Feedback isn't too slow

CSGnet Archive CSG1997
1

[From Bruce Abbott (971208.1210 EST)]

Bill, you've argued that in a control system "feedback isn't too slow," on
the basis that the system can generate larger accelerations than an open
loop system can, because it can also apply braking as the CV approaches its
new steady-state value. While this is true, I'm wondering whether your
argument misses its mark.

It seems to me that the reason some claim that feedback is too slow is not
that control systems change their outputs too slowly compared to open-loop
systems (i.e., slew rate limited by feedback), but rather, that in some
cases the information about the CV needed to "guide" the output is not
available at the time it is needed, because of the lag involved. As a
result, one takes action based on a preformulated "plan" that makes
assumptions about what should be done, given previous experience with
similar situations. The unfolding "plan" sets the references for faster
lower-level control systems operating in the usual closed-loop mode.

As an example, consider the housefly narrowly averting death by leaping into
the air as a rolled newspaper crashes down upon the table where it had been
standing microseconds before. The fly's visual system detected the rapidly
expanding image of the newspaper, which triggered a sudden contraction of
muscles attached to the middle legs, flinging the fly upward and backward
and pull-starting the flight-motor. This direction of movement usually
brings the fly out from under the newspaper before it strikes the table,
thus saving the fly's life. It is (apparently) not determined by feedback
from the visual system by which the fly could steer itself away from the
approaching missile.

I am not arguing for this interpretation as a general case, but doesn't the
"feedback is too slow" position rest on the presumed need to begin action in
the absence of timely feedback, rather than on speed of response?

Regards,

Bruce

2

[From Bill Powers (971208.1457 MST)]

Bruce Abbott (971208.1210 EST) --

Bill, you've argued that in a control system "feedback isn't too slow," on
the basis that the system can generate larger accelerations than an open
loop system can, because it can also apply braking as the CV approaches its
new steady-state value. While this is true, I'm wondering whether your
argument misses its mark.

It seems to me that the reason some claim that feedback is too slow is not
that control systems change their outputs too slowly compared to open-loop
systems (i.e., slew rate limited by feedback), but rather, that in some
cases the information about the CV needed to "guide" the output is not
available at the time it is needed, because of the lag involved.

Wouldn't that argument apply just as well to the open-loop system? The lag
would be just the same, wouldn't it? Why should perceptual or output
functions work any faster in an open-loop system? The only difference is
that in a control system there is a feedback path that can "catch" a
too-large response and keep it from overshooting, or make corrections in a
trajectory that is initially aimed a little (or a lot) wrong.

The difference I was discussing with Rupert concerned how fast a
_specific_ motion can be produced -- not just a sudden all-out spasm, but
something requiring a degree of precision, such as reaching out to catch a
fast-approaching baseball. You can get your hand to the appropriate
position fastest if it's moving under feedback control, because the initial
acceleration can be larger than in the open-loop case.

As a
result, one takes action based on a preformulated "plan" that makes
assumptions about what should be done, given previous experience with
similar situations. The unfolding "plan" sets the references for faster
lower-level control systems operating in the usual closed-loop mode.

Yes, I'm sure this is conceivable, but you have to remember that most of
the examples people present have been modeled as open-loop processes from
the start and nobody has looked at a control-system model for the same
examples. What's always forgotten is that before you can select a plan, you
have to perceive the current state of the world, analyze it and deduce the
appropriate plan, and then put the plan into action. This process requires
time and introduces delays of its own. The net result is, I contend, always
slower than the operation of an equivalent control system that simply
monitors some aspect of the environment and is already set up to correct
errors by altering lower-level reference signals.

Consider your fly example:

As an example, consider the housefly narrowly averting death by leaping into
the air as a rolled newspaper crashes down upon the table where it had been
standing microseconds before. The fly's visual system detected the rapidly
expanding image of the newspaper, which triggered a sudden contraction of
muscles attached to the middle legs, flinging the fly upward and backward
and pull-starting the flight-motor. This direction of movement usually
brings the fly out from under the newspaper before it strikes the table,
thus saving the fly's life. It is (apparently) not determined by feedback
from the visual system by which the fly could steer itself away from the
approaching missile.

But what you have just described is exactly a control process, which is
hooked up to keep the rate of approach of large objects below some
reference level. Exactly the same control system would be used in
approaching a large object to land on it. When the reference level for rate
of approach is set to zero, as the large object approaches the
leaping-flying mechanism is activated to keep the rate of approach as near
as possible to zero. The faster the approach, the larger the error and the
more energetic the flying. In order to land on the newspaper or flyswatter,
the fly simply sets a somewhat positive reference level for rate of
approach, which is reduced as the fly nears the surface -- by the same
control system.

This control process does not begin "microseconds" before the newspaper
crashes down. It starts while the newspaper is still a foot or so, many
milliseconds, away. You can verify this by moving your hand slowly until
it's about two feet from the fly, and then making a sudden move of six
inches toward the fly. The fly will most likely depart. A faster-moving
object will get closer to the fly before it can initiate the output action
to start correcting the "looming" error. A flyswatter will get there before
the output has time to act.

When you say this process is not determined by "visual feedback," what do
you mean? Isn't vision itself a feedback path? If the visual controlled
variable is disturbed by a rapid change in the visual image, this creates
an immediate error which produces an output that tends to oppose the
disturbance. How is that different from what you described? Just visualize
the standard PCT diagram: a disturbance propagates to the comparator and
the output before the first feedback change appears. How does that feedback
slow the system down?

I am not arguing for this interpretation as a general case, but doesn't the
"feedback is too slow" position rest on the presumed need to begin action in
the absence of timely feedback, rather than on speed of response?

If there is a "need to begin action," it has to be known through some
sensory pathway, doesn't it? And whether a change in a perception signals a
need to act depends on the reference signal with which the sensory
information is being compared. It seems to me that the connections in the
control loop couldn't be more direct. Perceptual signal, comparator,
action. The only way to get faster than that would be to hook the sensory
nerve directly to the muscle with no comparator -- and that would prevent
the fly from ever flying toward a large object on purpose. Feedback doesn't
introduce any delays that wouldn't be there anyway.

This whole "feedback is too slow" myth was invented by people who simply
didn't understand how feedback control systems work.

BTW, have you played around with the "illusion" program any more? Have you
tried breaking the loop to see what happens? I think I'll post a
modification in which the 6th choice is simply to eliminate the feedback
path -- unless you beat me to it. With no feedback function, the output
will be a wildly different function of the disturbance.

Best,

Bill P.

3

[From Rupert Young (971018.1300 BST)]

(Bill Powers (971208.1457 MST))

The difference I was discussing with Rupert concerned how fast a
_specific_ motion can be produced -- not just a sudden all-out spasm, but
something requiring a degree of precision, such as reaching out to catch a
fast-approaching baseball. You can get your hand to the appropriate
position fastest if it's moving under feedback control, because the initial
acceleration can be larger than in the open-loop case.

Bruce Abbott (971208.1210 EST) --

It seems to me that the reason some claim that feedback is too slow is not
that control systems change their outputs too slowly compared to open-loop
systems (i.e., slew rate limited by feedback), but rather, that in some
cases the information about the CV needed to "guide" the output is not
available at the time it is needed, because of the lag involved.

Yes, I think there are two issues here. One of closed-loop vs. open-loop
regarding speed of response, which Bill addressed. The other concerns the need
for an emulator (model-based control) due to lags in the system. The reasoning
for this that was given to me recently (by Rick Grush) was that because of
feedback delays what you perceive reflects the situation as it WAS not as it
IS. So, for example if you are driving it won't take you long to crash if you
are perceiving the past not the present.

···

--
Regards,
Rupert

4

[From Bill Powers (971209.0359 MST)]

Rupert Young (971018.1300 BST) --

Yes, I think there are two issues here. One of closed-loop vs. open-loop
regarding speed of response, which Bill addressed. The other concerns the

need

for an emulator (model-based control) due to lags in the system. The

reasoning

for this that was given to me recently (by Rick Grush) was that because of
feedback delays what you perceive reflects the situation as it WAS not as it
IS. So, for example if you are driving it won't take you long to crash if you
are perceiving the past not the present.

That statement requires some conditions on it. When your spinal control
systems sense the tension in a tendon, the signal that reaches the spinal
motor neuron reflects the tension not as it IS, but as it WAS -- about 2
milliseconds ago. This, of course, makes absolutely no difference when the
muscle tension can't change much in 10 milliseconds. You have to consider
the delays in relation to the speed with which normal disturbances can
change, and with which the actions that are produced can change.

With respect to keeping a car on the road, Grush is simply demonstrating
ignorance. Even with a delay of about 0.3 seconds in the visual system
(which is erroneously assumed by most people talking about "feedback is too
slow" -- it's really about half of that), a negative feedback control
system will work perfectly well. It can be stabilized by using an
integrating output function with the right sensitivity factor. There is no
need at all for prediction. There may well be cases in which prediction
would improve matters, but that isn't one of them.

In many cases where prediction might seem to be needed, this need can be
eliminated by changing the definition of the controlled variable. If you
define the controlled variable involved in steering a car as the position
of the car in its lane, then it's hard to explain why the car doesn't swing
wide at the start of every curve in the road because of the driver's
reaction time. So people who are into prediction say that the driver has to
extrapolate the position of the car ahead, so as to anticipate the curve
before the car gets to it. Sound familiar?

Well, that's not necessary at all. If you simply line up the nose of the
car with a point on the road some distance ahead, and control _that_
perception, the road will do all the predicting for you. You will see the
controlled variable start to change before the car itself has reached the
point you're watching. If you've picked the right point, your reaction
delay will just equal the time it takes the car to get to that point, and
the effective delay of your action will be zero. An experienced driver
looks farther ahead at higher speeds. He or she doesn't have to do any
predicting.

What about avoiding a collision with another car approaching on a side
road? In the psychological literature concerning control theory, the only
solution I've seen is for the driver to extrapolate his position and the
other car's position into the future to predict a hit or a miss. But every
sailor not only knows this is a poor way to avoid a collision, but that
there is a much simpler solution which people actually use. It's called the
"constant-bearing" method. If the angle of the other vehicle from your
direction of travel is constant, there is going to be a collision. So all
you have to control for, to prevent a collision, is a continuously
increasing or decreasing angle to the other vehicle. This same principle
applies to running to catch a ball, or to intercepting a prey, where you
want a collision and deliberately maintain a constant bearing. Absolutely
no calculation of trajectories is required; this method will get you there
even if the target takes evasive action (unless it's faster than you are,
but then prediction wouldn't help you catch it, either).

Prediction is always the poorest choice, often even when it's the only
choice. There is only a very narrow time-span over which prediction is
worth the trouble. For any shorter time, prediction itself just wastes time
that could be better used in feedback control. For any time beyond this
span, the accuracy of predictions rapidly falls off, to the point where you
begin taking more wrong actions than right ones.

Perhaps the most important fact in this discussion is that even predictive
control is feedback control. It's very, very seldom that you make a
prediction, carry out the calculated action, and then simply rely on the
result. Most of the time you make a prediction, check up on it a short time
later by perceiving the new situation, make a new prediction based on the
new information, check it again, and so on in a series of successive
approximations. The frequency with which you must update your perceptions
and make new predictions depends on the accuracy of your predictions, the
speed with which disturbances can come and go, and the speed with your
actions can change. When the intervals at which new predictions must be
made becomes shorter than the time it takes to develop the prediction and
carry out the implied action, prediction is no longer of any use.

I do wish that people who talk about what can and can't be accomplished
with feedback control would learn something about it before they open their
mouths.

Best,

Bill P.

5

[From Bruce gregory (971209.1100 EST)]

Rupert Young (971018.1300 BST)

Yes, I think there are two issues here. One of closed-loop vs. open-loop
regarding speed of response, which Bill addressed. The other concerns the need
for an emulator (model-based control) due to lags in the system. The reasoning
for this that was given to me recently (by Rick Grush) was that because of
feedback delays what you perceive reflects the situation as it WAS not as it
IS. So, for example if you are driving it won't take you long to crash if you
are perceiving the past not the present.

Notice that you don't crash very often. Notice that if you close
your eyes, you quickly crash. Does this say anything about the
likely role of model-based control in driving?

Bruce

6

[From Bruce Gregory (971209.1130 EST)]

Bill Powers (971209.0359 MST)

I do wish that people who talk about what can and can't be accomplished
with feedback control would learn something about it before they open their
mouths.

Why exactly is it that you think feedback control should be
treated differently than anything else? :wink:

Bruce

7

[Martin Taylor 971209 00:05]

Bill Powers (971208.1457 MST) to Bruce Abbott (971208.1210 EST) --

I agree with Bill's comments to Bruce, but would like to add something.

It seems to me that the reason some claim that feedback is too slow is not
that control systems change their outputs too slowly compared to open-loop
systems (i.e., slew rate limited by feedback), but rather, that in some
cases the information about the CV needed to "guide" the output is not
available at the time it is needed, because of the lag involved.

Wouldn't that argument apply just as well to the open-loop system? The lag
would be just the same, wouldn't it?
...

What I think gets missed is that in neither open or closed loop systems
is feedback lag concerned in the time to _first_ reaction to a step
disturbance. That lag, for both, is the processing time in the S-R sense,
through the perceptual input function, the comparator, the output
function and lower-level control loops to the muscles. The _feedback_
lag is from the output to the CEV, and that's important only as a lag
when there is a step change in the _reference_ signal. Otherwise, the
lag affects the permissible bandwidth of the feedback loop.

If the feedback lag affects the time it takes the perceptual signal to
change following a change of reference signal, it can be compensated if
the system "knows" the characteristics of the environmental feedback
function. The reference can change earlier, to synchronize the physical
event at the right moment. Baseball batters begin to swing before the
ball arrives, and before they know whether it will arrive in the strike
zone. I read an article in an old Science recently in which it was
electrophysiologically discovered that a running turkey (bird, not human)
tenses its leg muscle just _before_ the foot hits the ground, allowing the
tendon to act like a spring, conserving energy. (Sounds a bit like Hans
Blom's pre-hop when a person decides to hop rather than having to hop
as a consequence of being pushed, doesn't it?).

If there is feedback delay, there is a phase shift of any given frequency
around the loop. If this phase shift is large enough, and the gain high
enough, the loop will go into oscillation rather than controlling the
perceptual signal. That's why feedback lag limits the bandwidth of the
control loop. What it means for control is that the loop will not be able
to control against disturbances that vary unpredictably too rapidly (i.e.
that have a bandwidth greater than the bandwidth of the control loop).

But an open-loop system can't control against such disturbances at all
(if they are observed, and the system's output opposes them, the loop
isn't open; if they are not observed, there's no way any planning system
knows what they are to oppose them, by definition of "unpredictable").

What seems to come out of this is that the problem of "too slow" arises
only when the event requiring action comes from _within_ the organism,
as a change in the reference signal. But since that change comes from
within, it could be made to come earlier if the lag is something constant
enough for the reorganizing system (and devices like Bill's Artificial
Cerebellum?) to have compensated for. After all, a child's first swings
with a baseball bat usually come ludicrously late, well after the ball
has gone by, but the grown child earning millions of dollars for hitting
the ball has no such problem.

Martin

8

[From Bill Powers (971209.1551 MST)]

Martin Taylor 971209 00:05 --

What I think gets missed is that in neither open or closed loop systems
is feedback lag concerned in the time to _first_ reaction to a step
disturbance. That lag, for both, is the processing time in the S-R sense,
through the perceptual input function, the comparator, the output
function and lower-level control loops to the muscles. The _feedback_
lag is from the output to the CEV, and that's important only as a lag
when there is a step change in the _reference_ signal. Otherwise, the
lag affects the permissible bandwidth of the feedback loop.

My very point, and well said.

If the feedback lag affects the time it takes the perceptual signal to
change following a change of reference signal, it can be compensated if
the system "knows" the characteristics of the environmental feedback
function. The reference can change earlier, to synchronize the physical
event at the right moment. Baseball batters begin to swing before the
ball arrives, and before they know whether it will arrive in the strike
zone. I read an article in an old Science recently in which it was
electrophysiologically discovered that a running turkey (bird, not human)
tenses its leg muscle just _before_ the foot hits the ground, allowing the
tendon to act like a spring, conserving energy. (Sounds a bit like Hans
Blom's pre-hop when a person decides to hop rather than having to hop
as a consequence of being pushed, doesn't it?).

I think you're making the point that while the reference signal can be
altered prior to some event, this by no means eliminates the lag, because
something has to decide when to alter the reference signal and that
decision must be based on perception of the environment in which the
behavior occurs. The effect is similar to my driving example: one learns to
control a variable taken from "upstream" in the chain of events, which
effectively compensates for any lag in the control system. It's as though
one were swinging at the ball just as it leaves the pitcher's hand,
although the actual swing doesn't occur until the ball reaches the plate.
In the case of baseball, the control that results is not terribly good. One
can only vary the height of the swing above the ground slowly, over many
trials, as a function of the visual appearance of the ball half a second or
less after it is released. But at least one can learn to swing just as the
ball is passing by, by observing its position at some time prior to that
event. Of course pitchers know this, and counteract the strategy by using
the changeup pitch -- a pitch that seems like any other, but that delivers
the ball more slowly. The batter "gets ahead of the pitch" and swings
before the ball gets there.

What seems to come out of this is that the problem of "too slow" arises
only when the event requiring action comes from _within_ the organism,
as a change in the reference signal. But since that change comes from
within, it could be made to come earlier if the lag is something constant
enough for the reorganizing system (and devices like Bill's Artificial
Cerebellum?) to have compensated for. After all, a child's first swings
with a baseball bat usually come ludicrously late, well after the ball
has gone by, but the grown child earning millions of dollars for hitting
the ball has no such problem.

Right. This analysis needs some work, but I'm sure it's on the right track.

Best,

Bill P.

9

[From Rupert Young (971211.1430 UT)]

(Martin Taylor 971209 00:05)

The _feedback_
lag is from the output to the CEV, and that's important only as a lag
when there is a step change in the _reference_ signal. Otherwise, the
lag affects the permissible bandwidth of the feedback loop.

By this do you mean, to put it technically, that if the thing you're trying to
control is changing faster than the delay around the loop then control will
fail ? eg. Like trying to track the erratic flight of a butterfly. If the
delays in our visual system were smaller we would be more successful at
tracking the butterfly.

···

--
Regards,
Rupert