Feedforward doesn't care

[From Rick Marken (931104.1100)]

I said:

I'm sure Bill Powers will "do his dooty" and point out (for the
400,000th time) all the problems with feedforward models of
behavior.

Hans Blom (931104) replies:

Rick, didn't I point out the disadvantages of feedforward control
sufficiently?"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""
""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""From @vm.utcc.utoronto.ca:owner-csg-l@VMD.CSO.UIUC.EDU Thu Nov 4 17:54:56 1993
Return-Path: <@vm.utcc.utoronto.ca:owner-csg-l@VMD.CSO.UIUC.EDU>
Received: from vm.utcc.utoronto.ca by ben.dciem.dnd.ca (4.1/SMI-3.2)
  id AA04785; Thu, 4 Nov 93 17:54:38 EST
Message-Id: <9311042254.AA04785@ben.dciem.dnd.ca>
Received: from vm.utcc.utoronto.ca by vm.utcc.utoronto.ca (IBM VM SMTP V2R2)
   with BSMTP id 1546; Thu, 04 Nov 93 17:53:10 EST
Received: from UTORONTO.BITNET by vm.utcc.utoronto.ca (Mailer R2.10 ptf000)
with BSMTP id 2546; Thu, 04 Nov 93 17:53:04 EST
Sender: "Control Systems Group Network (CSGnet)" <CSG-L@UIUCVMD.BITNET>
X-To: CSG-L%UIUCVMD.bitnet@vm42.cso.uiuc.edu
Status: R

From Tom Bourbon [931104.1345]

[Hans Blom, 931104]

(Rick Marken (931103.1200))

I'm sure Bill Powers will "do his dooty" and point out (for the
400,000th time) all the problems with feedforward models of
behavior.

Rick, didn't I point out the disadvantages of feedforward control
sufficiently? Do you think that I was unfair or unrealistic in positing
that there are circumstances where they DO work?

Hans, I think the problem was not that you said there are circumstances
where ballistic behavior might work, but that your examples were not
convincing to Rick, or to Bill as he showed in his long reply, or to me.
Each of us has also said ballistic actions would work in a disturbance-free
world. The problem is, we don't live in that kind of world. Pre-planned
ballistic actions do indeed produce reliable results in the clean rooms
where wafers are manufactured, but that is not our home and robots do that
work.

Hans:

Stimulus-response behavior, motor or mental, often takes over
after learning is complete

Rick:

How long does this takeover last after a person has completely
learned to toss a ball at a target in a highly reliable
environment? How come when Gary Cziko applies a disturbance
to the throwing arm immediately after this learning has taken
place the S-R behavior seems to relinquish its "take over" (and
the throwing arm resists the disturbance)?

Hans:

How about this addition: feedback controlled behavior, motor or mental,
again takes over as soon as we discover that the stimuli that our senses
receive do not correspond anymore to well-known patterns that we have
developed adequate reaction patterns to. This presupposes a "sensor" or
"recognizer", at some level, that discriminates familiar from unfamiliar,
or maybe danger from safety or some such. There is strong evidence that
such a "device" exists and is well developed already in small babies.

Tom:
Hans, that would need to be an extremely efficient sensor-recognizer system.
The kinds of "corrections" Rick described occur very rapidly and the
results are most often "just right," leading to continuous production
of the person's intended results. And that kind of correction seems to
occur all the time.

Hans:

In this sense, you might say that feedforward is for experts and
that feedback is for beginners :slight_smile:

Rick:

Yes, a wind up doll is a real expert, all right. And it can walk
in the dark, too! Imagine that :wink:

Hans:

Aren't we all wind up dolls in a great many respects?

Tom:
No.

(Bill Powers (931103.1340 MST))

Bill:

4. Under HPCT, it is not necessary that a perceptual signal be
available to consciousness or awareness in order for it to play a
part in feedback control.

Hans:

Sure. I do not assume consciousness or awareness in feedforward systems
either. In many respects, feedforward responses can be compared to uncon-
ditionally evoked reflexes that run off once started. This typifies the
behavior of most of us under most circumstances, doesn't it?

Hans, this comment reflects a radical difference in our understandings of
what we see in behavior. Certainly actions that are commonly called
"reflexive" can be seen in much of what we do, but even the concept of a
blind unconditioned reflex, as an explanation of behavior, is called into
question when you recognize that the reflex varies just as needed to
maintain some specified physiological or perceptual state. Reflexes are not
ballistic actions.

Bill:

6. The range of situations under which feedforward control can
perform usefully is limited.

Hans:

Yes, I thought I mentioned that.

Tom:
You did. That is not the point Bill questioned. He was speaking of your
remarks when you described examples of human actions that you said were
instances of feedforward control. He questioned your interpretation of
those actions as ballistic. So do I.

Bill:

7. The kinds of variables amenable to feedforward control are
those that naturally change slowly, over seconds or longer, even
in the presence of normal disturbances.

Hans:

Not necessarily. We want predictability, not constancy. After you have
thrown a football, your idea of where it might land might be sufficiently
accurate to continue the play of the game with other concerns in mind. Of
course, the ball might hit an overflying bird, be picked up by a UFO, or
hit a shoe thrown by a female fan. But those are things that you usually
can disregard.

Tom:
Hans, in American professional football (at both levels -- college and the
National Football League) a quarterback must do more than produce
predictable results when he passes the ball. And the potential sources of
failure if he tries to perform ballistically are not nearly so far fetched as
the examples you gave. The unavoidable variability of the environment, and
of his own muscles, posture and the like, assure that a ballistic
quarterback would soon be looking for another job. I live in Houston -- I
know whereof I speak!

Also, you did not address Bill's distinction between continuous and sampled
control. Many of the actions you describe as ballistic are nice examples of
sampled control.

Bill:

9. An adaptive compensatory system can't learn to control
"better" unless its outer loop is a feedback control system.

Hans:

Right. But once it has learned, it might not need feedback anymore, or
only occasionally, or only under special circumstances.

Tom:
Occasionally, as in moment to moment. Special conditions, as in while
living in a variable body in this variable world.

Bill:

While, as you say, apparent cases of feedforward abound, a great
many of them are based on superstition and after-the-fact
"prediction."

Hans:

Skinner talked about "superstitious learning". Now I could easily show you
that in any model of learning that I am aware of, ALL learning is necessa-
rily superstitious in the sense that it takes into account only the evid-
ence at hand -- a more or less reliable correlation between action and
outcome. If pecking at that bar is effective to keep the grain coming, do
it -- even though, unknown to you, the grain comes at random intervals
that cannot be influenced by pecking. It IS a solution, although a sub-
optimal one. But how would YOU act if in all previous experiments your
only method of control was pecking at bars? Finding _optimal_ solutions is
difficult in any walk of life, even for a pigeon.

Tom:
Finding optimal solutions is indeed difficult. Fortunately, living control
systems don't need to do that, so they don't do it -- unless, of course, they
also happen to be control engineers who earn their livings building optimal
controllers.

Bill:

I can agree with almost everything you say about the nature and
characteristics of feedforward control. I just can't agree that
it accounts for very much of human behavior.

Hans:

OK, thank you. We both think that feedforward / stimulus-response control
might work. We just differ in our estimate of how much it accounts for
human behavior.

Tom:
Yes, it might work. And, yes, we differ in our estimates of how much it
accounts for human behavior. Those estimates range from, "This
typifies the behavior of most of us under most circumstances," to "this has
very little to do with human behavior."

(Bill Powers (931103.1530 MST))

Your use of the term "reference signal" in a completely open-loop
system is somewhat strange to me. Why not just call it the
"input?"

Hans:

I don't care much for names. "Input" is fine with me, if you insist. I
just wanted to stress the similarity with a reference signal.

Tom:
Say what? Bill was implying something more than a difference in name.

Hans:

ref -------- action
------>|+ |------> world
  --->|- | |
  > -------- |
  > -------- |
  ----|model |<---
      --------

where the model is purely internal; it only "knows" our actions / output,
not their effects on the outside world. In fact, the model PREDICTS those
effects given the actions and thus does something like your "imagination
mode". Am I clear enough or are all those transformations between methods
of drawing the control diagram too confusing? Mathematically / functional-
ly, they can be shown to be equivalent.

Tom:
Is there a reference signal in the "model" and if so what is its relation
to the one we would call an "input?"

(Tom Bourbon [931103.1641])

You say that your actions during your close encounter with the metal
obstruction exemplify feedforward and did not entail feedback. Did you
close your eyes (to eliminate visual perceptions), release the wheel and
remove your feet from the pedals and the floor (to eliminate kinesthetic,
proprioceptive, and tactual perceptions), and cover your ears (to
eliminate auditory perceptions)? Do you do similar things to eliminate
all other perceptions when you walk through the darkened room at night?

Tom, you must discriminate those perceptions that are used to provide the
feedback information from all the other old perceptions that we have but
that are not relevant to the task at hand. When I navigate my way through
a darkened room, I have a great many perceptions, but if the floor is even
-- no female shoes lying around -- I have no indication as to where I am,
except maybe from some things other than my immediate perceptions, such as
an internal model.

We agree on this.

This is a very different statement from saying that when we walk through a
darkened room we go ballistic and do not use feedback control. We can
achieve perceptual control of felt movements relative to imagined or
remembered movements, or relative to remembered or imagined perceptions of
an environment we cannot now see. The absence of visual perceptions in no
way requires a shift from feedback control to ballistic actions.

In the dark, you do have indications of where you are in the room as it is
remembered or imagined. You feel yourself move through your model of the
room.

Until later,

Tom

···

Date: Tue, 2 Nov 1993 15:41:16 CST
Reply-To: "Control Systems Group Network (CSGnet)" <CSG-L@UIUCVMD.BITNET>
From: Tom Bourbon <tbourbon@HEART.MED.UTH.TMC.EDU>
Subject: Re: feedforward, once again
To: Multiple recipients of list CSG-L <CSG-L@UIUCVMD.BITNET>