model-based behavior

[From Bill Powers (950509.0815 MDT)]

Bill Leach (950509.01:42 U.S. Eastern Time Zone) --

RE: world-model-based control

The idea behind world-model-based control is fairly simple, although
implementing it can become complex. If you have a mental world-model
which contains ALL the properties of the environment relevant to
controlling a particular variable, then an output signal that drives the
world-model can also be applied to the environment, and the two should
behave in the same way.

However, the basic defect in this approach is a practical one: it isn't
really possible to model the environment in all the revelant aspects. A
true world-model would, for example, have to include all the inverse
kinematics and dynamics of the body and limbs, plus all the physical
laws that describe effects of body and limbs on other objects in the
environment, plus the physics of the affected objects, plus world-models
of the perceptual system through which we monitor effects of behavior.
And even if that could be done, we would still have the problem of
unpredictable disturbances that are left out of the world-model. In
order to use world-model-based control to accomplish umbrella-opening-
to-keep-the-rain-off, we would need a successful weather-predicting
world-model.

In world-models like the one Hans submitted, most of these problems are
bypassed. If there is a signal u that affects the world-model, it also
affects the real world -- but nothing is said about the means by which
it affects the real world, for example the properties of the actuators
(which the real world-model in a real organism would also have to
contain). Likewise, on the input side, there is a perceptual function
which is represented only as a noise generator adding to the effect of
the controlled variable on an internal representation. In a true world-
model, the model of the external world would have to include the
properties of the sensors -- for example, the lens of the eye, the
retinal computations, and any higher computations required to convert
the optical inputs into measures of variables like position or velocity.
Of course nobody can include such details in a model (anybody's) that we
can actually program and run, but we have to consider these details in
judging the practicality of any proposed model of behavior. The concept
of world-model-based control necessarily implies that all these details
must be present in the organism's world-model; I am dubious about that.
The real system is not allowed to make the simplifications and
approximations that we make in our computer models.

One way around this problem is to say that world-models come into play
only in the higher levels of organization, with the lower levels being
real-time negative feedback control systems that don't need all the
detailed modeling. At the lower levels, a limb can be positioned just by
varying a reference signal; the lower control system will overcome
disturbances without having to anticipate them. At the higher levels,
the model doesn't need the details because the controlled variables are
more abstract. But however the world-model is implemented, and at
whatever levels, behavior based on internal models is still limited even
in principle because of the impossibility of predicting all disturbances
that will make a difference. If I decide to buy IBM shares on a Monday
afternoon, I know what to do ordinarily, but there's no way I can model
the fact that the senior partner has died and my broker is at his
funeral.

What we need is data, not just blind modeling based on thought-
experiments. It's much too easy to imagine doing something that one
can't actually do in real behavior. And it's much too easy to remember
only positive instances of behaviors that fit a model and to bend the
memory to favor the model. That's why I wrote the little program about
blind tracking. This will let us see just how good a model-based
behavior can be, and will also help us to notice facts that might be
used for an alternative explanation of the behavior (such as the fact
that I, at least, control a different perceptual variable when I can't
see the cursor). I currently favor your explanation:

     The example of walking through a dark room seems like a description
     of what one would think of for "model based control" however, even
     for those of us that "briskly" walk through pitch dark places, I
     believe that is still an example of PCT. A "model" is indeed
     involved but the model is setting references for current
     perceptions.

Right, kinesthetic and vestibular perceptions. If there were a total
loss of perceptions of all kinds, I doubt very much that we would be
able to walk through a pitch-dark room -- or walk at all.

On the other hand, we can't just discount the idea of mental world-
models. There are too many things we do that involve making predictions
and running imaginary scenes through our heads in which remembered
properties of the world play a part. When one person calls across the
room to another person "Knight to King's Bishop six" there are mental
chessboards inside the players and onlookers, and a quite incredibly
accurate model of the state of the board at any time. Notice, however,
that these models don't have to include the details of HOW the knight
gets moved. The players know which knight is meant from the (imagined)
state of the board and the state of play. The player making the move
specifies the new reference position, but he doesn't have to specify the
means by which the new position is to be achieved. The player doesn't
have to add, " ... and don't knock the Queen over on the way." The lower
control systems will take care of that if the board is real; in an
imaginary board, collisions don't even have to be modeled.

Sitting around and making up stories isn't going to help us come up with
the right model. We must actually try the things we're talking about, in
a setting where we can record what actually happens.

···

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

Bill P.

<[Bill Leach 950509.18:34 U.S. Eastern Time Zone]

[From Bill Powers (950509.0815 MDT)]

I think that "gets" somewhere, at least for me. I believe as I think
that you do, that it is probably impossible that "Model Based" control
could exist as a complete "control loop" for even the simplest behaviour
(at least if involving the external world).

Thus, the "Model" starting up and actually generating final output device
operations is not a consideration for us.

The partial "Model" operation is, I think, much easier for us than for
any other group. In HPCT it is not at all unreasonable to consider that
we probably do actually use such models as a matter of course. They use
is not likely to be noticed until the model fails to produce the correct
results FOR SOME HIGHER PERCEPTION.

I really believe that stumbling upon a missplaced stair is specifically
an example of a "failure" with such a system in operation. Obviously
(again I think), the "model" system is only a "minor" part of the whole
control loop handling the goal of perceiving oneself at the top of the
stairs. Do you agree that this is possibly true?

-bill

[Hans Blom 950510]

butting in on (Bill Powers (950509.0815 MDT)) to Bill Leach (950509.01:42)

RE: world-model-based control

However, the basic defect in this approach is a practical one: it isn't
really possible to model the environment in all the revelant aspects. A
true world-model would, for example, have to include all the inverse
kinematics and dynamics of the body and limbs ...

Where did you get this idea? Let me correct it immediately. A world-model is
the internal FORWARD kinematics and dynamics representation of its external
counterpart. The model runs IN PARALLEL WITH the world. Inverse models and/
or computations enter in some CONTROL schemes, but not in IDENTIFICATION
schemes.

                                               , plus all the physical
laws that describe effects of body and limbs on other objects in the
environment, plus the physics of the affected objects, plus world-models
of the perceptual system through which we monitor effects of behavior.
And even if that could be done, we would still have the problem of
unpredictable disturbances that are left out of the world-model. In
order to use world-model-based control to accomplish umbrella-opening-
to-keep-the-rain-off, we would need a successful weather-predicting
world-model.

Who says that a model is always accurate? My weather-predicting world-model
isn't very accurate, and neither is my umbrella-opening control system. Who
says that models accurately model ALL OF the world out there? They don't.
Our three pounds of brain are just not enough to model all that goes on out
there. Barely enough, I would say, to keep the human race propagating its
genes for a few million years!

In world-models like the one Hans submitted, most of these problems are
bypassed.

No, they are not. It is just that a simple model is simple: only accurate
enough to model a very simple "world", not the real thing out there. The
model in our brains is enormously more complex than the one in my demo,
hence its explaining/predictive capabilities enormously more extended.
That is not to say that even a model in the brain is a COMPLETE model. No
model ever is.

         If there is a signal u that affects the world-model, it also
affects the real world -- but nothing is said about the means by which
it affects the real world, for example the properties of the actuators
(which the real world-model in a real organism would also have to
contain).

The transfer function of actuators can be modelled as well. Or, preferably,
measure the output of the actuators as near to the skin -- the interface to
the outside world -- as possible.

        Likewise, on the input side, there is a perceptual function
which is represented only as a noise generator adding to the effect of
the controlled variable on an internal representation. In a true world-
model, the model of the external world would have to include the
properties of the sensors -- for example, the lens of the eye, the
retinal computations, and any higher computations required to convert
the optical inputs into measures of variables like position or velocity.

The transfer functions of sensors can be modelled as well as those of
actuators. Why not?

Of course nobody can include such details in a model (anybody's) that we
can actually program and run, but we have to consider these details in
judging the practicality of any proposed model of behavior. The concept
of world-model-based control necessarily implies that all these details
must be present in the organism's world-model; I am dubious about that.
The real system is not allowed to make the simplifications and
approximations that we make in our computer models.

Why not? It does. You do not drive EXACTLY in the center of your lane. In
everything you do, you keep a margin of safety. It is often surprising for
modellers to discover how simple (low order) models can be used (in control
tasks, say) to represent complex (high order) systems. The reason is that
that margin of safety is usually there. Small errors usually do not matter,
as you know from cursor tracking experiments. Being off by a few pixels
doesn't kill you. If it did, you wouldn't do cursor tracking experiments.

... But however the world-model is implemented, and at
whatever levels, behavior based on internal models is still limited even
in principle because of the impossibility of predicting all disturbances
that will make a difference. If I decide to buy IBM shares on a Monday
afternoon, I know what to do ordinarily, but there's no way I can model
the fact that the senior partner has died and my broker is at his
funeral.

You can't. You don't. You cannot predict to full accuracy. You make mis-
takes. Models cannot be perfect, except those of very small artificial
worlds.

Sitting around and making up stories isn't going to help us come up with
the right model. We must actually try the things we're talking about, in
a setting where we can record what actually happens.

The "right" model? No model of human behavior can ever be right. A model is,
by definition, a SIMPLIFICATION of something else. You can only hope that it
captures all significant aspects and leaves out only insignificant details.
But look at your mental model versus mine: what is important/significant for
you is different from what is important/significant for me. In many ways,
not just in our discussions about modelling and control. How can we ever
agree about what the "right" model is?

Greetings,

Hans

<[Bill Leach 950511.01:17 U.S. Eastern Time Zone]

[Hans Blom 950510]

To my thinking you quite literally can not be butting in on a dialog in
which you have been a primary participant for over a week.

Putting my thinking "on the line"... A "pure" model based system can
_ONLY_ work on paper, thought experiment, or computer simulation. As a
practical matter it can also "work" in a purely electronic system (ie:
No "mechanical" input or output).

In absolutely _NO_ real world application can a purely model based system
achieve control (that is work inspite of the presence of disturbance) --
at the _very_ least ALL mechanical output must be either negative
feedback control or always operated against fixed physical limits.

Such system can and indeed have proven to be very useful and productive
but they are "control systems" only in the very loosest use of the term
and are not related to living system control.

While I do believe that at least part of what you have demonstrated can
be inferred to exist in living system control it must be at the most only
a part of any complete control loop. I might be "all wet" but I can not
imagine that it would be possible for a "model based controller" to be
actually providing muscle "drive" signals for example.

Where did you get this idea? Let me correct it immediately. A
world-model is the internal FORWARD kinematics and dynamics
representation of its external counterpart. The model runs IN PARALLEL
WITH the world. Inverse models and/ or computations enter in some
CONTROL schemes, but not in IDENTIFICATION schemes.

I may be misunderstanding you here but I do have a couple of problems
trying to figure out just what you mean...

In the first place, as I envision possible "model control", the model
would not do any "calculations". The "model" is probably created "as a
matter of course" just from the fact that the organism IS normally
controlling perception. Though as Bill P. has already pointed out,
"switching modes" might well be anything but trivial, it _seems_ as
though the "model" can provide reference signal outputs at the "proper"
amplitudes and rates as long as some perceptual inputs are received and
correctly "match up with" experience. The systems receiving the model's
references WILL control against any and all disturbances affecting their
perceptions. Disturbances that affect the model's goal and with
sufficient magnitude in a dimension not perceived by any of the feedback
control loop will _Always_ result in failure OF THE MODEL to achieve the
desired perception.

I suppose that Rick (being a simple sort of guy) might phrase this as "a
control system that does not control against disturbance is not a control
system" or maybe the even more obvious "a system that purports to control
a perception but does not sense the perception is not a control system".

What do you mean by "FORWARD kinematics and dynamics representation ..."?
You seem to be saying that the model must do output calculations and then
that it does not have to do so.

accurate

I think that you might be missing Bill P.'s point here. The issue is
"accurate with respect to what?". Achieving a "perfect" output value
that does not achieve the desired perception maybe considered to be very
accurate in some sense but that sense is essentially useless.

COMPLETE model

You are undoubted right that there are _no_ complete models for the "real
thing" and thus only closed loop negative feedback systems with a random
or limited random trial methods can possibly "cope".

The transfer function of actuators can be modelled as well. Or,
preferably, measure the output of the actuators as near to the skin --
the interface to the outside world -- as possible.

The transfer function can be modelled but since the actuators are not
perfect and vary in their characteristics based upon both internal and
external influences the modeling will not work. Of course your
suggestion of sensors suggests that the transfer functions could be
ignored IF the model only sets references for lower level perceptual
control loops (as I and at least Bill P. have suggested is probable).

Bill P.>

Of course nobody can include such details in a model (anybody's) that
we can actually program and run, but we have to consider these details
in judging the practicality of any proposed model of behavior. The
concept of world-model-based control necessarily implies that all these
details must be present in the organism's world-model; I am dubious
about that. The real system is not allowed to make the simplifications
and approximations that we make in our computer models.

Why not? It does. You do not drive EXACTLY in the center of your lane.
In everything you do, you keep a margin of safety. It is often
surprising for modellers to discover how simple (low order) models can
be used (in control tasks, say) to represent complex (high order)
systems. The reason is that that margin of safety is usually there.
Small errors usually do not matter, as you know from cursor tracking
experiments. Being off by a few pixels doesn't kill you. If it did, you
wouldn't do cursor tracking experiments.

I don't believe your argument supports your contention. Again, accuracy
is not an issue beyond achieving the desired perception but is irrelevent
to the issue. It is _FLATLY_ impossible to CONTROL for something that is
not sensed but is subject to disturbance. The idea that we might really
use a model based system to _attempt_ to achieve (or maintain) a
perception in absence of the perceptual input, even for short periods of
time does not in any way suggest that living being are not fundamentally
and basically negative feedback control systems. Model based "control"
isn't... and if Hiesenburg (sp?), Bohr and the likes of QED are even
close in principle then IT CAN NOT control.

Of course at least in a sense you acknowldge this with:

You can't. You don't. You cannot predict to full accuracy. You make mis-
takes. Models cannot be perfect, except those of very small artificial
worlds.

The problem however is that model based systems have no way of knowing
when they fail since they do not perceive the controlled variable --
sorta like the psychotic living in non-existant castles or the acid head
"flying like a bird" from atop a twenty story building.

The "right" model? No model of human behavior can ever be right. A model
is, by definition, a SIMPLIFICATION of something else.

Maybe by your definition but I know of no _requirement_ that a model be a
simplification even if that truely is usually the case.

You can only hope that it captures all significant aspects and leaves
out only insignificant details.

No you don't just hope... You do what Rick has been literally begging
people to do and that is beat upon the model to get it right. Yes, it is
true that except in rather trivial situations, no model is KNOWN to be
absolutely correct -- in good models all we know is that the model
performs the same as "reality" in every test that we have so far been
able to apply.

But look at your mental model versus mine: what is important/significant
for you is different from what is important/significant for me. In many
ways, not just in our discussions about modelling and control. How can
we ever agree about what the "right" model is?

Bill's MODEL is not a mental model. Bill's "mental models" OTOH are just
that until the become "real" by actually doing what they are supposed to
do _independently_ of his thoughts about how he thinks that they work.

As to the "right" model...

It must behave as does the organism being modeled behaves under the same
circumstances.

To be a really REALLY "right" model, it should have the same internal
organization as the modeled organism.

As it stands today, if a non negative feedback control system model can
behave the same as living beings behave then it is just as "right" as PCT
until more is known about the "real" internal structure. That however
appears to be one hell of an "if"!

-bill