Gretzy, Self-reporting, catching flyballs, etc

When asked to account for his skill in playing hockey, Wayne Gretzy reportedly said, "I skate to where the puck will be." That, of course, implies prediction. It's also an instance of self-reporting. Rick Marken has suggested that baseball players don't run to where the ball will be, instead they control their perception of the ball's angle and velocity (I think I have that right). I've been watching a lot of football lately - and a lot of passing. It seems to me that Gretzy, football receivers and baseball players who catch fly balls have something in common, something to do with perceptions of angle and velocity. Do I have that right?

···

--
Regards,

Fred Nickols
Managing Partner
Distance Consulting, LLC
nickols@att.net
www.nickols.us

"Assistance at A Distance"

[From Bill Powers (2009.12.17.1600 MST)]

When asked to account for his skill in playing hockey, Wayne Gretzy reportedly said, "I skate to where the puck will be." That, of course, implies prediction. It's also an instance of self-reporting. Rick Marken has suggested that baseball players don't run to where the ball will be, instead they control their perception of the ball's angle and velocity (I think I have that right). I've been watching a lot of football lately - and a lot of passing. It seems to me that Gretzy, football receivers and baseball players who catch fly balls have something in common, something to do with perceptions of angle and velocity. Do I have that right?

You'd think the person doing the catching would know, but as far as I can see "skating to where the puck will be," or running to where the football or the baseball will be, is probably wrong. The problem is that whatever the method is, the behavior does move the catcher to where the ball or puck will be when finally caught (if not, it isn't caught). So the description is really based on the outcome, not on knowledge of how the outcome is produced. If you asked Gretzky "HOW do you skate to where the puck will be?" he probably couldn't tell you, or he'd give you some theory he heard from someone else. Or he'd just say "practice."

In football, it's the quarterback who has to try to estimate where the ball will come down, and aim so the receiver can get there. But the receiver does a lot of adjusting to make up for the quarterback's miscalculations, and normally does it by keeping his eye on the ball over most of the trajectory. When the receiver doesn't look soon enough, he just misses the ball, or it hits him in the back. It happens a lot.

When I first read about the way baseball players catch balls (article in Science by Chapman: keep the ball slowly rising in the visual field) I didn't believe it until I actually got a friend to throw some balls to me. I thought I just ran to the place where the ball was going to come down. So I tried doing it the way the article said they really do it -- and it worked perfectly! Then I realized that I had probably been doing that all along, though there were some cases where I really did turn and run, then turn around, locate the ball, and catch it. But I also realized that those were the cases in which I couldn't do it the other way because the ball was too high and fast and I couldn't run backward fast enough while keeping my eye on the ball. And I didn't catch anywhere near all of those. When I turned around, I wasn't always quite close enough to intercept the ball once I found it again. Sometimes it worked, as you'd expect, but it was always a near thing. "Watching the ball into the glove," as the coaches put it, works far better. The guy who catches it while running blindly or behind his back is looked down upon as a hot-dog -- risking the team's win just to show off. The players know that hot-dogging is unreliable even if it does work sometimes. Willy Mays was just plain lucky.

Best,

Bill P.

···

At 07:55 PM 12/17/2009 +0000, Fred Nickols wrote:

Hi everybody !

BP : Then I realized that I had probably been doing that all along, though
there were some cases where I really did turn and run, then turn around,
locate the ball, and catch it. But I also realized that those were the cases
in which I couldn't do it the other way because the ball was too high and
fast and I couldn't run backward fast enough while keeping my eye on the ball.

BP and MT (much, much earlier) :
BP wrote : "In my mind's eye I can see the imagined path of a served
ping-pong ball very soon after it leaves the opponent's paddle.."
MT wrote : "That's prediction in my book"
BP wrote : "Yes I agree. We're not disccusing whether prediction happens --
we both know it does.

BH : Now finally we have some experienced "playing" situation that could
maybe help us get on "higher level" to solve the conflict.
I remember when I talked once with Richard, we had much less problems,
because we were talking while "experiancing" the commnon playing situation,
I hope I remember right, in some rocket game. It wasn't difficult to
establish what is happening.

The most important thing I see till now is that we established the existance
of feedforward and that combination with feed-back exists.

I don't say that Bill is wrong. I hope I never said that. He is right about
that keeping eye on the ball (keep the loop closed) is probably the best way
if we want to achieve whatever our goals are.

But I don't see that as a problem. A problem I see is not whether people's
or animal's behavior has more feedforward or more feedback, or what is more
important or efficiant. The problem I see is why it spontaneously occurs
"switching", when control is impossible as Bill said, and why this mechanism
was "developed" through evolution if it is so unusefull and should be
replced by feedback as PCT predicts ?

I hope we are not talking about how behavior will suit the most to theory of
PCT, but how theory of PCT will succesfuly explain what is really happening
in the behavior and why we make so many mistakes in the game and in life.
Probably because we don't control actual perceptions all the time. So I'm
asking myself why people are using that "alternative mechanism" in behavior
if it's not so efficiant as control is or is maybe effeciant only in
combination with control.

I never tried to say that Bill is not a great inventor and even less that
PCT is not a great theory. But I'll always say that nothing could be so good
that it couldn't be better. If we'll try to fit all behaviors into PCT, it
will never advance. And I think this is also the case with feedforward. Why
not trying to explain why feedforward occurs in HPCT and use it as something
that has the best explanation in PCT.

People observe and obviously common experiances show that there is something
"strange" happening. Now I think we have to explain that. Just answer, it's
not good to use feedforward, it's better that we use control, because it's
the most effective, will not help much.

So I don't see the problem in question whether closed loop is the best to
use or not. I'm asking myself why people use feed-forward so much if it's
not neccesary in some situations or even something what result with bad
consequences ?

I agree with Bill that the best way to play is to be all the time in control
mode. That's why coaches yell. But goals of players which unexpectadly
occurs during the game differs from those in coaches minds, because they are
not experiencing the playing situation.

The question for me was always why I can't do "what's right" in the critical
situation, and I ussually always realized after, that I made a mistake and
that it would be better if I've done as coach said. But in "that moment" I
just "couldn't remember", although I've heard that hundred times on
trainings and games.

So I think that we must explain why feedforward is happening during the game
even to players (professionales) who train 4,5,6 and even more hours a day
just for not doing that mistakes. But they still do. They fail to catch the
ball, they fail to receive pack, they fail to receive ball, they fail to
switch over on right place, they fail to score from 2-meters from goal and
40.000 or more spectators think "what an idiot", even I could score that
ball. Well I'd like to see them in that situation.

So I'm asking myself all the time why feed-forward exists in behavior and
why organisms use it probably spontaneuos or why evolution "develope" that
mechanism and kept it in living organisms if it's not so usefull or even
neccesary for behavior ? Maybe it's some kind of mechanism in control
systems that occurs automatically ?

So I'm aksing myself if we could turn this converastion into advantage for
PCT and explain why the combination of feedforward and feedback occurs and
how to use it the training to make players more efficiant and finally to win
the game. That could by my oppinion make PCT usefull aplication. "Guessing"
what is better won't do any advance.

I've heard some good theoretical back-ground for feed-forward and it seems
to me quite exceptable from PCT view, but I must ask the member of CSG net
if we can introduce our conversation. I respect the intimacy if others do.
But if they don't, I don't either.

Best,

Boris

[From Bill Powers (2009.12.22.0730 MDT)]

BH: The most important thing I see till now is that we established the existance of feedforward and that combination with feed-back exists.

BP: As your citations show, I have never claimed that prediction and "feedforward" don't exist. I posted a diagram showing how they would work together in the same system.

Before any feedback effects could have existed (at the start of evolution), there must have been effects of environmental stimuli on the interaction of proto-organisms with their environments (ordinary causation). When negative feedback arose, however, it would quickly have replaced feedforward or S-R relationships, leaving them to be important only when feedback control is impossible. And saying "important" is not the same as saying that feedforward would work very well. Feedforward can't deal with any changes in the output functions of a system, or changes in the properties of the world being acted upon, or occurrance of disturbances of kinds that can be detected only through their effects on a critical variable.

BH: But I don't see that as a problem. A problem I see is not whether people's or animal's behavior has more feedforward or more feedback, or what is more important or efficiant. The problem I see is why it spontaneously occurs "switching", when control is impossible as Bill said, and why this mechanism was "developed" through evolution if it is so unusefull and should be replced by feedback as PCT predicts?

BP: My guess is that it wasn't "developed" through evolution, but was there from the start, and can also result from reorganization. Disturbances of proto-organisms (organic molecules) altered the interactions with the environment according to physical and chemical laws. They still do. Effects of that kind that reduce the chances of continued existence eliminate that line of development: that is natural selection. The entities that remain are those which don't react, or react less, to external influences in ways that reduce their chances of surviving. That process of elimination is random, slow, and inefficient. But over a billion years, it works. It also works in a single lifetime as we become organized.

When negative feedback was developed through natural selection, there must have been an enormous increase in efficiency: negative feedback control will replace feedforward everywhere that is possible.

BH: I hope we are not talking about how behavior will suit the most to theory of PCT, but how theory of PCT will succesfuly explain what is really happening in the behavior and why we make so many mistakes in the game and in life. Probably because we don't control actual perceptions all the time. So I'm asking myself why people are using that "alternative mechanism" in behavior if it's not so efficiant as control is or is maybe effeciant only in
combination with control.

BP: My guess is that we use feedforward when there is no way to get immediate and accurate feedback about the state of whatever we're controlling (or when the effort to do so isn't worth the trouble because it's not an important variable). But "mistakes in the game of life" are a different matter. Even if we're controlling a variable, we may not pick a good variable to control -- think of people who want to make billions of dollars, and do it, and are laughed at or hated by everyone instead of being admired. It doesn't help them to be control systems.

BH: I never tried to say that Bill is not a great inventor and even less that
PCT is not a great theory. But I'll always say that nothing could be so good
that it couldn't be better. If we'll try to fit all behaviors into PCT, it
will never advance. And I think this is also the case with feedforward. Why
not trying to explain why feedforward occurs in HPCT and use it as something
that has the best explanation in PCT.

BP: First, it would be good to know exactly what you mean by the term "feedforward." I have my own understanding of it, but perhaps you see something there that I don't see. Some examples would be useful. I have described one (the fielder running to where the ball will be without looking at the ball) and have noted that in my limited experience with watching outfielders on television, this method doesn't appear to work very well. It's used mostly when the more reliable method, watching the ball all the time, can't be used. Also, I've commented on a related problem: collision avoidance. It is possible to predict where another ship is going to be, and where your own ship will be at various future times, and decide in this way whether both ships will arrive at the same place at the same time. But it is far easier simply to take repeated bearings on the other ship. If the bearing (the angle relative to your own ship's direction) is constant, and the distance to the other ship is decreasing, there will be a collision. No need to predict specific future positions at all. If you control to keep the bearing changing, you will avoid the collision. It's just a matter of controlling the right present-time variable.

What kinds of examples are you thinking of?

BH: People observe and obviously common experiances show that there is something "strange" happening. Now I think we have to explain that. Just answer, it's not good to use feedforward, it's better that we use control, because it's the most effective, will not help much.

BP: What's strange about feedforward? It's just a blind reaction to a sudden change. You can organize yourself that way if you want to; there's no law against it. A small child sees a fire and reaches out toward it, perhaps "instinctively." But not more than once. We all quickly learn the advantages of feedback control, and fortunately evolution has provided the necessary equipment to acquire control systems.

BH: So I don't see the problem in question whether closed loop is the best to
use or not. I'm asking myself why people use feed-forward so much if it's
not neccesary in some situations or even something what result with bad
consequences ?

BP: Since there's no instruction book, each of us has to learn through experience what variables we need to control, and how to control them. No two of us experience identical environments, but we learn to control anyway, because we can reorganize. I don't think people use feedforward very often, compared with the number of situations in which they use feedback. But people do use feedforward, which proves that reorganization can produce that result just as it can produce control systems. But usually feedforward leaves enough error so reorganization will continue, leading to the development of a control system instead. It's possible that the result will be a combination of feedforward and feedback, but in my experience that's not usually needed.

I actually tried using some feedforward in a tracking task, displaying a representation of the disturbance that was affecting the cursor (normally, the disturbing variable can't be sensed in my tracking tasks). I wanted to see if it would improve control. The result was to make control worse, because attention was taken away from the target movements. So that wasn't a good way to introduce feedforward. A better way would be to make it internal to the subject, connecting a perception of the moving target directly to the circuits that operate the muscles. But the feedback model we use already accounts for more than 99% of the variance of the behavior, so adding feedforward couldn't improve control by a large amount. Maybe a little bit.

BH: I agree with Bill that the best way to play is to be all the time in control mode. That's why coaches yell. But goals of players which unexpectadly occurs during the game differs from those in coaches minds, because they are not experiencing the playing situation.

BP: Well, coaches don't yell at players who catch the ball every time, so if feedforward works as well as feedback, I'm sure the coaches would permit it unless they like losing games.

BH: The question for me was always why I can't do "what's right" in the critical situation, and I ussually always realized after, that I made a mistake and that it would be better if I've done as coach said. But in "that moment" I just "couldn't remember", although I've heard that hundred times on
trainings and games.

BP: Hearing and doing are two different things. I had a football coach who would yell "Put your back into it." I could never figure out how to do that.

BH: So I think that we must explain why feedforward is happening during the game even to players (professionales) who train 4,5,6 and even more hours a day just for not doing that mistakes. But they still do. They fail to catch the ball, they fail to receive pack, they fail to receive ball, they fail to
switch over on right place, they fail to score from 2-meters from goal and
40.000 or more spectators think "what an idiot", even I could score that
ball. Well I'd like to see them in that situation.

BP: What you're saying, if I understand you, is that in the actual situation, either you can't remember what you're supposed to do, or the behavior you've been told to use wouldn't work. Of course if you're supposed to react to a particular situation automatically by doing a particular behavior, that is already feedforward or stimulus-response, so you're describing a choice between two feedforward strategies, the one you automatically use and the one you're told to use. Maybe a control process would be better than either one. Good coaching tells you what variables to control, not what behaviors to carry out. But not many coaches know that, apparently.

BH: So I'm asking myself all the time why feed-forward exists in behavior and
why organisms use it probably spontaneuos or why evolution "develope" that
mechanism and kept it in living organisms if it's not so usefull or even
neccesary for behavior ? Maybe it's some kind of mechanism in control
systems that occurs automatically ?

BP: I think it might be just a control system at the sequence level: if event or situation A occurs, the next thing must be behavior B. If we're predisposed to develop control systems, that's one way we can imitate a stimulus-response system. Tom Bourbon wrote a paper about that -- the way a control hierarchy can imitate either a stimulus-response system or a top-down cognitive system. [If anyone has a copy, you might post it]

BH: So I'm asking myself if we could turn this conversation into advantage for PCT and explain why the combination of feedforward and feedback occurs and how to use it the training to make players more efficiant and finally to win the game. That could by my oppinion make PCT usefull aplication. "Guessing" what is better won't do any advance.

BP: Right. The best thing to do is to stop arguing and get out on the playing field and test your ideas. I always do that if it's possible. Either that or try to devise a working simulation to test the idea. I don't take abstract arguments very seriously.

BH: I've heard some good theoretical back-ground for feed-forward and it seems to me quite exceptable from PCT view, but I must ask the member of CSG net if we can introduce our conversation. I respect the intimacy if others do. But if they don't, I don't either.

BP: You don't have to ask anyone's permission to introduce a new topic. The worst that can happen is that if nobody is interested in it, nobody will comment.

Best,

Bill P.

···

At 10:46 AM 12/19/2009 -0600, Boris Hartman wrote:

[From Bill Powers (2009.12.22.1605 MDT)]

BH: The most important thing I see till now is that we established the existance of feedforward and that combination with feed-back exists.

I've been away exercising to improve my lung function, and musing about this clear example of feedforward. Since I can detect no effect on my lung functions at all, I am an example of a feedforward system while doing that, at least at one level. There is no difference between a feedforward system and a feedback system except for the lack of a feedback effect through the environment on the variable that is giving rise to the behavior. Feedforward just means lack of feedback.

Every morning, I take a vitamin pill, which I have always thought was good for me. However, I can detect no effect whatsoever from taking the pill. There is no feedback from taking the pill, so this qualifies as a feedforward kind of behavior. At a higher level, of course, there is feedback from doing this because taking the pill gives me the perception that I'm taking care of myself, which matches a reference perception. What appears to be feedforward at a lower level turns out to be instigated by a feedback system at a higher level. And of course at lower levels the action (picking up pill, putting it in mouth, drinking, swallowing) is carried out by lower-level feedback control systems as usual.

Behavior, according to PCT, is caused in two ways. One way is for a disturbance to alter a perceptual signal and make it different from a reference signal. That produces an error signal which causes lower-order behaviors to occur. If those behaviors affect the disturbed perceptual signal in the direction that changes it back toward the reference state, the behavior is called control, and feedback (negative) exists. If the behavior has no effect on that perceptual signal, we call it a "response" and classify the system as a feedforward or stimulus-response system. There is no substantial difference between feedforward and feedback inside the behaving system. The difference is in the environment -- whether the action affects the same perception that appeared to cause the action.

The second way in which behavior is caused is for the reference signal to change while the perceptual signal is undisturbed. This change in the reference signal generates an error signal which causes the output function to act by changing the reference signals for lower-order systems or by operating muscles. That is clearly a feedforward connection if the action has no effect on the perceptual signal. However, the action might affect the perception so as to make it change in the same direction that the reference signal changed. That would make the system into a feedback control system. The feedforward path is the path from reference signal, through comparator to error signal, and through the output function to the action (whether a muscle action or the action of a lower control system or feedforward system). In the absence of feedback, the action takes place but has no effect on the perceptual signal in the system that responded to the change in the reference signal.

Again, the higher system that changed the reference signal might still be a control system if the action affects some other perception that the higher system is controlling. And the lower-order systems can still be feedback systems of the kind we talk about in PCT. Once more, the only difference between feedforward and feedback is the absence or presence of an external feedback effect on the perception involved.

I think it will be very hard to find any behavior which does not have effects on perceptions. How would a person even know that a behavior had taken place?
And since feedforward behavior would not alter any perception that led to the behavior, what would be the point of the behavior? A sensation of hunger might cause the muscles to blink the left eyelid, but so what? If the behavior had no effect on the hunger, it would certainly be called feedforward behavior, but why would anyone learn to behave that way?

Are there other examples of feedforward that don't fit the above cases?

Best,

Bill P.

···

At 10:46 AM 12/19/2009 -0600, Boris Hartman wrote:

[From Rick Marken (2009.12.22.1920)]

Bill Powers (2009.12.22.1605 MDT)--

I've been away exercising to improve my lung function, and musing about this
clear example of feedforward. Since I can detect no effect on my lung
functions at all, I am an example of a feedforward system while doing that,
at least at one level.

Every morning, I take a vitamin pill, which I have always thought was good
for me. However, I can detect no effect whatsoever from taking the pill.
There is no feedback from taking the pill, so this qualifies as a
feedforward kind of behavior...

Are there other examples of feedforward that don't fit the above cases?

Here a few:

Virtually any religious ritual: praying, human sacrifice, etc. These
actions have no feedback effect on the perceptions, like crop yield or
getting into heaven, that people are trying to control by doing them.

Presenting evidence to ideologues. Absolutely no feedback effect on
the controlled perception: what they say they believe.

Writing papers on PCT to change psychology;-)

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Rick Marken (2009.12.22.2010)]

Rick Marken (2009.12.22.1920)]

Bill Powers (2009.12.22.1605 MDT)--

Are there other examples of feedforward that don't fit the above cases?

Here a few:

Virtually any religious ritual:

Oops, sorry. You were asking for examples that _don't_ fit yours. My
examples do fit yours.

Never mind;-)

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Rick Marken (2009.12.22.2010)]

Rick Marken (2009.12.22.1920)]

Bill Powers (2009.12.22.1605 MDT)--

Are there other examples of feedforward that don't fit the above cases?

Here a few:

Virtually any religious ritual:

Oops, sorry. You were asking for examples that _don't_ fit yours. My
examples do fit yours.

Never mind;-)

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bruce Abbott (2009.12.13.0820 EST)]

Bill Powers (2009.12.22.1605 MDT) --

BH: The most important thing I see till now is that we established
the existance of feedforward and that combination with feed-back exists.

BP: I've been away exercising to improve my lung function, and musing
about this clear example of feedforward. Since I can detect no effect
on my lung functions at all, I am an example of a feedforward system
while doing that, at least at one level. There is no difference
between a feedforward system and a feedback system except for the
lack of a feedback effect through the environment on the variable
that is giving rise to the behavior. Feedforward just means lack of
feedback.

BP: Every morning, I take a vitamin pill, which I have always thought was
good for me. However, I can detect no effect whatsoever from taking
the pill. There is no feedback from taking the pill, so this
qualifies as a feedforward kind of behavior. At a higher level, of
course, there is feedback from doing this because taking the pill
gives me the perception that I'm taking care of myself, which matches
a reference perception. What appears to be feedforward at a lower
level turns out to be instigated by a feedback system at a higher
level. And of course at lower levels the action (picking up pill,
putting it in mouth, drinking, swallowing) is carried out by
lower-level feedback control systems as usual.

BP: Behavior, according to PCT, is caused in two ways. One way is for a
disturbance to alter a perceptual signal and make it different from a
reference signal. That produces an error signal which causes
lower-order behaviors to occur. If those behaviors affect the
disturbed perceptual signal in the direction that changes it back
toward the reference state, the behavior is called control, and
feedback (negative) exists. If the behavior has no effect on that
perceptual signal, we call it a "response" and classify the system as
a feedforward or stimulus-response system. There is no substantial
difference between feedforward and feedback inside the behaving
system. The difference is in the environment -- whether the action
affects the same perception that appeared to cause the action.

A system employing pure feedforward senses a disturbance to the variable in
question in advance of its effect on that variable, and acts in ways
designed to oppose that effect. (The state of the variable itself is not
sensed.) Maybe I'm missing them, but I'm not seeing any feedforward signals
in your examples. Although it is true that a pure feedforward system
operates in open-loop fashion, not all open-loop systems employ feedforward,
so the mere fact that a system is open loop does not automatically qualify
it as a feedforward system.

Pure feedforward systems can work well if the effects of the disturbance on
the variable to be controlled are predictable and there are no other
significant disturbances that go undetected and thus uncorrected. That's a
tall order in most cases, however; consequently a practical system is likely
to include both feedforward and feedback.

When I prepare to leave the house, I put on my coat and hat because I know
that it's cold outside and I don't want to freeze (feedforward). I don't
wait until I've been outside for awhile and my teeth are starting to chatter
before taking this action (feedback). But once outside, I may discover that
it's actually warmer than I thought it would be and, as I begin to heat up,
I take the coat off. I would venture to guess that most examples of
feedforward are like this. That is, feedforward is piggybacked onto a
negative feedback control system.

Another common example involving feedforward is classical conditioning. For
example, in eyeblink conditioning, a tone immediately precedes a puff of air
directed onto the cornea of the eye. The puff of air stimulates sensors in
the cornea, and ordinary negative feedback control acts to close the eyelid
to counteract the effect of this disturbance to the corneal sensations. But
after a sufficient number of pairings between the signal predicting the puff
of air (the tone) and the air puff, the tone becomes effective at eliciting
the same closure of the eyelid, in advance of the air puff. The tone has
become a feedforward signal and the result is to further limit the effect of
the air puff on those corneal sensations.

In both examples, learning was involved. If the current actions based on
feedforward are not doing their job effectively, the individual may begin to
vary those actions until an effective solution is found. Thus the
feedforward aspect of the system is actually involved in a longer-term
reorganization; the actions that result from feedforward are varied within
the negative-feedback reorganization control system, helping to assure that
feedforward, if employed, will usually be effective in helping to control
the controlled variable.

Bruce

···

At 10:46 AM 12/19/2009 -0600, Boris Hartman wrote:

[From Rick Marken (2009.12.23.0820)]

Bruce Abbott (2009.12.13.0820 EST)--

A system employing pure feedforward senses a disturbance to the variable in
question in advance of its effect on that variable, and acts in ways
designed to oppose that effect.

Sounds like pure feedback control to me; the perception under feedback
control is a perception of the disturbance. The feedforward part is
the imagined effect of your control of this perception on some other
variable. Your example of putting a coat on before going out in the
cold involves both feedback and feedforward control. The feedback
control is putting the coat on when you hear that the weather outside
is cold; the perception under feedback control is that of the
relationship between the whether report and putting on (or not) the
coat. The feedforward control is the imagined effect of controlling
this variable on your perception of how cold you are when you go
outside.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bill Powers (2009.12.23.0812 MDT)]

Bruce Abbott (2009.12.13.0820 EST) --

BA: A system employing pure feedforward senses a disturbance to the variable in question in advance of its effect on that variable, and acts in ways
designed to oppose that effect. (The state of the variable itself is not
sensed.) Maybe I'm missing them, but I'm not seeing any feedforward signals
in your examples. Although it is true that a pure feedforward system
operates in open-loop fashion, not all open-loop systems employ feedforward,
so the mere fact that a system is open loop does not automatically qualify
it as a feedforward system.

BP: Feedforward includes more than sensing disturbances. For example, if the reference signal is coupled directly to the output function (as well as entering the comparator), its effects "feed forward" to produce action that isn't based on an error signal (that is, feedback can't affect it at that level). Of course you could say that the reference signal is a disturbance affecting the control loop.

In my vitamin example, the input to the feedforward system is a difference between how I feel and how I want to feel. However, taking the vitamins makes no difference in how I feel, so there is no feedback.

Feedforward exists when there is an action apparently intended to affect something, but no feedback indicating whether that effect, or any effect, occurred. And of course feedforward exists if there is a response to perception of a stimulus, but the response has no effect on the perception.

BA: Pure feedforward systems can work well if the effects of the disturbance on the variable to be controlled are predictable and there are no other
significant disturbances that go undetected and thus uncorrected. That's a
tall order in most cases, however; consequently a practical system is likely
to include both feedforward and feedback.

BP: Well, yes, it is a tall order -- you might as well say "... can work well if everything in the environment is arranged so it can work well." I wouldn't dispute that. Unfortunately, that usually takes a lot of arranging as you note: a tall order.

The temperature-compensated pendulum clock was a very nice application of feedforward. A column of mercury (as in a thermometer) mounted on a clock pendulum would expand upward, moving mass toward the pivot to compensate for the increase in length of the other material in the pendulum as temperature increased. That was in 1721 (good old Google). It produced clocks accurate to one second a week instead of one second per day.

What is generally overlooked is that there must be a very good negative feedback control system in operation before feedforward can be made to work at all. The negative feedback system does sense the variable that is to be controlled, and over time makes finer and finer adjustments in the feedforward system -- the amount of mercury in the compensated pendulum -- until no error can be detected. If great care is taken with this adjustment, and the feedforward system is protected against any disturbances, and there is no wear in the mechanism or any change of properties such as lubrication, the feedforward system will work for a while, if not too much accuracy is demanded of it. When I speak of the negative feedback control system, you know who I am talking about.

BA: When I prepare to leave the house, I put on my coat and hat because I know that it's cold outside and I don't want to freeze (feedforward). I don't
wait until I've been outside for awhile and my teeth are starting to chatter
before taking this action (feedback).

BP: True. It took your mother a while to persuade you to put your coat on, but after you had gone out a few times in winter without it, you realized that you would freeze, and made putting on a coat part of the going-outside sequence.

This is certainly not feedforward. If it were feedforward, you would never know whether putting a coat on had any effect on feeling warm or cold. How do you know you will freeze if you don't put on a coat? By freezing, and not liking it. There is feedback; it's just delayed a bit and averaged over a number of trials. A feedforward system has no feedback; its input is not affected by its output.

BA: But once outside, I may discover that it's actually warmer than I thought it would be and, as I begin to heat up, I take the coat off. I would venture to guess that most examples of feedforward are like this. That is, feedforward is piggybacked onto a negative feedback control system.

BP: That's more likely. But I have yet to find a case of control with any significant ability to resist disturbances or correct errors where feedforward can lead to much improvement of control. Feedback control is always faster and more accurate, though feedfoward can, in principle, fill in the gap when there are transport lags and when disturbances can occur almost instantaneously. In our tracking experiments, for example, feedforward might start the output changing during the 0.13 seconds between the onset of a step-disturbance and the time it takes for the perceptual system, comparator, and output function to begin to act. Of course you'd have to find components for the feedfoward part of the system that didn't also require 0.13 second to begin to act after sensing a sudden disturbance. And if this were a model of the nervous system, eventually you'd have to find the places in the nervous system where this superfast, superaccurate computation took place.

BA: Another common example involving feedforward is classical conditioning. For example, in eyeblink conditioning, a tone immediately precedes a puff of air directed onto the cornea of the eye. The puff of air stimulates sensors in the cornea, and ordinary negative feedback control acts to close the eyelid to counteract the effect of this disturbance to the corneal sensations. But after a sufficient number of pairings between the signal predicting the puff of air (the tone) and the air puff, the tone becomes effective at eliciting the same closure of the eyelid, in advance of the air puff. The tone has become a feedforward signal and the result is to further limit the effect of the air puff on those corneal sensations.

BP: I would say that the perceptual input function has been reorganized to add the tone to its definition of the controlled variable. It perceives and controls "puff + tone --> food" instead of just "puff --> food". This is not an open-loop system: the reaction to "puff + tone" definitely feeds back to affect "puff + tone." A disturbance of either individual variable will result in action that opposes the change in the composite variable. And of course the initial reaction to "puff" is probably a reaction to the drying effect of the puff, with the blink restoring the wetness. Resistance to cooling also might play a part.

I expect Dick Robertson to chime in here with his observations showing that the unconditional eye-blink reflex is not all that myth-makers would have us believe about it.

BA: In both examples, learning was involved. If the current actions based on
feedforward are not doing their job effectively, the individual may begin to
vary those actions until an effective solution is found.

BP: An example of this is filling a sink with warm water before washing your hands. After long experience you learn how much to turn the faucet handles to get the temperature about right, and you don't need to keep feeling the water temperature. Infrequent sampling is good enough, the final error being small enough to correct with a small addition of straight hot or cold water. So you can multitask.

Note, however, that a pure feedforward system can't learn, because it never senses the effects of its output actions. It has to be adjusted by some other system, a negative feedback system that can sense the effectiveness of the solution and adjust the feedforward system's parameters as required. This is similar to the way we assume the reorganizing system adjusts the parameters of a control system.

Thus the feedforward aspect of the system is actually involved in a longer-term reorganization; the actions that result from feedforward are varied within the negative-feedback reorganization control system, helping to assure that feedforward, if employed, will usually be effective in helping to control the controlled variable.

Yes, we agree on this.

I'll admit that all this is a reasonable thing to imagine if you'll admit that we're both imagining it. What we really need here is a model of a negative feedback control system that will fit real behavior better if we add feedforward to it. So far I haven't found one: it would be pretty hard to get the tracking model to fit the behavior any better, and any improvement that did occur would be extremely small, hardly worth the bother in most situations. My few attempts to add feed-forward actually made performance worse, but we can't conclude that this would always happen.

At the moment, in my book feedforward is just a conjecture. I really don't understand why anyone is interested in it, but that's just me.

Best,

Bill P.

[From Bill Powers (2009.12.23.1625 MDT)]

Bruce Abbott (2009.12.13.0820 EST) --

BP earlier: I would say that the perceptual input function has been reorganized to add the tone to its definition of the controlled variable. It perceives and controls "puff + tone --> food" instead of just "puff --> food".

BP now: I don't know why I said "food" here -- some idea that has slipped away, if any. Food obviously has nothing to do with blinking, unless you get it in your eye. No, that couldn't have been it. Oh, well.

Best,

Bill P.