World models

CSGnet Archive CSG1994
1

<Martin Taylor 940221 15:30>

Bill Powers (940221.0930 MST)

Martin Taylor (940221.0945)

(1) Exaggeration in the use of the words "world model." The
"world" in question is the world relevant to the ECS in
question, namely the effect of output on the perception of the
ECS. The world is that of the ECS.

Even a limited world model requires much more than that. A
control system that controls the position of a mass in space
perceives only position (and perhaps velocity).

The "world model" that could be useful to an ECS doesn't need any of that.
All it needs is some kind of time waveform that corresponds to the way
the CEV would react to the ECS output in the absence of disturbance. Any
more, and you are not applying the words "world model" as I intend them.

And I still think that this form of usage is quite parallel to the usage
that allows you to talk about "perception" when dealing with the time
waveform of a scalar perceptual signal.

To build a tuning system that would alter the model to make its
behavior like that of the real perceptual signal, someone or
something would have to decide to represent the properties of the
environment as two integrators in tandem, with or without springs
and friction.

Why? What is needed is a Finite Impulse Response (FIR) filter that has
the right temporal characteristics. If these characteristics could be
constructed more simply by modelling the processes in the outer world,
so be it. Nothing available to the ECS would allow it to know about that
possibility. And there are methods of optimizing the match of an FIR
filter to the anticipated waveforms; they are described under the rubric
of "self-tuning systems" in control theory courses. I don't know, but I
anticipate that in the simpler systems with which we deal, sub-optimum
but trivial algorithms would work quite well for most cases. Where they
might not is in situations where the nonlinearity of the world is intrinsic
(such as a bifurcations--being "over" a tabletop when putting something
down is quite different from being "under" even when the positions are
quite close).

No, that's comparing apples and apple strudels. A world model of
cups and cupboard requires building a model with properties like
that of the real world

An ECS knows nothing of cups and cupboards. Its perceptual signals are
waveforms of values. If an outside observer identifies changes in those
values with changes in numbers and locations of cups and cupboards, so be
it. The ECS knows none of that. And if the "world models" in the various
ECSs happen to correspond (for the outside observer) to changes in cups
and cupboards, so it may be. The ECSs in question don't know.

To propose that perceptual signals are the elements of subjective
perception is simply to propose an identification: the subjective
perception is perception of a neural signal. The glorious complex
that we subjectively experience is made of individual perceptual
signals, as we can see by disturbing the complex in selective
ways. If we could not disturb it along independent dimensions, we
wouldn't think of it as multidimensional.

I think you take my point. All I am saying is that the same applies both
to the real (from the senses) signals and the imagined (from the world
models) signals.

I have yet to see a self-tuning system that can decide it is
perceiving the wrong variable, or producing the wrong kind of
output.

I thought that's what reorganization was about.

No, the reorganizing system doesn't have any idea what's wrong
with the system it's reorganizing.

No, and neither does the "self-tuning" thing in a self-tuning system.
Of course it is external in a sense. The self-tuning occurs as a result
of external signals. If the external signals match the anticipated signals
that are the output of the world model, no tuning takes place.

My meaning in referring to the reorganization process was that through
reorganization, new forms of perceptual function and new linkages among
existing ECSs come to pass. If the reason for failure of the hierarchy
to maintain intrinsic variables was that some ineffective (i.e. "wrong")
variables were being perceived, reorganization will, with luck, change
them.

One could use exactly the same approach to developing world models. An
e-coli system should work reasonably well for short impulse responses.
But since gradient-based methods exist and are plausible, it seems more
effective that they should be used. Actually, I'd be quite interested
to try a random-walk approach to the self-tuning problem.

It depends on what you see as "the system." If "the system"
includes the reorganizing subsystem, then it is bootstrapped.

I hoped you wouldn't say that. I consider "a control system" to
be the least semi-permanent functioning unit that can control.
It's not hard to separate a self-tuning "system" into the system
that tunes and the system that gets tuned. Any broader use of the
term system just opens the doors for the lawyers.

There are elementary systems (minimal control units--ECSs), hierarchic
systems (the learning hierarchy, as you sometimes call it), and organic
systems (things in skin bags, which are the overt things that behave).
You can separate out subsets of all of them, but the total organism is
legitimately described as a system, I think. It includes a reorganizing
process.

I have not seen a proposal for an external entity that changes
the nature or the parameters of the reorganizing system.

I call it DNA, which varies the settings (and nature) of
intrinsic reference levels as a way of coping with evolutionary
selection pressures (on days when I am promoting that model of
evolution).

Sure, but now who is extending the meaning of terms. Now you have a
system that includes all organisms back to before the first single cell.
DNA doesn't change the settings, so much as record them in such a way
that ones that work continue to be observed. DNA is itself changed from
outside, whether by cosmic rays or by mustard. When you get to that level
of discussion, all life is one of Rick's detested "self-organizing systems."

(Rick: how about reading something on self-organizing systems before making
silly comments? Discussions are much more valuable when based on at least
minimal common understanding.)

Martin

2

[From Rick Marken (950830.0830)]

Bill Powers (950829.0930 MDT) to Hans Blom --

I notice that you make frequent reference to your "world-model"
model, but that's just a thought, too, isn't it?

Hans Blom (950830) --

Sure, a contruction, a model. My internal "world-model" tells me
that we can do no more than construct "world-models" as a way to
understand and deal with the world. That's kind of recursive, I
know. But it's a construct that I'm fairly happy with, because
for me it has a lot of explanatory power and thus far I haven't
discovered internal inconsistencies (conflicts) in this model.

I think it is this attitude that nearly killed the patient who was
attached to your model-based blood pressure control system. A scientific
theory can be completely internally consistent (no conflicts in the model)
and completely wrong (inconsistent with external reality). Bill Powers
showed that your model-based controller controls no better (and in most cases
far worse) than a simple perceptual control model of the same situation; so
your model is neither a parsimonous nor an accurate representation of the
real controlling done by living systems. And, thus far, we have seen no
observations of real controlling that CAN be explained by your version of a
model-based controller and CANNOT be explained by a simple perceptual
control system.

I am aware that, as any model, it is incomplete and susceptible to
change at any time upon new insights (perceptions).

I guess "change" doesn't include "rejection". I have reported experiments on
"predictive control" where the behavior is explained perfectly by a
perceptual control system; apparent "anticipation" (cursor "leading" target)
occurs when the perception controlled includes the derivative of target
position. A model-based controller did no better than the perceptual
controller and, indeed, completely failed when an unpredictable disturbance
was added to the effect of the output on the cursor. I reported these results
on the net but I have detected no change in your world model; you still seem
to believe that a controller must include a model of the nature of the
environment ("world") that relates the controller to the controller variable;
that is, you assume that a controller must have a model of the feedback
function in order to control. You believe this despite proof that it does
not.

I guess that I just want to say that it works, for me, as a high level
concept that is able to give a inclusive, though hazy, picture.

If it works for you, that's great. But I'm more interested in models that
really work -- in the sense that they are "externally" consistent, ie.
consistent with what we actually observe.

Best

Rick

3

[Martin Taylor 950901 11:40]

Rick Marken (950830.0830)

And, thus far, we have seen no
observations of real controlling that CAN be explained by your version of a
model-based controller and CANNOT be explained by a simple perceptual
control system.

At periodic intervals, I pose the following pair of questions for theorists
and control engineers. Thus far, nobody has suggested even an approach to
an answer. In the questions, "model" is taken to be an isolable module
with content that can readily be explicitly expressed.

Perhaps this time???

The questions:

1. Is it possible to show that there CAN BE any model-based control system
whose behaviour cannot be reproduced in exact detail by a non-model-based
perceptual control system?

2. Is it possible to show that there CAN BE any perceptual control system
whose behaviour cannot be reproduced in detail by some model-based control
system?

I have reported experiments on
"predictive control" where the behavior is explained perfectly by a
perceptual control system; apparent "anticipation" (cursor "leading" target)
occurs when the perception controlled includes the derivative of target
position. A model-based controller did no better than the perceptual
controller and, indeed, completely failed when an unpredictable disturbance
was added to the effect of the output on the cursor.

The fact that a particular model-based controller did not replicate the
behaviour of a particular perceptual controller is kind of irrelevant to
the two questions. And it is the second of those questions that has to
be answered "YES" for your comment to be a valid criticism:

you still seem
to believe that a controller must include a model of the nature of the
environment ("world") that relates the controller to the controller variable;
that is, you assume that a controller must have a model of the feedback
function in order to control. You believe this despite proof that it does
not.

(That last sentence should read: "despite proof that a particular model does
not, in a particular case").

So do you, but the controllers you deal with hide their "model" within
their structure of perceptual functions, output functions, and linkage
weights. Bill's Artificial Cerebellum improves control some, when the output
function learns a specific model of the feedback function. But even the
simplest "control" system will not on average control without at least enough
of a model of the feedback function to get the sign of the output right.

You are arguing not about the existence of a model, but about where in the
control system the model resides, about how it is used, and about whether the
model can be abstracted as an identifiable module of the control structure.

Martin