Feedforward: Out of control

CSGnet Archive CSG1993
1

[From Rick Marken (931109.2200)]

The feedforward discussion is getting completely out of control.
I have no idea what people mean by the term "feedforward" anymore.
I detect at least three distinct meanings:

1. The direct connection between reference and output in the
H-L model is being called "feedforward". But this is just a
connection in a closed loop feedback control system. As Bill
Powers just showed (931109.1330 MST) the H-L model connections
produce virtually the same behavior as the canonical control
system. The "feedforward" connection in this model is not
having its influence "open loop". The system still controls its
perceptual signal -- it's a feedback control system.

2. Hans Blom (eg. 931110) is using "feedforward" to refer to any
behavior that appears to be the same whether it is done with the
eyes open or closed. Thus, if it "appears" to be done without sensory
input then it is feedforward. Given this superficial (literally)
definition of feedforward behavior, there seems to be virtually no
way to convince Hans that what he might actually be seeing is the
side effects of closed loop control; I have still not heard how
Hans knows that a behavior is under "feedforward control" other
than by just looking and seeing that the person is "doing it
with their eyes closed".

3. I was under the impression that feedforward refers to a means of
determining behavioral results through computation of the outputs
that produce those results. This would mean that the result produced
by a feedforward process is NOT under control -- by definition. Thus,
the first step in evaluating the value of a feedforward explanation
of a behavioral phenomenon would be to determine whether or not the
phenomenon involves control. If it DOES involve control then no
feedforward explanation is possible; if it DOES NOT involve control
then some feedforward model will do the trick. It is entirely possible
that some behavioral results are NOT controlled (even if they are
PART of control loops). But this must be determined by experimentation
-- not ratiocination.

The square wave tracking task can serve as a simple example of the
problems I am have with the term "feedforward". Suppose you ask
a person to keep the cursor aligned with the square wave movement
of the target, even if the display disappears. It is easy to show
that the target/cursor relationship is under control while the
display is visible -- disturbances to the cursor's position
are resisted, as are disturbances to variables that influence the
cursor's movement, like hand position. When the display disappears,
it is easy to show that the target/cursor relationship is no longer
under control -- disturbances to the cursor are not resisted. But
disturbances to the hand still ARE resisted. This means is that
hand position is still being controlled; it also means that cursor
movements are being produced open loop (by a feedforward process).
The model of this feedforward process is rather simple -- the invisible
cursor movement is just a side effect of the control of hand (and, hence,
handle) position. This "feedforward-as-side-effect-of-control" model
would describe the results of this experiment perfectly -- predicting
the invisible cursor position to a tee -- including it's position when
disturbed. This is a true feedforward model -- but pretty uninteresting,
it seems to me, since it says nothing about the nature of the
processes IN THE PERSON that produced this result (the feedback model
handled that).

Before we continue with the discussion of feedforward, can we agree
on what we mean by the term. I propose that only meaning (3)
above make sense -- meaning (1) is the same as "feedback" and (2) is
the same as "feedforward is whatever it looks like to me". At least
with (3) we have an objective means of determining whether the value of a
variable is the result of a feedforward process: such a variable is one
that is NOT under control.

Best

Rick

2

[Martin Taylor 931110 10:45]
(Rick Marken 931109.2200) and below (Bill Powers 931109.1330)

Rick asks about what he sees as three different concepts of "feedforward."
I must confess that I do not see these as different concepts, but maybe they
are.

1. The direct connection between reference and output in the
H-L model is being called "feedforward". But this is just a
connection in a closed loop feedback control system. As Bill
Powers just showed (931109.1330 MST) the H-L model connections
produce virtually the same behavior as the canonical control
system.

(See comment below on Powers 931109.1330)

The H-L connections were INTENDED to produce EXACTLY the same behaviour
as the canonical control system when the perception is being controlled
against disturbance. The "feedforward" provides an entirely different,
and separately constructed, dynamical behaviour in respect of a change
of reference. It has nothing to do with "control" in the sense we usually
take it--resistance to disturbance. It permits a first shot at getting
the perception to a new reference level, based on long-term stability
in the environment's reaction to output signals.

2. Hans Blom (eg. 931110) is using "feedforward" to refer to any
behavior that appears to be the same whether it is done with the
eyes open or closed. Thus, if it "appears" to be done without sensory
input then it is feedforward.

As I read Hans's original posting on this, a reading confirmed by Hans,
he was talking about the action output of ONE ECS in a hierarchy when
sensory input was blanked out TO IT. Under those conditions, a "canonical"
ECS will provide a large step output unless the reference level for
its perceptual signal is whatever the PIF does with zero inputs. Hans
argued that the output would be maintained at some level, or varied
according to some model of what the world would be. An ECS whose
sensory input is blanked out CANNOT control, but it is still in the
hierarchy, even if its gain is somehow set to zero when its sensory
input vanishes.

I see this kind of "feedforward" as essentially identical to type 1.
The output is more or less appropriate to bring the perceptual signal
somewhere near its reference level, provided the environment is
reasonably stable and in the absence of major disturbance. To do this
is better than either zeroing the output or acting as if the "true"
perceptual signal is zero (or whatever the PIF does with zero inputs).

3. I was under the impression that feedforward refers to a means of
determining behavioral results through computation of the outputs
that produce those results. This would mean that the result produced
by a feedforward process is NOT under control -- by definition.

Right. Both the other constructs agree with this. I don't think
anyone has suggested otherwise, though I can see that you might have
implied it from some of the wordings used.

...the first step in evaluating the value of a feedforward explanation
of a behavioral phenomenon would be to determine whether or not the
phenomenon involves control. If it DOES involve control then no
feedforward explanation is possible;

Substitute "is complete" for "is possible." Then I'll agree with you,
and so, I presume, would Hans.

Before we continue with the discussion of feedforward, can we agree
on what we mean by the term. I propose that only meaning (3)
above make sense -- meaning (1) is the same as "feedback" and (2) is
the same as "feedforward is whatever it looks like to me". At least
with (3) we have an objective means of determining whether the value of a
variable is the result of a feedforward process: such a variable is one
that is NOT under control.

Whether we can agree is a matter for experimentation. I propose that all
three meanings are the same, and that in particular circumstances
feedforward may matter not at all, to some degree, or may be the only
way any half-way sensible output can be produced. In none of these cases
is feedforward related to "control" against disturbance. Only perceptual
control, which means a closed loop feedback system, can deal with disturbance.

ยทยทยท

============
(Bill Powers 931109.1330)

I buy your much improved analysis for the optimum values of the A() function
as 1/EP and B = 1 in the original Lang-Ham configuration. But when you
put a revised A() function A=1/E inside the loop, you lose somthing. It
is the same if you look only at the dynamics of the reference->CEV pathway,
but the dynamics of the perceptual control loop have changed. The point
of the L-H configuration was to separate these two dynamical issues, because
the environment constrains the possible dynamics of the loop in a way that
it does not constrain the open-loop ref->CEV path.

Consider a CEV that imposes a fixed transport delay. If the rest of the
loop (PIF, comparator, output function) is wide-band, the loop will oscillate.
That can be eliminated by putting a comb filter somwhere in the loop, if
it is well tuned to the inverse of the delay. This solution will not work
if the delay is variable. You have to restrict the loop bandwidth. But
in either case the A() function can be wide-band if it is outside the loop,
not if it is inside.

I think it likely that your solution of putting A inside the loop while
setting B = 1 + 1/PEG might solve the problem by putting effect of the comb
filter into the reference signal as a compensatory additive waveform, but
it looks a bit dicey, since if there is any error (or worse, uncertainty)
in E, the added filter will introduce its own oscillations as a forced
function into the loop. At least so my intuition says.

A properly designed system with pure feedback can do just as well
as a feedforward system, and its operation is not materially
affected by reasonable changes in its own output parameters or of
the external feedback loop.

If my argument above is correct, this is not so. But then, the L-H
configuration does not SUBSTITUTE feedforward for feedback. It provides
a feedforward addition to jump-start the error-correction process that
is performed by the feedback system after a change in reference level.

So far I have seen no advantages of feedforward over feedback,
and there seem to be a number of obvious drawbacks.

The advantages I perceive are not of feedforward over feedback, but of
joint feedforward-and-back over pure feedback. Those advantages are, as
Lang and Ham say, that one can separately optimize the response to a
reference step and the response to a disturbance step, without losing
the ability of feedback to correct for perceptual errors, whether they
be induced by disturbance or by miscomputation of the reaction of the
world to changes of output. Both optimizations can be done by the same
kind of reorganization (which, if the environment is stable enough in its
behaviour, could be on an evolutionary time scale).

Martin