Much ado about feedforward

CSGnet Archive CSG1993
1

[From Rick Marken (931108.2100)]

I am becoming more and more convinced that feedforward solves
a problem that doesn't exist; at least, a problem that
doesn't exist for the hierarchical control model of behavior
developed by William T. Powers. Tom Bourbon (931108.1005)
clinched it for me when he said:

Biological systems (such as we) probably can accommodate the loss
of perceptual signals in one of the several sensory systems that contribute
to a higher-level perceptual function by simply continuing to control the
signal emerging from that function. (See a simplified model below)

I tested Tom's idea this evening using my three level spreadsheet
hierarchy (I had to change it a bit to get the actual states of
the controlled variables when the systems went "open loop').
I put individual systems into open loop mode by turning on the
"imagination" connection for the system. This takes perception out
of the loop and pegs the output at its current level; so the control
system is blindly generating a constant output. The result of taking
out one or two lower level systems depends on a number of things.
If the disturbance variations are slow, there is surprisingly
little effect on higher order control systems if two or even
three low level systems go "open loop". This lack of effect on higher
level control lasts for quite some time, but eventually, of course,
higher level control is lost. The lower level systems that are
"open loop" lose control immediately, of course. But the model can
still control a level three variable (a logical relationship) for some
time after it has lost control of some of the lower level components of
the perceptual variable (as Tom had suggested).

If the "open loop" control systems are themselves higher in the
hierarchy (level 2, say) then it takes some time for these
systems to lose control -- and even longer for the higher level
systems that depend on them to lose control.

The lessons from this little exercise are 1) Tom Bourbon is right
nearly as often as I am and 2) a hierarchy of negative feedback
control systems can survive fairly long term losses of lower level
perceptions (even if there ARE disturbances -- if there are NO
disturbances it can survive indefinitely long losses of lower
level perceptions).

So the perceptual control hierarchy seems to act just like a person
walking across the room in the dark. The higher order variable
(the path across the room) can be controlled for some time despite loss
of the ability to perceive lower level controlled variables (visual
variables like distances, configurations and transitions that
could be controlled while carrying out the "crossing the room" event). The
behavior of the control hierarchy looks like it must be a result
of feedforward processes (to an outside observer who notices
the loss of visual feedback, say) but it's still just good old
hierarchical control of perception -- the perception that is still
available, that is.

Best

Rick

2

[Martin Taylor 931109 11:00]
(Rick Marken 931108.2100)

I tested Tom's idea this evening using my three level spreadsheet
hierarchy (I had to change it a bit to get the actual states of
the controlled variables when the systems went "open loop').
I put individual systems into open loop mode by turning on the
"imagination" connection for the system. This takes perception out
of the loop and pegs the output at its current level; so the control
system is blindly generating a constant output. The result of taking
out one or two lower level systems depends on a number of things.
If the disturbance variations are slow, there is surprisingly
little effect on higher order control systems if two or even
three low level systems go "open loop". This lack of effect on higher
level control lasts for quite some time, but eventually, of course,
higher level control is lost.

Rick, dies this "quite some time" allow for the very large slowing factors
that you use in the spreadsheet model? In other words, are you asserting
that the high-level ECSs could have lost control much more rapidly than
they did, given the parameters that were set in the spreadsheet?

As I understand the spreadsheet model, each perceptual signal is a defined
fuinction of all its input variables. If some of the input variables
don't correspond to "reality", the perceptual signal ought to be giving
the wrong answer, and the error ought to be causing the wrong output to
be generated. This is different from the situation Tom was describing,
in which one information source can be used when another, that usually
provides better information about the same CEV, has been disrupted.

Am I misunderstanding what the spreadsheet does?

Martin

3

From Tom Bourbon [931109.1125]

[Martin Taylor 931109 11:00]
(Rick Marken 931108.2100)

I tested Tom's idea this evening using my three level spreadsheet
hierarchy (I had to change it a bit to get the actual states of
the controlled variables when the systems went "open loop').
I put individual systems into open loop mode by turning on the
"imagination" connection for the system. This takes perception out
of the loop and pegs the output at its current level; so the control
system is blindly generating a constant output. The result of taking
out one or two lower level systems depends on a number of things.
If the disturbance variations are slow, there is surprisingly
little effect on higher order control systems if two or even
three low level systems go "open loop". This lack of effect on higher
level control lasts for quite some time, but eventually, of course,
higher level control is lost.

Rick, dies this "quite some time" allow for the very large slowing factors
that you use in the spreadsheet model? In other words, are you asserting
that the high-level ECSs could have lost control much more rapidly than
they did, given the parameters that were set in the spreadsheet?

Wouldn't that always be the case, in any hierarchical control system? Are
you imagining a significance that you did not state, Martin?

As I understand the spreadsheet model, each perceptual signal is a defined
fuinction of all its input variables. If some of the input variables
don't correspond to "reality", the perceptual signal ought to be giving
the wrong answer, and the error ought to be causing the wrong output to
be generated. This is different from the situation Tom was describing,
in which one information source can be used when another, that usually
provides better information about the same CEV, has been disrupted.

Martin, my reading of Rick's simulation has him using higher-order
perceptual functions in the way I imagined. Those functions do not give a
"wrong" output, just an output. Imagine an "angle-detector" cell in the
visual cortex. It gives virtually the same output across many
transformations of the proximal stimulus pattern on the retina. Under many
cases when the retinal pattern is other than the maximally effective one,
the cortical cell shows a near-maximal response. That is not a "wrong"
response.

In his simulations, Rick found the upper level in the hierarchical-parallel
model continued to control, using the available lower level loops. He found
what I suspected.

Am I misunderstanding what the spreadsheet does?

Or am I, Rick?

[Martin Taylor 931108 19:40]
(Tom Bourbon 931108.1005)

Martin:

They (and I) were not talking about *modelling* control, but about
*controlling*.

Tom:>

Fine. But the discussion initiated by Hans was about control by organisms
(including Hans). So was my reply. You say Lang and Ham were talking about
designing optimal, engineered control systems. That is often a different
subject, as it was in this instance.

Martin:

I thought Hans was talking about how humans might be controlling. Humans
having evolved over some considerable time as control systems, it occurred
to me that Nature might have discovered good systems, if humans could find
them.

Tom:
I would be the last to argue that Nature has not "discovered" all of the
tricks employed by control engineers. I even think engineers are part of
Nature, along with the rest of our species. But I intend to go as far as I
can, given the limits on my abilities, to determine the extent to which
Nature can be modeled as though it had opted for the simple, rather than
the ingenious. (But I DO think a simple PCT loop, in hierarchical-parallel
systems, is pretty ingenious!) I want to see how far we can go using as our
model the simplest PCT loop, then that loop in a hierarchical-parallel
system. The fact that our species has devised other ingenious processes is
interesting, but most of the instances presented as evidence that those
processes are at work in living control systems seem to be of the "if we
build them, they are us" variety.

You wouldn't see the feedforward in your tracking studies (much), because
the feedback does a good job. But try the sawtooth tracking studies with
irregular blanking of the target, and see what happens. Will the feedback
only models still account for 99% of the variance? (This might be worth
trying in the sleep experiment).

I've tried the triangular tracking studies with the screen going blank after
the initial run-in period. For one minute, I "track" blind, with no
disturbance. I have tracked that triangular path many hundreds of times,
during more than nine years. In a recent bout in which I performed the task
sixteen times in succession, each time blanking the screen after the run-in
period, the mean correlation between my achieved handle positions (1800
of them during each run) and the ideal ones was -0.003 (S.D. = .118, range =
.390 to -.223). It is not easy to _quantitatively_ reproduce the same
pattern of movements for even one minute. _Qualitatively_, I always did
"almost the same thing." But as Bill Powers has pointed out during this
round of discussion, we must look very closely at all alleged examples of
actions that repeat identically, as in the allegedly ballistic actions that
some people want to call examples of feedforward. For the 16 runs I just
summarized, *any* incorrectly timed reversal of the direction of my hand
movements led to disastrous consequences quantitatively, but not so,
qualitatively. I suspect that the "sort of correct" patterns I created
without vision are a lot like the "sort of the same" movements produced by
"deafferented" animals -- and like people walking trough darkened, but
familiar, rooms.

Tom:

Yes, Avery, your option (b) was what I had in mind.

Martin:

No problems here. My base assumption is always that one uses whatever
information one can. That includes all kinds of sensory input as well
as remembered and deduced information about "how the world works." There's
no incompatibility.

Then do you still think the higher level perceptual functions in a
hierarchical-parallel control system (as in hierarchical-parallel perceptual
control theory -- HPPCT :-)) give the wrong signal when some of their
contributing lower-level perceptual signals are missing? Aren't perceptual
signals "about" their accompanying perceptual functions, rather than about
external "reality?"

Tom:

Much of the current
round of discussion about feedback-feedforward seems to hinge on the idea
that something special must happen that (at least some of the time) allows a
person to walk successfully through a familiar but now darkened room.

Martin:

No. It is that the same things that allow it are happening normally.

That is what I said, and what Rick simulated.

Martin:

. . . One
uses the best information that is at hand. If you have direct perception of
the current situation, that's usually the best you can get. But it isn't
always.

Direct perception of the current situation. I'm not sure what that is. If
you mean simply, "current perceptions," I agree that's *always* the best you
can get.

Tom:

The revised Lang-Ham model *might* be *useful* under some conditions, but I
don't think it will prove *necessary* when *part of* the feedback signal is
cut off.

Martin:

This is a direct cue to get back into the "information in perception"
discussion. I don't want that tar-baby in my hands right now. But it
will come.

I don't know what you mean. Is it that you see a cue there, even though I
had no such topic in mind? If you don't want to grab the tar baby, that's
fine by me -- I never placed one on the path. All I was saying was that
feedforward is not *necessarily* involved when we control in the absence of
part of our normal complement of perceptual signals. My remarks did not
rule out the possibility that feedforward might be used, sometimes,
somewhere. It is just that no example offerred during this round
of discussion (playing tennis, driving on ice or sand, playing chess,
walking to the john in the dark, etc.) *requires* that we abandon the
attempt to use a feedback PCT model and that we incorporate feedforward.

Martin:

Down with all-or-none thinking!

Have you seen any of that around here?

Until later,

Br'er Bear (aka Tom)

4

[Martin Taylor 931109 16:40]
(Tom Bourbon 931109.1125)

Too many misunderstandings at once for me to deal with. I'll confine
myself to one in this posting. Maybe the others will resolve themselves
through parallel postings.

It is just that no example offerred during this round
of discussion (playing tennis, driving on ice or sand, playing chess,
walking to the john in the dark, etc.) *requires* that we abandon the
attempt to use a feedback PCT model and that we incorporate feedforward.

Martin:

Down with all-or-none thinking!

Have you seen any of that around here?

Right there. "Abandon."

Another example: sawtooth tracking with NO display as a contrast to
CONTINUOUS display. I asked about momentary blanking out, envisaging
as much as 1/3 of the length of one peak-to-peak motion. You replied
with an all-or-none answer. All I wanted was to know the behaviour
of the *model* when the perceptual signal goes to zero for a short
period, and to find out whether this matched what the human does.

I see all-or-none thinking in numerous postings to this net. "There is
NO information about the fluctuation [replacing "disturbance" with the
word agreed by Bill and me] in perception" is claimed as the ONLY possible
alternative to "Perception has perfect information about the fluctuation."

But I intend to go as far as I
can, given the limits on my abilities, to determine the extent to which
Nature can be modeled as though it had opted for the simple, rather than
the ingenious. (But I DO think a simple PCT loop, in hierarchical-parallel
systems, is pretty ingenious!) I want to see how far we can go using as our
model the simplest PCT loop, then that loop in a hierarchical-parallel
system.

An approach I applaud. But that doesn't stop me from considering simple
alterations to the basic structure. If the simple structure, in as simple
a connection (i.e. not using more of them), using as simple a learning
procedure, can do as good a job, then we should certainly believe in the
simple structure preferentially. And unless we know just what the simple
structure can do, we can't defend it against claims that an altered
structure in a simpler configuration could do something it can't. So I
ask without prejudging the case--does the tracking model do what humans
do when the display is momentarily blanked out, using 1/3 of the height
of the sawtooth as a reasonable instantiation of "momentarily." It would
be particularly interesting to do this around the time of the sawtooth
peak, given that the sawtooth is regular and doesn not reverse at random.
But anywhere in the track would be worth trying.

Martin

5

From Tom Bourbon [931109.1730]

[Martin Taylor 931109 16:40]
(Tom Bourbon 931109.1125)

Too many misunderstandings at once for me to deal with. I'll confine
myself to one in this posting. Maybe the others will resolve themselves
through parallel postings.

It is just that no example offerred during this round
of discussion (playing tennis, driving on ice or sand, playing chess,
walking to the john in the dark, etc.) *requires* that we abandon the
attempt to use a feedback PCT model and that we incorporate feedforward.

Martin:

Down with all-or-none thinking!

Have you seen any of that around here?

Right there. "Abandon."

"Abandon?" Hmm. I thought that by placing the qualifying emphasis on
*requires*, I was leaving open the possibility of the other option. And it
is a fact that the examples I listed have been offerred as evidence that we
operate much or most of the time in a feedforward mode. The selection and
citation of those examples was not my doing.

. . .

I see all-or-none thinking in numerous postings to this net.

So do I. for example, when we look at the post that started this
round of discussion about feedforward, we find the following, some of which
seems to deserve the label you suggest:

ยทยทยท

======================================
[Hans Blom, 931103]

Sometimes there is no feedback signal, either because it just isn't there,
or because you do not pay attention to it. Some more examples. (1) After I
switch off the light in my bedroom at night with the switch that is locat-
ed next to the bedroom door, I walk some 10 feet towards the bed, grab the
bedcover and get into the bed, all in total darkness. All the while, there
is no feedback whatsoever about my position relative to the bed. Yet I
grab and lift the bedcover accurately without fail.

That seems pretty all-or-none-ish to me. Much of the discussion that
has followed has been directed at that remark, and at others just as stark.
I'm sorry, but this is a line of talk that cannot be laid at the feet of the
usual "culprits" on this net.
. . .

. . . As the saying goes, experts do not think, they ACT. In
many cases, they have a fail-safe strategy for action that will work
reliably at all times.

. . .

It should by now be clear that "feedforward" is the theory that underlies
stimulus-response systems. Just like feedforward controllers have their
place, stimulus-response descriptions of behavior have their place in
psychology. Stimulus-response behavior, motor or mental, often takes over
after learning is complete -- whenever learning can be complete, which is
only if the learning is with respect to reliable, "lawful" aspects of the
world. Just like in industrial control systems, where you frequently can
choose between a feedback or a feedforward approach, the same is true in
people. Some people act in very stereotyped ways on many occasions. Others
are less predictable, maybe because they take more aspects of the situat-
ion into account. In particular, experts -- or those who think that they
are experts -- frequently behave according to very stereotyped (but some-
times idiosyncratic) patterns. It has been estimated that many family
doctors or general practitioners (now extinct in the US?) accurately
recognize about 80 percent of patient's problems within 30 seconds of
meeting the patient. That's pattern matching/feedforward. It's the other
20 percent, which comprises cases that they seldomly see and where they
have not built up enough expertise, that are often difficult for them. In
these cases they may have to fall back on a feedback-like process of dif-
ferential diagnosis.

. . .

. . . When driving on snow, I still have the same goal
as always: to follow a certain course. But THE WAY IN WHICH I follow that
course should be different, and that is something that I get no feedback
about -- or too late.

. . .

To sum up: feedforward is everywhere. My guess is that the purpose of much
of the learning that we do over our lifetime -- in becoming an expert in
living -- has to do with replacing feedback with feedforward control.

=========================================

That is the post that set the tone of the present discussion. I won't
track down and repeat the assertions that people go ballistic to play chess
or tennis.

Martin:

Another example: (Of my all-or-none thinking, Martin?) sawtooth tracking

with NO display as a contrast to

CONTINUOUS display. I asked about momentary blanking out, envisaging
as much as 1/3 of the length of one peak-to-peak motion. You replied
with an all-or-none answer. All I wanted was to know the behaviour
of the *model* when the perceptual signal goes to zero for a short
period, and to find out whether this matched what the human does.

Of course, a one-level, one reference-signal control loop will not duplicate
the performance of a person who uses nunerous perceptual systems in parallel
and in a hierarchy. I will never claim otherwise. Neither will any other
PCT modeler I know.

I assure you it was not my intention to offend you, Martin. I thought it
might be of some value to describe data on behavior in a circumstance
where a person tries to recreate a highly practiced pattern of movements
while deprived only of the accustomed visual information about the pattern.
I believe those data have at least a little bearing on the subject of how
well people perform "without feedback" and that in turn has some relevance
for what we might want to model. I never thought any person would take my
example as evidence that CSG-L is a forum for PCT people to engage in
all-or-none thinking. I am disappointed to learn otherwise.

Tom