loss of feedback; cherry pie

[From Bill Powers (950605.1415 MDT)]

Martin Taylor (950605.1100)--

     "But if the PIF works only on the current values of the input
     signal, then when the input values change dramatically (such as by
     vanishing), the perceptual signal will change dramatically,
     creating probably large error values, and ineffective but large
     changes in output."

     Is that a premise or a consequence of talking about the "standard
     model," the form of which is shown in the CSG logo?

There is nothing in the "standard model" that defines the input function
to have only a single input. The logo is only a symbolic representation
of the typical control system. In reality, both the input and output
lines should be multiple, with many paths through the environment (as in
the figure in my 1973 Science paper or the cover of LSC I). So yes,
assuming that cutting off visual inputs cuts off all inputs used by a
given perceptual function is an assumption.

     >The "use only the current value of the inputs" kind of PIF is
     >tuned for all possible unpredictability. It results in immediate
     >loss of control when the input is cut off.

If this is a quote of something I wrote, as the single greater-than
signs suggest, I must have been under anaesthesia when I wrote it,
because I don't understand it at all. What does "tuned for all possible
unpredictability" mean?

It's true that if you cut off all lower-level signals of which a
controlled perception is a function, there will be immediate loss of
control. But in suggesting that perceived cursor position might be a
function of kinesthetic inputs as well as visual ones, I was proposing a
way in which loss of _visual_ inputs would not amount to total loss of
all signals on which the perceptual signal standing for cursor position
depends: you would have to lose the kinesthetic inputs, too, to produce
total loss of control. I said, trying to make this clear,

In the case of visual tracking, for example, we tend to assume that the
sense of cursor position is derived entirely from visual information.
But we also know that while we are tracking, we can feel our arms and
hands moving. What kind of input function could use both the visual and
the kinesthetic feedback information to produce a cursor position
signal?

You reply:

     This refers NOT to a case in which the input is cut off, but to a
     case in which the input comes from a variety of lower-level
     perceptions, some of which no longer have access to the real world
     data (no telling whether they use prediction, but assume they
     don't), and some of which continue to have access.

I think the problem here is that you've got into the habit of thinking
of the standard control system in a certain special way, so when I refer
to a somewhat more general case you don't recognize it as an example of
the standard model. You sometimes write as if you think that for every
perceptual signal there is some real CEV in the environment that
corresponds to it. When you refer to "the" input being cut off, you seem
to mean all contributing signals wherever they come from, which would be
a pretty unusual circumstance under my proposal. And it would probably
result in total loss of control.

In talking with Hans Blom I've been playing his game by his rules. In
offering some ideas about how control might continue even if one
component of the input data is cut off, I was raising a question about a
possible alternative to his model that does not require either
prediction or a world-model to explain what we observe in one specific
example of real behavior. I'm only _suggesting_ this alternative,
because I haven't tried out a model of it.

     Of course I agree with the statements in your message, about the
     ability to use multiple sources of input. But they are irrelevant
     to the question that was at issue: what happens when the input is
     cut off, with a "predicting" PIF or a "use current values only"
     PIF.

If the currently-perceived cursor position is a function of current
visual inputs and current kinesthetic inputs, then the PIF belongs to
the "use current values only" type of input function. I don't see what
your problem is, unless unconsciously you were thinking of a "use
current values FROM ONE SENSORY MODALITY only" PIF.

     My point was that there is a conceptual distinction between "having
     a value of zero" and "having an unspecified value," and that the
     "standard model" control loop has no way for that distinction to be
     represented, either in the PIF or elsewhere.

That's true, and neither does Hans' model. When the input is lost in
Hans' model, the expected variance pvv of the perceptual variable y has
to be set extremely large (10,000), to stop the Kalman filter from
continuing to make adjustments based on a spurious value of y. This is
the equivalent of an external intelligence that realizes that whatever
value y may now have, it no longer represents an environmental variable
and must not be allowed to participate in computations. If you supplied
a standard PCT model with help from a similar external intelligence,
there would be no problem there, either. In fact, in Hans' model, when
the system is instructed to "go blind," not only is the expected
variance of pvv set to a large number, but y itself is arbitrarily set
to 10.0, the average value of xopt. So I suspect that the distinction
you make, which is a valid one, caused problems for the adaptive model,
too.

In fact, I believe, any perceptual input function will produce an output
signal in the absence of all input, and that signal will have a meaning
in terms of the perception (even if the output signal is zero). The only
way to prevent unwanted actions from ensuing would be for a higher
system to disable the control system entirely, in any one of several
ways that have been suggested for doing so. This is as true for the
Kalman filter adaptive model as for the PCT model.

     A one degree-of-freedom system in which the single degree of
     freedom is dedicated to the value of a variable cannot
     simultaneously represent the validity of that variable.

I agree entirely. Something must continually monitor the relationship
between the signal and other signals at the input of the PIF, judging
not the magnitude of the signal but its validity, and taking appropriate
steps to shut down the control process (or the adaptive process).

···

-----------------------------------------
Your reply to Bill Leach's comment on choosing between unwanted
alternatives used an inappropriate example:

     And given free control, I would drive straight from work to home,
     if my control systems were not overwhelmed by the difficulty of
     driving over kerbs and through houses that intervene. So I control
     what I can, and arrive home by a more circuitous route.

Bill's comment was that _neither_ choice presented to the rats was one
it would select by itself. A more appropriate example would be, "Would
you rather be able to blow your car's horn or to have it silent while we
shove the car over a cliff with you in it?" Certainly you have a right
to study the effects of that choice if you want to, but it would be
pretty silly.
-----------------------------------------------------------------------
Best,

Bill P.

[Martin Taylor 950606 11:00]

Bill Powers (950605.1415 MDT)

I debated whether to respond to Bill's message, on the "assumption" that
he would see from rereading my original that his criticisms were unfounded,
and that he more or less rewords what I wrote in agreement with his earlier
posting. So why respond? Because I have found in the past that
misunderstandings left silently alone tend to be taken as understandings,
which leads to later problems.

No, I do NOT think that there is one single input to a PIF. What good
would such a control system be, other than at the very periphery? It would
simply be controlling a perception already influenced by the output of the
control system whose perceptual signal WAS that input. A pointless
connection, I would think, UNLESS the higher PIF involved some time-binding
operator such as a derivative or an integral. Higher PIFs ordinarily
accept as inputs the perceptual signals from multiple lower level ECUs.

But in suggesting that perceived cursor position might be a
function of kinesthetic inputs as well as visual ones, I was proposing a
way in which loss of _visual_ inputs would not amount to total loss of
all signals on which the perceptual signal standing for cursor position
depends: you would have to lose the kinesthetic inputs, too, to produce
total loss of control. I said, trying to make this clear,

, preaching to the choir.

    >The "use only the current value of the inputs" kind of PIF is
    >tuned for all possible unpredictability. It results in immediate
    >loss of control when the input is cut off.

If this is a quote of something I wrote, as the single greater-than
signs suggest, I must have been under anaesthesia when I wrote it,
because I don't understand it at all. What does "tuned for all possible
unpredictability" mean?

You quote using tabs. I quote using ">". So I quoted your copy of my
writing. So it behooves me to explain.

If the PIF uses ONLY the current value of its inputs, then it can and will
alter its output as fast as the input values change, without limit. The
input values are completely unpredicted, and the PIF is tuned for all
possible unpredictability. If, on the other hand, the PIF is, say, a
narrow-band filter, its output can change only at a rate inversely
proportional to its bandwidth (I'm not talking, of course, about the
oscillatory component of its output, which changes at the rate corresponding
to the filter's centre frequency--to head off a misunderstanding that just
might have been possible). Such a PIF is tuned for a small range of
unpredictability, namely that which affects the component of the input
that is within the filter's band-shape.

You sometimes write as if you think that for every
perceptual signal there is some real CEV in the environment that
corresponds to it.

Should we get out all the postings from the archive in which this was all
thrashed out? To head that off, perhaps I should once again list some of
my "assumptions," with which I think you have (had) no quarrel.

1. There is some reality through which the outputs affect the perceptual
signals.

2. There is no way for the organism to determine ANYTHING about the
reality except through the perceptual signals.

3. Perceptual functions are functions of inputs, at least some of which
are influenced by the external reality.

4. Each Perceptual Function (PIF, the "I" implying that the function is
the input component of a control system) defines an aspect of the external
reality.

5. The aspect of reality defined by the PIF is unknowable to any other
organism, but from the viewpoint of the defining control system, it "exists"
and we call it the Complex Environmental Variable defined by the PIF.

6. Output from the control unit controlling the perceptual signal from a PIF
affects something in reality, but only the parts of reality incorporated
in the CEV are involved in control. All other effects of output are
"side effects".

7. "Side effects" are totally irrelevant to the Elementary Control Unit
(ECU) that generates them, but they may well influence the CEVs of other
ECUs, either in the same organism or in another ("observer") organism.

8. All signal paths between the output of an ECU and the PIF may branch
indefinitely, as seen by an outside analyst, but those that are part
of the control loop (i.e. not "side effects") converge on the CEV and
diverge from it (an "observer" cannot see the CEV, but an "analyst" can
impute its existence.

9. There is no way for the ECU to determine the branching pattern of the
various signal paths, since all that happens is that there is a (one, scalar,
singular) perceptual signal and an (one, scalar, singular) output signal,
somehow related. All effects of one on the other could be produced equally
well by one possibly complicated signal path or by many parallel signal
paths.

When you refer to "the" input being cut off, you seem
to mean all contributing signals wherever they come from, which would be
a pretty unusual circumstance under my proposal. And it would probably
result in total loss of control.

We are agreed on that. The question was how fast would control be lost
with different kinds of PIF. Without being able to run Hans' model, I
have not tried to hang any arguments on it. Rather, perhaps ill-advisedly
because of the apparent misunderstandings, I have tried to use his verbal
discussions as a peg on which to hang the simple question.

In fact, I believe, any perceptual input function will produce an output
signal in the absence of all input, and that signal will have a meaning
in terms of the perception (even if the output signal is zero).

We are agreed on that, too.

The only
way to prevent unwanted actions from ensuing would be for a higher
system to disable the control system entirely, in any one of several
ways that have been suggested for doing so.

It is probably most effective to do it that way, but it may not be the only
way. If the ECU is made less "E"lementary, and provided with a second
degree of freedom by way of a "perceptual validity" signal, it could
shut itself off. How such a signal could be produced, if indeed it could
at all, is another question entirely.

Your reply to Bill Leach's comment on choosing between unwanted
alternatives used an inappropriate example:

    And given free control, I would drive straight from work to home,
    if my control systems were not overwhelmed by the difficulty of
    driving over kerbs and through houses that intervene. So I control
    what I can, and arrive home by a more circuitous route.

Bill's comment was that _neither_ choice presented to the rats was one
it would select by itself.

All right. I'll give an example that faces me this Thursday. In Canada
we have a riding-based election system. The person with most votes in a
riding gets to be the member for that riding. You vote in the riding in
which you live. There are three major parties, each of which has formed
the Government of Ontario over the last decade. I greatly prefer party A,
but in my riding, the candidate for party A has never had the slightest
hope of winning. So I can't effectively vote for party A, and must choose
between two alternatives, neither of which would be the one I would select
by myself.

This seems to me to be an exact analogy with the rat. In another riding
(not constrained by the cage) I could choose a candidate that would make
me happy, but in my riding (forced to be exposed to shock) I must choose
the less bad of two alternatives.

A more appropriate example would be, "Would
you rather be able to blow your car's horn or to have it silent while we
shove the car over a cliff with you in it?" Certainly you have a right
to study the effects of that choice if you want to, but it would be
pretty silly.

That's a bit extreme, but not unlike what I feel my voting choices are.
It is not unusual to be in a position in which you can't choose what
in other circumstances you would, and in which you must choose between
distasteful remaining possibilities.

Thank goodness even our more rabid right wingers are more like US Democrats
than Republicans--they're bad enough, but not nearly as frightening.

Martin

<[Bill Leach 950607.01:31 U.S. Eastern Time Zone]

[Martin Taylor 950606 11:00]

I debated whether to respond to Bill's message, on the "assumption" that
he would see from rereading my original that his criticisms were
unfounded,

Though it has probably happened, I am not sure that I recall you ever
saying that one (or more) of Bill's criticisms had valid basis.

After having been on this list for greater than a year I conclude that
you have a very high reference for not perceiving yourself to have made
an error. One of the highest in fact that I think I have ever observed.

Have you ever had a student "prove" you wrong in any point? If so, did
you admit to that student that you had been in error or did you just
quietly accept the new information and go on as though that is how you
had viewed the matter all along?

I admit that I don't thing that ANYONE _likes_ to be "wrong".
I especially don't think that anyone likes to discover that they have
been shown to be "wrong" in public. I also don't know of anyone that has
not been both from time to time.

Even the immortal in his own time, Bill Powers, has made mistakes...
graciously admitted them when pointed out (and many times admitted them
even when no one did point them out) and continued on (providing no end
of pleasure - at the demonstration that he too is human - to at least a
few of us that tend to make a great deal more mistakes than does he).

Inspite of this rather blunt posting, I do admit that you are, in my
opinion of course, more often misunderstood than anyone else here.

-bill