Questions about an incident

CSGnet Archive CSG1993
1

[From Bill Powers (931104.1900 MST)]

Hans Blom (931103,04) --

We've come down to "yes it is" and "no it isn't", so it's time to
get more specific.

Let's use some real data:

It was in a shallow curve to the right that the direction of
the car did not follow the command of the steering wheel
anymore. I was in the left- most of three lanes, and the car
was -- very, very slowly, it now seemed -- coming closer to the
left side of the road, where an unappetizingly sturdy metal
construction would meet the car if control was not regained in
time. Thoughts about what might happen raced through me. Most
probably the car would ricochet off the dividing barrier and
crash into cars on the rightmost two lanes; traffic was rather

heavy at this time of the after- noon.

No, I did not turn the steering wheel more and more to the
right in order to keep on course. Instead, I kept the steering
wheel in the same ineffective position.

I'm sure you can remember this rather memorable incident quite
well, so perhaps you could tell me some more details about it.

How did you know that your hands were holding the steering wheel,
and how did you know that you were keeping it in the same
position?

I braked, but just a little.

Can you remember how you knew that you were braking, and that it
was "just a little?"

Little by little the car came closer to the barrier.

How did you know that the car was coming closer to the barrier?

These are meant to be simple questions. I'm just trying to find
out how you knew of these things that were going on.

···

---------------------------------------------------------------
Best,

Bill P.

2

From Tom Bourbon [931105.0931]

[From Bill Powers (931104.1900 MST)]

Hans Blom (931103,04) --

We've come down to "yes it is" and "no it isn't", so it's time to
get more specific.

Let's use some real data:

It was in a shallow curve to the right that the direction of
the car did not follow the command of the steering wheel
anymore. I was in the left- most of three lanes, and the car
was -- very, very slowly, it now seemed -- coming closer to the
left side of the road, where an unappetizingly sturdy metal
construction would meet the car if control was not regained in
time. Thoughts about what might happen raced through me. Most
probably the car would ricochet off the dividing barrier and
crash into cars on the rightmost two lanes; traffic was rather

heavy at this time of the after- noon.

No, I did not turn the steering wheel more and more to the
right in order to keep on course. Instead, I kept the steering
wheel in the same ineffective position.

Bill then asked some specific questions about how you came to know about
the occurrence of each event you described. His questions go directly to
the points I addressed more indirectly, and perhaps not obviously, in my
reply (Tom Bourbon [931103.1641]) to your original post on the incident. I
asked if you released the wheel, eliminated as much contact as with the
automobile as you could, and closed your eyes. My obscure purpose, and
Bill's direct one, is to ask how you knew each part of the incident was
occurring at the time, how you knew you were taking the actions you
described in your post, and how you knew the consequences of your actions.
Your answers to those questions might help us better understand where our
use of control theory is similar to and different from yours.

Until later,

Tom

3

[Martin Taylor 931105 12:10]
(Hans Blom, Bill Powers, Rick Marken, Tom Bourbon--various postings)

Sorry to butt in, but I perceive a classic case of mutual misunderstanding
in the discussions of "feedforward."

I read Hans as saying that in some conditions, conditions that are not
uncommon, there may exist within the hierarchy some ECS that is deprived
of perceptual input and nevertheless manages to produce output that
maintains the CEV at a tolerable level (one that would be produce only
a small error if the perceptual signal were available). It can do this
because the world as seen from its viewpoint is reasonably stable and
predictable. (Its "world" includes all of the lower ECSs to which it
sends output and from which it receives input. Part of the stability
of that world is due to the feedback control performed by lower systems.)

I read Bill, Tom, and Rick as imputing to Hans the claim that all ECSs
in the hierarchy are simultaneously deprived of sensory input, and that
this condition happens often. They ask questions such as:

(Bill Powers (931104.1900)

How did you know that your hands were holding the steering wheel,
and how did you know that you were keeping it in the same
position?

I braked, but just a little.

Can you remember how you knew that you were braking, and that it
was "just a little?"

These questions seem to me irrelevant to the point Hans was trying to
raise.

Hans, on the other hand, ignores the normal function of reorganization
in a highly non-linear world, which is to provide mechanisms for dealing
with the situation when the world's behaviour changes (the wheels no
longer grip the road). I think it was Bill who pointed this out.

It's clear, and I think generally agreed, that some "control in imagination"
is normal, particularly in the higher layers of the hierarchy. To function,
there must be something that looks like a model of how the world behaves,
whether that model be explicit or implicit in the structure of the hierarchy.
"Feedforwrd" involves using that model in lieu of the absent perceptual
signal, as I believe Hans has previously said. And as Lang and Ham showed
as long ago as the mid-50's even a linear one-level control system with
normal sensory input can work a lot better with the inclusion of a world
model than without it.

(Bill, do you have any comment on that paper, which I sent you some time
ago?)

Martin

4

[Martin Taylor 931106 20:45]
(Tom Bourbon 931105.1328)

And as Lang and Ham showed
as long ago as the mid-50's even a linear one-level control system with
normal sensory input can work a lot better with the inclusion of a world
model than without it.

Does this mean, as an example, that Lang and Ham would produce a model that
accounts for more than the 99.2% of the variance, or greater, accounted for
when two unsophisticated, one-level, world-model-free PCT models predict the
results of interactions during a task with 2 controllable dimensions? How
much more of the variance would be accounted for if each of the PCT models
were equipped with its own (predictive?) world model?

They (and I) were not talking about *modelling* control, but about
*controlling*. The simple linear model without feedforward in their
demonstration brought its perceptual signal to its reference in a normal
exponential approach (looking at the figure), whereas the feedforward
plus feedback system provided a graph that looked much more like two
linear segments--a fast slew to bring the perception to the reference,
with no overshoot when it got there.

···

********
                            ****
                         **
                      **
                    *
                   *
                          without feedforward

                   **********************
                  *
                 *
                *
               *
              *
                         with feedforward

I'm sure modern control engineers must be able to do much better than
was possible 40 years ago.

This doesn't say that biological systems do the same, but it is a usual
presumption that evolution is pretty good at finding successful tricks,
so it would not be surprising if they do. That doesn't mean that a
feed-both-ways system would predict better than a simple feedback system
does.

I would have posted the reference to Lang and Ham, but I seem to have mislaid
the reprint some time in the past week. There are some copies floating
around DCIEM, and Bill's got one, so the reference can be recovered if
it is wanted.

The Lang-Ham system can be described quite easily. They called it a
"conditional control system." Think of the usual ECS. Now put a box
between the reference input and the comparator, called "A". Take
another line from the reference input through a box called "B" and
add it to the error before feeding the result to the output function.
That's it. "A" and "B" between them represent the world model, and
as I understand it, if the world behaves like the model, there will be
no error, and the output would be driven (correctly, as it happens)
by the planned signal. But if the plan fails, there will be error, and
the output will be driven in the normal way, as in the simple ECS.

I would be able to say this with less hesitance if I had the paper in
front of me. I hope that one of our control engineers can develop a
more modern view of the same idea.

Martin

5

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

Martin:

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

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.

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.

It's a constant theme in what used to be called "bionics" before Hollywood
perverted the term: See how one might solve a problem mechanically, see if
you can see whether Nature used that solution; See what problem Nature is
solving and how, see if you can implement the solution mechanically. There's
an ongoing interplay (it is how a lot of sonar systems were developed, and
how we came to understand how bats use the weird signals they sometimes
emit.)

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 am in no way trying to pre-empt the use of the imagination loop as
ordinarily conceived (reconnecting output to perceptual input through
a model). That's used under different conditions, when there is no
intended output, though its model might well be precisely the A() function
in the Powers-modified Lang-Ham system. The same functional entity
might serve both purposes--feedforward and imagination.

Check the figure in my 981108 17:15 posting of a few minutes ago, and
think of the A() input as coming from a higher-level ECS's output, and
the A() output as going to that higher ECS's sensory input. (I know
this violates the structure of many-to-many connection between levels,
so it has to be wrong in detail. But something of the kind might work,
such as having the A() functions from many higher ECSs combined in the same
way as outputs are combined to form reference signals. This has to be
thought out before being proposed seriously).

···

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

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

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.

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.

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

Hans
suggests that feedforward is necessary. Martin agrees, to the extent
that a Lang-Ham world-model-based control model could control "better than"
a simple single control loop.

To clarify (also see my 931108 17:15), The Lang-Ham model doesn't *control*
better, if by that you mean "resist disturbances." It does that identically
the same as the simple single control loop. What it does is to *control*
better the shift of the perceptual signal to a new reference level.

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.

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.

Down with all-or-none thinking!

Martin