feedforward: prospects dimming

[From Bill Powers (931109.0830 MST)]

Martin Taylor (931108.1915) --

... the
_effective_ reference signal is no longer the direct standard
against which the perceptual signal is compared. The engineer,
standing off to one side, has to adjust this function and the
others in the circuit to achieve the desired relationship
between the reference signal and the OUTPUT.

Yes, that's true. But in a biological system, it would not be
an engineer who would make this adjustment. It would be
evolution ...

You miss my point. The L-H design is intended to produce a known
effector output, not a known perceptual signal. Feedback from the
perceptual signal is just a way of adding corrections to the
basic input-output feedforward chain. In the L-H design, the
perceptual signal doesn't even match the reference signal,
showing how unimportant they consider it.

Thus they can say "Two distinct control problems have been
identified: the servo or signal transmission problem and the
regulator or disturbance suppression problem." This amounts to
saying that we are concerned separately about the effector output
and about the error signals resulting from disturbances.

In PCT these are the same problem: if the perceptual signal is
controlled, that is all that the system needs to do. There is no
concern, with respect to any one control system, with producing a
given effector output, too. The only reason a given effector
output would be desired would be that it has OTHER perceptual
effects which are under control at the same time. And those can
be taken care of by another independent control system.

I realized this morning that the L-H design is a solution of a
problem that does not exist in PCT. In an organism, there can be
no control of objective effects of effector outputs, because
there is nobody who knows what those objective effects are. The
only variables that can be controlled are _subjective_ effects of
effector outputs: perceptions, or intrinsic variables. If those
effects are maintained near internally-specified levels, the
organism couldn't care less what its effector outputs are.

The L-H design results from the engineering approach which fails
to distinguish outputs from outcomes. Once it is recognized that
only outcomes matter, then the perceptual signal can be seen as a
direct representation of the outcome to be controlled, and the
output from the effectors can be left free to vary as the means
of accomplishing this end. There is no "transmission" problem
separate from the problem of countering disturbances.

As to "evolution" taking care of the adjustment problem, I object
to using evolution as a catch-all answer. In cases like this it
is no better than assigning responsibility to God's Will. If we
leave the achievement of organized behavior up to evolution we
might as well give up on modeling altogether. Far better to come
up with a model that doesn't need these adjustments. Any system
that depends on accurate calibrations to achieve its functions is
unlikely to be suitable as a model for a biological system. I
count the need for these adjustments as a heavy mark against the
L-H design. Or let me put it this way: a model that requires
accurate calibration to produce accurate control is far inferior
to a model that can produce accurate control without those
calibrations.

···

-----------------------------------------------------------
I think it's significant that the L-H model from 1955 dropped out
of circulation. I expect that what happened was that some other
engineer realized that it could be reduced to a classical model
simply by redefining functions, and thus introduced no new
considerations. Look at the top part of your diagram:

                > reference
                 V
             -------------------
            V V
          ----- -----
         > fcn | | fcn |
         > B() | | A() |
          ----- -----
            > >
            V V
      ---------- ------- -
     >comparator>-> |output |->|+|
     > > >fcn G()| -
      ---------- ------ |

The pathway through fcn A() adds to the output of the output
function fcn G(). This effect could be achieved by defining a new
function B'() such that the reference signal produced the same
added effect via the comparator and output function G(). So this
part of the diagram reduces to the canonical diagram, except for
the useless function in the reference path:

            > reference
            >
            >
            V
          -----
         > fcn |
         > B'()|
          -----
            >
            V
      ---------- -------
     >comparator>-> |output |->-
     > > >fcn G()| |
      ---------- ------ |
          >

Incidentally, your diagram omits an output function for the
direct connection of the reference signal to the CEV.
------------------------------------------------------------
I have an increasing (or is it sinking) feeling that we're up
against unshakeable tradition, here. Some people are just
insisting that the old input-output model is perfectly good, so
why do we need this new-fangled PCT model? No matter how
elaborate the old model has to be, no matter how much it relies
on calibration by some entity who knows the objective properties
of the environment, no matter how complex the mathematics gets,
no matter what problems have to be ignored, the old model is
still preferable.

The PCT model doesn't need special calibrations; all the output
pathways can vary their characteristics over a wide range without
materially affecting the controlled variable. Control is achieved
by an organism that has no knowledge of objective effects in the
external world. PCT redefines the nature of the problem we're
trying to solve: organisms don't produce specified behavioral
outputs at all, which is the problem the control engineers have
been trying to solve. They control a world of perception. The
engineers keep wanting to solve the other problem, which is not
the problem that organisms face.
--------------------------------------------------------------
We should, of course, continue to examine how feedforward might
play a part. Tom Bourbon has introduced a new consideration,
however, which makes the prospects of feedforward dimmer. We do
indeed have many perceptions which provide alternative ways of
controlling. With this in mind, it's perfectly clear that walking
in the dark is not open loop; one searches desperately for SOME
information about position. Walking in the dark is very different
from walking with the lights on. One walks more slowly, feeling
about with the hands, searching for an identifiable glimmer of
light. When no light exists, the mode of control changes
entirely; one controls for feel, not sight. If even feel is
missing, one doesn't control at all. At best, one can take some
steps under kinesthetic control. To view this transition merely
as a continuation of the old mode of control through feedforward
is to distort grossly what actually happens, and the distortion
is in just the direction needed to make feedforward seem still
adequate. We don't just stride on ahead when the lights go out.
We go into a drastically different mode of behavior. Let's
remember that we're trying to explain what actually happens, not
change what happens to make it fit the concept of feedforward.

Note to Rick Marken: you might try making a perceptual signal
depend on more lower-level signals than are needed to define it
unambiguously. Then control (of the perceptual signal) could
continue indefinitely when some lower-level signals are lost. The
actions might change, but the perception would still be
controlled.
---------------------------------------------------------------
Best,

Bill P.

[Martin Taylor 931109 15:30]
(Bill Powers 931109.0830)

Martin Taylor (931108.1915) --

... the
_effective_ reference signal is no longer the direct standard
against which the perceptual signal is compared. The engineer,
standing off to one side, has to adjust this function and the
others in the circuit to achieve the desired relationship
between the reference signal and the OUTPUT.

Yes, that's true. But in a biological system, it would not be
an engineer who would make this adjustment. It would be
evolution ...

You miss my point. The L-H design is intended to produce a known
effector output, not a known perceptual signal.

I didn't miss your point at all (unless I do so now as well). Maybe
you missed mine (and theirs). If the current perceptual signal and
reference are known, and IF the behaviour of the world is reasonably
stable AND the bandwidth (or expectation) of the disturbance is relatively
low, then the production of a predetermined effector output is likely to
bring the CEV close to the condition that corresponds to the reference
perception. It doesn't always work (consider the "flying suitcase"
situation, which, incidentally I had confirmed by a colleague as
something that literally happened to him), but in many circumstances
it works well enough to provide a first cut at bringing the perception
to match the reference.

Feedback from the perceptual signal is just a way of adding corrections
to the basic input-output feedforward chain. In the L-H design, the
perceptual signal doesn't even match the reference signal,
showing how unimportant they consider it.

Right. They hadn't heard of perceptual control theory. But the first
of those two sentences is a matter of emphasis. I would say that the
input-output element is a way of speeding the initial effect of a change
of reference, and it seems to work, according to the Lang-Ham simulations.

I realized this morning that the L-H design is a solution of a
problem that does not exist in PCT. In an organism, there can be
no control of objective effects of effector outputs, because
there is nobody who knows what those objective effects are. The
only variables that can be controlled are _subjective_ effects of
effector outputs: perceptions, or intrinsic variables. If those
effects are maintained near internally-specified levels, the
organism couldn't care less what its effector outputs are.

Lang and Ham would agree with you (as I do) about "are maintained,"
and (I suspect) about "can be controlled."

You accept that reorganization provides a set of output linkages
that lead to negative rather than positive feedback for all ECSs. You
accept that reorganization does not normally occur during the initial
stages of bringing a perceptual signal to match a new reference value.
It follows that the initial effector outputs (i.e. the signals to
any immediately lower-level ECSs) are going to be altered in a
predetermined way by a step change in the reference signal. The
output function is a function, not an on-line choice mechanism (at
least at the lower levels of the hierarchy--I have questions about
the program level).

Nobody KNOWS what the objective effects are of particular outputs. Right.
But reorganization has so arranged it that they are usually in the correct
direction (though they won't necessarily be if the environment is behaving
differently). Why should one not also believe that reorganization might
have so arranged it that the magnitude of the output is usually somewhere
near correct? There's no guarantee either about magnitude or about
direction. Both depend on the stability of the world, and both are subject
to perceptual control if the first move turns out to be wrong--think of
the experiment of reversing the handle-cursor movement relation.

Any system
that depends on accurate calibrations to achieve its functions is
unlikely to be suitable as a model for a biological system. I
count the need for these adjustments as a heavy mark against the
L-H design. Or let me put it this way: a model that requires
accurate calibration to produce accurate control is far inferior
to a model that can produce accurate control without those
calibrations.

If the A() function (the feedforward function) of the Lang-Ham
system is set to zero (poorly calibrated), and the B() function to
unity (equally poorly calibrated), the system works as well as the
standard PCT configuration. That's because it IS the standard PCT
configuration. So I wouldn't say that it depends on accurate calibration
to any greater extent than does the standard PCT configuration. If
setting the A() function to model (invert) the average behaviour of
the environment improves the speed of response to a change of reference,
then it would be better. But you can't lay a charge of "requiring
accurate calibration to work" on it.

···

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

I think it's significant that the L-H model from 1955 dropped out
of circulation.

Did it? Did it ever get into circulation, and if so, when did it
drop out? I stopped being professionally interested in control
engineering around 1956, so I don't know the answer to either
question. I'd be interested in knowing, though.

The pathway through fcn A() adds to the output of the output
function fcn G(). This effect could be achieved by defining a new
function B'() such that the reference signal produced the same
added effect via the comparator and output function G().

Lang-Ham Figure 4 shows this.

I have an increasing (or is it sinking) feeling that we're up
against unshakeable tradition, here. Some people are just
insisting that the old input-output model is perfectly good, so
why do we need this new-fangled PCT model?? No matter how
elaborate the old model has to be, no matter how much it relies
on calibration by some entity who knows the objective properties
of the environment, no matter how complex the mathematics gets,
no matter what problems have to be ignored, the old model is
still preferable.

I resent this slur. Sorry, but I do. It doesn't represent either
my thinking or the configuration we are supposed to be discussing.

It may represent the way Lang and Ham were thinking in 1954. As to
that I can't say. But they were not exposed to PCT and could not
have posed the question "why do we need this new-fangled PCT model?"

If you followed the equations that I hope I did right, You can see
that (a) Both the Lang-Ham and your B=1 modification bring the
perceptual signal to the reference level, but that your modification
improves the long-term accuracy by roughly the factor G (output gain).
(b) the dynamic behaviour after the initial effector output due to the
reference step is that of the PCT system.

The PCT model doesn't need special calibrations; all the output
pathways can vary their characteristics over a wide range without
materially affecting the controlled variable. Control is achieved
by an organism that has no knowledge of objective effects in the
external world. PCT redefines the nature of the problem we're
trying to solve: organisms don't produce specified behavioral
outputs at all, which is the problem the control engineers have
been trying to solve. They control a world of perception. The
engineers keep wanting to solve the other problem, which is not
the problem that organisms face.

The truth of this paragraph is no reason to go out of your way to
avoid thinking about modifications to the standard configuration.

We should, of course, continue to examine how feedforward might
play a part. Tom Bourbon has introduced a new consideration,
however, which makes the prospects of feedforward dimmer. We do
indeed have many perceptions which provide alternative ways of
controlling.

We use whatever is available, to the extent that it is currently
available. We don't stop walking or go by feel when someone turns
the lights out, at least not for a few paces. Of course we do use
kinaesthetic sensation to judge where we are; we even need it simply to
stand. I don't think anyone was claiming we didn't (Hans said I
read him correctly on this).

To view this transition merely
as a continuation of the old mode of control through feedforward
is to distort grossly what actually happens,

Who views it so?

and the distortion
is in just the direction needed to make feedforward seem still
adequate. We don't just stride on ahead when the lights go out.

Not for more than a few paces, unless we are on a flat, open floor.
Why not? For all the reasons you adduce, and to which Hans has
pointed--integrators lose accuracy and so forth, so that the model
room no longer relates to one's perceived (in imagination) position.

Let's go a little further. I'll make the straight out claim that there
is NO benefit in perceptual control IF the result of output is guaranteed.
(No claim about how often this situation occurs in the world. We probably
wouldn't know if there are any actions we perform of this kind.) In human
factors work this can be important. You don't want to force people to
look at stuff that they should expect to know already.

For example, if a speech recognizer makes a high proportion of errors
(say 5%), the user wants to know what has been recognized. But if the
recognizer is very accurate (say 2% or less word error), users often
ignore word-by-word feedback and are surprised when the results of their
commands are not as intended. It is better not to provide the perceptual
feedback at the word level, but to provide it at a level where failure
is more probable and more costly.

So the above claim can be softened and extended: there is little benefit
in perceptual control if the result of an output is highly probable to
put the CEV into the desired state.

Perceptual control is needed for any behaviour in an unstable world, but
not all of the world is unstable on the time-scale over which control
might be exercised.

Martin

[From Rick Marken (931109.1330)]

Bill Powers (931109.0830 MST)--

I have an increasing (or is it sinking) feeling that we're up
against unshakeable tradition, here. Some people are just
insisting that the old input-output model is perfectly good, so
why do we need this new-fangled PCT model? No matter how
elaborate the old model has to be, no matter how much it relies
on calibration by some entity who knows the objective properties
of the environment, no matter how complex the mathematics gets,
no matter what problems have to be ignored, the old model is
still preferable.

I agree with this, of course, but I would point out that defense
of the input-output model is not confined to those who question
the need for PCT. It seems to be possible to see the need for
"this new-fangled PCT model" of behavior and at the SAME TIME see
the need for the very model of behavior that PCT denies. It's a
peculiar phenomenon -- rather like seeing the need for a sun-centered
model of planetary motion and at the same time seeing the need for the
Ptolmaic one.

I think that there are many people in this boat -- people who
don't necessarily think that the input-output model is preferable
to PCT; just that the input-output model has its place in our
explanations of behavior (possibly a very BIG place). I think
that a lot of this just has to do with a reverence for historical
authority; if the input-output model has been accepted for many
of years by famous smart people then there must be SOMETHING right
about it. I think many of us control for this "respect for historical
authority" along with (or sometimes instead of) scientific scepticism.
This is why all the empirical evidence that you cite (over-complexity,
reliance on external calibration, factual problems, etc) count for
nought. I don't think there is any way to overcome this resistence
except possibly by reassuring people that they are grown-ups now and
can evaluate things on their own now.

Parentally

Rick