Now, there's a convincing argument if I ever heard one!

No attempt was made to provide one. I just wanted to make you aware that I
have a different opinion on that score; therefore I do not view your
objection as problematic to my view.

Perhaps I did not make this sufficiently clear. In my proposal, the "cue"
serves as a _retrieval_ cue for associative memory; it retrieves the
parameters required to set up the next control process in the sequence
(e.g., the reference value).

Ah, I see. It does, somehow, whatever is required to make your idea work.
Asking what is required is not in the rules of the game the way you play
it. It is assumed that there is just one "next process in the sequence"
that will serve to achieve the higher-level goal under all circumstances.
And since you have said that, it must be true.

This is not my view and furthermore, I don't find sarcasm especially
illuminating. It says more about your state of mind than it does about my
position.

I was misled by your speaking as if you knew that reliable
sequence-production could occur open-loop. I now understand that you were
only conjecturing that it might be possible.

Yes.

doing something about it, and also showing that the elements of the
sequence are being generated independently.

Independently of what?

Of the environment. When the person repeats part of the sequence after the
target movement has stopped, this shows that there is a sequence-generator
inside the person.

If the cues arise as a result of completion of each behavioral act, then the
sequence would continue under the open-loop proposal as well.

Oh, that. What you're describing now is a closed loop system (an
oscillator) in which the _only_ source of the cue is the behavior itself.
The perception is produced by the action, and the action is the result of
comparing the perception with a reference perception (possibly zero). So
what's a "cue" that makes it different from a perception?

I explained above that it is simply a state of a perception that signals to
the system that the next control element can now begin.

We also know that from the fact that prior to the
target's stopping, the person is matching movements to the target movements
at a speed too great for simple tracking to work.

Please explain.

I move my finger up, right, down, up, left, down and so on in a continuing
pattern of quick moves, while you track it with your finger. If I go very
slowly, you don't need to recognize the pattern; you track just as if my
finger movements were random. As I speed up, however, there will come a
speed at which your ability to correct errors in that way starts to
deteriorate. At a slightly higher speed, assuming you recognize the
pattern, you will begin generating that pattern yourself, and adjusting it
so the amplitude, phase, and perhaps a few points along the way match what
I am doing.

In the slower regime, if I suddenly stop my finger, your finger will stop,
too, after a very short delay. In the faster regime, if I suddenly stop my
finger, your finger will execute one more and perhaps two or three more
elements of the up, left, down, up, right, down sequence. This shows that
sequence control is slower than position control.

Thanks. As I indicated in my prior post, learning to initiate the next
controled act in response to the appearance of the proper cue would yield
the same performance under that condition.

Sorry, I have no idea what that means. How is a reference perception
different from a comparison perception?

If a sequence of perceptions is being read out of memory, this may _supply_
the references for the lower-order systems. Alternatively, the sequence of
memory-perceptions may be compared in real time to the arriving sequence of
perceptions; the output of the perceptual function would be a series of
same-different judgements. The reference of this system might be for
"same," in which case a "different" perception would generate an immediate
error when it occurred.

Your model contains a sequence, but no sequence-perceiver. If you read out
of memory a set of elements one after another, you will get a set of
individual elements, one after another. What is needed is something that
will receive this set of elements as they occur and emit a signal if and
only if _one particular sequence is occurring_. That is what is required if
you are to explain how we can hear "Mary had a " in the right pitches, and
from that moment until the last stanza is sung perceive that _the same
sequence is in progress_. If, at any moment, you hear that something is
left out, changed, or inserted, that perceptual signal will change and will
cease, to some degree, to be that sequence.

My concept of sequence perception refers to that kind of phenomenon, which
differentiates between the _occurrance_ of a sequence (as an observer
perceives it) and the _perception_ of a sequence.

In my proposal the associative linkages pull up the appropriate memories as
the sequence unfolds (in reality or in memory). One can than make a point
by point comparison, noting where they are the same and where they are
different.

Again, I don't agree that the architecture has to be either/or. I
understand that this violates one of the HPCT design principles and that it
should not be accepted without solid empirical evidence. But it is a
testable proposition, and seems consistent with some of my own inner

experience.

The architecture determines what computations can be carried out, and that
depends on the type of neurons that are present and how connections can
grow and change among them. For comparators to exist, for example, we must
have neurons that receive both excitory signals and inhibitory signals from
the afferent and efferent sides of the brain, respectively (although which
is excitatory and which is inhibitory can differ at different levels).
These connections are there and well-identified. In the spinal cord we even
have both arrangements, interconnected, set up between antagonistic muscle
control systems. In the brain stem, the comparators are in a layer of the
output nuclei, at least in the Red Nucleus; the sensory feedback signals
are uniformly the excitatory ones and the reference signals the inhibitory
ones.

The physical arrangements and propensities, therefore, appear to be set up
to favor the development of negative feedback control systems; this
continues at least to the inner layer of the cerebral cortex, known to be
involved in motion or transition perception.

So?

Here is my final take on this matter, for now:

Suppose we have a control system with a reference signal of zero. How will
it behave when a sudden perceptual signal arises? The effect of the
perceptual signal will be to create an error signal, which passes through
the output function to become a change in lower-level reference signals. So
for a zero or constant reference signal, the control system will behave
exactly as if the change in perceptual signal were a "cue" and the change
in lower-level reference signal (or the effects thereof) were a "response"
to that "cue".

You will notice that _any_ control system at _any_ level will appear to be
a cued-response system if its reference signals remains constant and if its
perceptual signals change suddenly. So your proposal can be taken as a
subset of the control-system proposal when behavior is viewed under certain
conditions.

As to sequence control, sequence control as envisioned in HPCT requires not
just that sequences exist, but that a perceptual signal exist representing
the state of an ongoing sequence. A steady perceptual signal would indicate
that a particular sequence is in progress; that is required to explain how
we experience ongoing familiar sequences.

Your proposal does not address sequence control per se, but only the
appearance of producing particular reactions to particular stimuli. We can
explain those appearances in terms of disturbances of control systems with
constant reference signals. We do not need a different model for every
different situation; the whole point of modeling, from my point of view, is
to find a _single_ organization that will account for all the various
phenomena we observe.

This argument does not carry much weight with me. It's not a "different
model for every situation" I argue for, it's a single organization that
accounts for all of the phenomena. One of the phenomena, in my experience,
is that certain sequences seem to get executed open-loop, without
supervision or correction, unless and until the performance runs into
trouble, as when the next cue in the series fails to appear on schedule.
There's a lot to think about here and I don't pretend to have a serious,
complete proposal to offer. What we've seen from me thus far is essentially
me thinking aloud about the problem (with y'all listening in), and trying to
develop an approach that might work if developed far enough. I appreciate
the feedback.

Regards,

Bruce A.

[From Bill Powers (991111.0741 MDT)]

Bruce Abbott (991110.1430 EST)--

It was presented in B:CP to show how a sequence-perceiver might be built
that would produce a scalar output whose value was proportional to the
degree to which the input perceptions occurred in the specified sequence.
It was supposed to provide an existence proof that something capable of
recognizing a sequence and generating an output perceptual signal
proportional to the "degree of sequence" could be constructed from neural
elements.

Bill:
Gosh, you guys [Abbott and Marken] are pretty free about guessing my
intentions. I recommend that you read the book before trying to devine what
I meant.

On page 141, I say "Since these numbers [of sequence-recognizers needed for
a recognition vocabulary of 30000 words] seem a little outrageous, it may
be well to pause for a moment to consider how a word-recognizing input
function _might_ be organized, just as a check of feasibility, and not as a
guess as to how it is actually done."

That seems pretty plain to me.

The three-element recognizer proposed is designed to recognize the
appearance of three "phoneme" signals in the order j, oo, ss, for the
spoken word "juice". On page 143 I say "We will take a copy of the signal
from the third element and call this the perceptual signal emitted by this
word-recognizer, since the last signal will appear if, and only if, the
three phonemes occur in the order j, oo, ss."

On p. 145 I say, "We could make the operation even more realistic if we
added a summing neuron which received copies of all the recirculating
currents. As the sequence progressed, this sum signal would get larger and
larger as each storage element in turn went to the _on_ state, and for
incomplete sequences a person would have at least some perceptual signal."

Those two sentences are the total of what I said on that subject. I do not
discuss any design for a sequence controller in this chapter.

The point, I then remind the reader, is to show that having 30,000
sequence-recognizers is not at all an outrageous proposition, since in
principle each one could be made of just a handful of neurons. I say "If we
assume for discussion six phonemes per average word, this would mean that a
recognition vocabulary of 150,000 words would require a maximum of 900,000
neurons (better design would require far less), and this number of neurons
would occupy something like 150 cubic millimeters, or 1/6 cubic centimeter
of brain tissue."

I then comment that even a large amount of redundancy would not require an
unreasonable number of neurons.

Bruce:

In true control of the sequence, errors would be corrected _as they occur_
so that the intended sequence is produced each and every time, unless the
system is overwhelmed by disturbance.

If you would stop and think for a moment, you would realize that that is
impossible. As soon as the wrong sequence has occurred, it is too late to
make sure that only the right sequence is produced. The sequence "T E H" is
wrong (if you meant to type "the") as soon as the E occurs, and correcting
it only produces the sequence "T E backspace H E," which is also wrong. All
you can do is reset and start over, if you want to produce a pristine
sequence.

What is clear is that the sequence
perceiver sketched in B:CP cannot serve as the input function of a
traditional PCT-style control system (in which the error between perception
and reference gets amplified to drive the output that opposes the effect of
the disturbance) and do the job required of it for sequence control. The
error signal produced does not provide the information required by the
output function in order to do what is necessary to correct the error.

Judging by the amount of work you seem to have put in on constructing a
successful sequence control model, I would have to say that you don't know
what you're talking about. Of course an analog system would be unlikely to
work at this level -- if you thought it would, that's your mistake, not
mine. The canonical PCT diagram is a teaching tool, not a design for a brain.

This silly and acrimonious argument is unworthy of CSGnet.

Best,

Bill P.

[From Bill Powers (991111.0820 MDT)]

Bruce Abbott (991110.2110 EST)--

Bruce:

Is it _really_ necessary to label each paragraph with the name of the writer
(as above)? We got along fine without that device for years, now suddenly
it's become a fad. I think it lengthens posts unnecessarily.

Bill:
I was trying to suggest, subtly, that we do something to alleviate the
large numbers of misattributions of citations that are showing up in posts
lately. My way at least shows who I _thought_ wrote the material. My way
lengthens the posts by one line for every one to three paragraphs. It's a
nuisance, I agree. But I'm going to keep doing it.

Bruce:

Let me turn this around and ask how it works in sequence-control as you
envision it. How do the elements get stitched together so that the sequence
is followed, and what determines when one element in the control sequence
will be replaced by the next?

Bill:
Do it yourself. I haven't yet tried to "envision" sequence control in any
way I want to be held to.

Bruce:

I have nothing against the idea of sequence control. But this requires
advance knowledge of the sequence to be produced (to provide the means of
assessing that the proper sequence is in progress).

Bill:
Of course it does. What's wrong with that? Don't you know what song you're
about to sing, or what sentence you're about to say, or how you plan to get
the car out of the garage? Or are these things just a stimulus-driven
surprise to you every time they occur?

Bruce:

Yet it is possible to
produce -- and reproduce -- a given sequence without that knowledge, if the
perceptual state achieved at the end of one control process serves as the
retrieval cue that allows one to recall what control process should come
next, and with what reference.

Bill: Go ahead and design it and make it work. Then I'll believe you.

Bill:

If one driver who wishes to turn at Lincoln street approaches the
intersection with Washington street, why does not the cue "WASHINGTON ST"
causes him to turn, while a second driver who wishes to turn at Washington
Street does turn? Is the cue "WASHINGTON ST" not identical in both cases?
Is it something about the cue, or something about the driver that causes a
turn?

Bruce:

Both. If the driver is to turn when cued by WASHINGTON ST, then the driver
will turn whether it is appropriate to do so or not (given where the driver
intends to go). He or she will not turn when the sign reads MONROE ST. The
reason why WASHINGTON ST will cause him to turn, and not MONROE ST., is that
the driver has followed this sequence in the past and has now associated
seeing WASHINGTON ST. with turning left onto WASHINGTON ST. as a result of
previously having done so on several occasions while making his/her way to a
particular destination.

Bill:
So the driver's intentions never enter into it, is that what you're saying?
I hope you can find something interesting to do, standing there on the dock
watching the ship go over the horizon.

I just scanned down through the rest of the post. Geez. What a lot of
words. Clearly, you are never wrong and can explain at length why you are
not. That does not make for very interesting reading, so I guess I'll skip
it and just admit that you're right, whatever you said.

Best,

Bill P.

[From Rick Marken (991111.0840)]

Bill Powers (991111.0741 MDT) --

Gosh, you guys [Abbott and Marken] are pretty free about
guessing my intentions. I recommend that you read the book
before trying to devine what I meant.

Sorry. Which intention(s) did I too freely guess?

Bill Powers (991111.0839 MDT)--

The real question is this: Is it OK to threaten someone with
torture for saying things you don't want him to say? And actually
to torture people who go ahead and say what they want anyhow?
Whether you would call doing this "coercion" or "applesauce" is
not the point.

Right.

So Marc, Bruce G.; how do you guys feel about threatening people
with torture for saying things you don't want them to say and
actually torturing people who go ahead and say what they want
anyhow? Is whatever this is OK with you?

Best

Rick

···

--
Richard S. Marken Phone or Fax: 310 474-0313
Life Learning Associates mailto: rmarken@earthlink.net
http://home.earthlink.net/~rmarken

[From Bruce Gregory (991111.1211 EST)]

Rick Marken (991111.0840)

So Marc, Bruce G.; how do you guys feel about threatening people
with torture for saying things you don't want them to say and
actually torturing people who go ahead and say what they want
anyhow? Is whatever this is OK with you?

Don't tempt me...

Bruce Gregory

from [ Marc Abrams (991111.1353) ]

So Marc, Bruce G.; how do you guys feel about threatening people
with torture for saying things you don't want them to say and
actually torturing people who go ahead and say what they want
anyhow? Is whatever this is OK with you?

No it's not OK. But at what point does it become coercion? How can i ever
tell someone elses intentions? I think Rick is setting up artificial
categories that have little to do with a _scientific_ explanation of
coercion. Coercion is ( in my opinion ) is not a singular event, that
consists of singular variables and isolated control processes. I don't know
what the answer is. I don't feel comfortable with Rick's explanations and I
think they are only partially correct.

Marc

[From Rick Marken (991111.1230)]

Me:

So Marc, Bruce G.; how do you guys feel about threatening people
with torture for saying things you don't want them to say and
actually torturing people who go ahead and say what they want
anyhow? Is whatever this is OK with you?

Marc Abrams (991111.1353) --

No it's not OK.

That is -very_ reassuring.

But at what point does it become coercion?

It already is coercion. You just don't want to call it that
(for reasons that are still a mystery to me), so it will
never become "coercion" for you. But whatever it is, at least
you agree that it's a lousy way to deal with other people.

How can i ever tell someone elses intentions?

With people you can ask: do you want to recant? do you want
to be tortured? You can also use the test for the
controlled variable.

I don't feel comfortable with Rick's explanations and I
think they are only partially correct.

Sorry to be a source of discomfort. But I'm afraid that
your discomfort does not convince me that "giving a choice"
is non-coercive. What part of my explanations are incorrect
and why are they incorrect?

Best

Rick

···

--
Richard S. Marken Phone or Fax: 310 474-0313
Life Learning Associates mailto: rmarken@earthlink.net
http://home.earthlink.net/~rmarken

[Bruce Abbott (991111.1445 EST)]

Bill Powers (991111.0820 MDT)

I just scanned down through the rest of the post. Geez. What a lot of
words. Clearly, you are never wrong and can explain at length why you are
not. That does not make for very interesting reading, so I guess I'll skip
it and just admit that you're right, whatever you said.

O.K., let's just leave it at that. Sorry to disturb your slumber. Pleasant
dreams . . .

Regards,

Bruce A.