Blindt1a; 78K; Seat of pants

[From Bill Powers (950510.0930 MDT)]

Bruce Abbott (950509.1205 EST) --

Blindt1a is a nice development.

     My (brief) experience with the program agrees with yours; I find
     myself attending to proprioceptive arm signals and peripheral
     visual input of the mouse position as a substitute for the cursor
     when the cursor is absent. I attempt to produce a pattern of mouse
     movement (as sensed via these signals) that matches some internal
     representation I have of the way in which the cursor moves when it
      is visible. It's still control, but when deprived of cursor
     position input, I have to find substitutes.

Yes, that checks. I can feel myself switching back and forth between
visual tracking and kinesthetic tracking, in the "mixed" mode. At no
time does it seem that I am imagining a cursor and trying to control it.
But it does seem that I am using other real-time perceptions. Such
subjective impressions are subject to bias, of course, but I don't think
that a world-model of the cursor movements exists at least in a form
available to my awareness. Rick Marken, what's your experience? Hans
Blom, are you in a position to try this? Is anybody else working along
with us? Lurkers?

       I attempt to produce a pattern of mouse movement (as sensed via
     these signals) that matches some internal representation I have of
     the way in which the cursor moves when it is visible. It's still
     control, but when deprived of cursor position input, I have to find
     substitutes.

There's a difference -- could you check that again? I don't imagine a
cursor at all. I'm just trying to make the felt mouse move the way the
seen target moves. but maybe you really do imagine a cursor. I'll try
it.

     Because the motion of the cursor cannot be derived from my
     proprioceptive and mouse-visual inputs, the source of the
     references for the mouse positioning/velocity/acceleration system
     must come from imagination or habit, whatever those terms mean.

Even to use kinesthetic or visual information about mouse/hand
movements, it's necessary to come up with a scaling factor. This can be
obtained only through experience. When we track with the cursor visible,
the mouse necessarily goes through the movements that keep the cursor on
the target (because it does stay on the target) and we can sense the
mouse positions. Wayne Hershberger pointed out some time ago that in
this situation, where we can sense the action that is accomplishing
control, we are in effect sensing disturbances, too. There are no
disturbances in this case (except irregularities in our own movements
and in friction against the mouse pad), but we are still able to sense
our own action through kinesthetic and visual channels. So it's possible
that some system is busily constructing a proportionality factor which
says that a given amount of kinesthetic/visual signal is equivalent to a
given amount of cursor-position signal. It isn't very accurate, but it's
better than nothing. Relationship control?

This implies to me that the perceptions of visual space and of
kinesthetic space are being mapped into a common perceptual space. If
this were not true, setting up the equivalence between mouse movement as
sensed kinesthetically and the cursor position as sensed visually would
involve a very complex ad-hoc 3-D rotation-translation matrix that would
have to change with each change in mouse position (from one experiment
to the next). I think there's reason to believe (if not know) that all
exteroceptive perceptions are mapped, at a low level, into one common
"laboratory space" which we maintain as an "objective" frame of
reference independent of bodily position. It has to be updated
frequently using real-time perceptions, but while the calibration lasts
it allows us to control spatial relationships in a relatively simple
way, using various sensory modalities like vision, kinesthesia, touch,
and sound.

Incidentally, I seem to recall that people studying how blind people
make their way around in the world have found that sound reflection
plays a part even when there is no deliberate echo-ranging using a
stick. Sighted people paying attention to sound can do this, too. When
you stand closely facing a wall, your own breathing sounds louder.

Also, infra-red heat sensitivity in the face can give crude indications,
such as an indication of being next to a door opening or a window rather
than a wall. I have noticed the infra-red effect in the dark. It also
shows up when I stand near a stove on which a heating element has been
left on "low" -- not hot enough to see, but hot enough to feel on the
hands and face. And there is a crude directionality. I'm quite sure I
could find a warm heating element in a pitch-black room.

I just thought of a real-world example of Blindt1a. When you're driving,
you don't watch the road every second. While you're looking at a tape
cassette or putting it into the slot, or trying to read a map spread out
on the steering wheel, you control the position of the wheel
kinesthetically (hoping that there are no disturbances of the car). At
least you try to hold the wheel in one position (I doubt that anybody
still alive would try this on a curving two-lane road in traffic).

     When cursor input is lost only briefly, I may be able to handle the
     job through preservation [perseveration?]: simply continuing to do
     what I was doing when the input was lost: moving in a given
     direction, at a given velocity, and perhaps accelerating at some
     rate. How do I accomplish this?

One nice thing about pure integral-output control is that if the error
goes to zero, the integral in the output function holds its value at
least for a while. To that, add that in a neural one-way control system
with an excitatory reference signal, a reference level of zero assures
that the error signal will remain at zero (no amount of inhibitory
perceptual signal can make the error signal go negative). If the output
of the control system is specifying an acceleration or a velocity for a
lower-level system, that acceleration or velocity will remain constant
(or decay only slowly) until the higher reference signal is made non-
zero again. Of course you need a pair of systems to handle two-way
control, but the nice thing is that if BOTH reference signals are set to
zero, BOTH outputs hold their values. So maybe if the immediate response
to loss of signal is to set the immediately lower level reference
signals to zero, you would see the result you describe. A nice simple
explanation, if true.

In a way, such features amount to a kind of model of the environment,
but it's not as literal as the world-model Hans proposes.

I'm getting a feeling that if we could just keep on exploring all the
possibilities this way, doing experiments along the way to make sure the
effects we guess at really happen, we could actually start sorting out
the way the control systems are really hooked up. But we need a lot more
people thinking and experimenting.

···

-----------------------------------------------------------------------
Ron Blue (950505) --

Mary and I would like to get a copy of your "78K model". I hope it's in
C or Turbo Pascal.

Powers_w@fortlewis.edu

Thanks.
-----------------------------------------------------------------------
Bill Leach (950509.18:34 U.S. Eastern Time Zone)--

     I really believe that stumbling upon a missplaced stair is
     specifically an example of a "failure" with such a system in
     operation. Obviously (again I think), the "model" system is only a
     "minor" part of the whole control loop handling the goal of
     perceiving oneself at the top of the stairs. Do you agree that
     this is possibly true?

Yes I do. You and I seem to have similar "seat-of-the-pants" approaches
to control engineering.
-----------------------------------------------------------------------