Misunderstandings & miscellany

[From Bill Powers (920826.2000)]

Spent a fine morning at Hovenweep ("Deserted Place" in Ute) National
Monument just across the border into Utah. Variation on Chaco Canyon
construction, but still the same idea: pile up big and little rocks
with adobe mud between them. It struck me that having a Plan for
stacking up rocks would have been futile when the Anasazi builders put
up one of those round, square, or D-shaped towers that abound at
Hovenweep. How can you make a Plan of Action when you don't know
exactly how one stone is going to fit on another, or how many thin
stones you'll need to make the next course level? The only Plan you
can have is for how the thing is supposed to look when it's done, and
they clearly had that. When I look at those T-shaped doors, the little
windows between rooms, the slanted peepholes, the straight corners on
some buildings and nicely rounded walls on others, I get the sense of
seeing into minds nine centuries dead. They have left their reference
images behind for us to see, in stone.

···

----------------------------------------------------------------------
Avery Andrews (920825) --

Avery, you are becoming the voice of reason on this net. Keep it up.
You will get us back on track.

I have received your C program. It is easy to read and understand,
which to me is good programming. I'll try it tomorrow -- no reason why
it shouldn't work. Only a few very small changes ought to make it work
even better. Back to you later with some ideas about how to tie Astro
to language.
----------------------------------------------------------------------
Penni Sibun (previous posts) --

From all the comments by others, I suspect that I misunderstood your

last post in most of the possible ways. My apologies. As Avery said,
we are ALL arguing against somebody else's point of view without
really understanding it. You keep your temper remarkably well, which
should be a lesson to all of us.

Just to clear the air, I interpreted your rhetorical questions as if
they were addressed to PCTers, with the "why" part a (misplaced)
criticism of PCT and the "when" part your suggested remedy. Now I
realize that this was a quote and not your words at all, and that the
target was the old-fashioned AI weenies. So we really have a three-way
collision going on here (4 counting interactionists, if I understand
correctly). It will help if we can keep straight what aspect of whose
ideas is being discussed, and why. I'm an old, old, man, Penni, so you
have to forgive me (heh, heh).

With Avery's explanations, I can certainly go along with the idea of
looking for the minimum necessary machinery. My philosophy of modeling
exactly.
----------------------------------------------------------------------
--
Hank Folson (920826) --

Very nice elaboration on the riding-the-bicycle thread.
----------------------------------------------------------------------
-
Greg Williams (920826) --

There's one problem with the "self-consistent" behaviorist scheme that
you didn't mention: the reference signal. When a behaviorist trains a
rat to do something, what the rat is to do is already known to the
experimenter. As I've said before, a behaviorist who wants to
demonstrate the reality of stimulus control of behavior must state
BEFORE the demonstration what behavior is going to be brought about.
In other words, the experimenter must describe what will be perceived
when the process has finished.

Skinner himself discussed shaping in those terms. He said that the
experimenter should simply keep in mind the final pattern that is to
be performed by the animal, and reinforce any little move that looks
like a step in the right direction. He emphasized that there's no
standard procedure here; he said in so many words that the
experimenter must be prepared to improvise. So he was describing the
behavior of a control system, but trying to explain it without using
any CT concepts -- avoiding particularly any mention of intention or
purpose. Without using those concepts, of course, he would have been
unable to explain the actions of the experimenter.

When Skinner spoke about an experimenter controlling an animal's
behavior, his language always contained some way of saying that the
experimenter desired, specified, or intended that a certain behavior
come into being. In talking about the environment controlling
behavior, of course, he did not use such terms. He didn't say that the
environment had in mind a particular behavior of the organism, and
improvised rules and happenings that reinforced any little move toward
that behavior. In speaking of "organisms" (meaning non-experimenters),
he also had to avoid anything sounding like a purpose, such as
knowledge in advance of the outcome or a desire for a particular
outcome. Only the experimenter was allowed to behave in a purposive
way, knowing what was to happen and doing whatever was required to
make it happen. Therefore he attributed true control only to
behavioristic experimenters, and not to subjects or to environments.

When he use the term "control" with respect to non-scientific
organisms or environments, he did not mean goal-directed error-
correcting actions like those the experimenter used. He meant
"influence" or "determine" or "cause" -- purposeless cause and
purposeless effect. In that context, "control" was a code-word, a
euphemism. He was doing as he often did: he used a common word that to
most people implies intention or purpose, but with the understanding
that only his meaning was allowed. Something like the way certain
politicians are now talking about "family values."

So I dispute use of the term "self-consistent" in reference to
behavioristic notions of control.
----------------------------------------------------------------------
-
Martin Taylor (920826.1100) --

RE: dynamics. I was interpreting Penni's reference to putting a study
of dynamics first as meaning that we should observe only the dynamics
of interactions as an external observer would see them. I think Rick
may have put the same interpretion on this passage. In fact, Penni is
right. Before we propose any model, we should first observe the
behaving system in its environment, and note the dynamics of the
interaction (by which I don't mean "psychodynamics," which involves
heavy layers of theory, but simply the temporal aspects of the
physical organism-environment relationship). This, I still maintain,
amounts to laying out the phenomenon that needs an explanation.

Then, of course, a model is required to explain the observed dynamics.
This model will have to have certain internal dynamical properties of
its own in order to show the same external dynamics as the real system
interacting with a real environment. So if I have misunderstood what
Penni was saying, at least my misunderstanding should be clear now.
----------------------------------------------------------------------
RE: degrees of freedom.

I demonstrated the absolute need for some means to pass control from
one ECS to another (or, equivalently, a means to change the

perceptual >input function of controlling ECSs).

I think we will all continue to be frustrated on this subject until we
come up with experiments that demonstrate this "absolute need." You
are talking about a HYPOTHETICALLY absolute need, conditional on
certain premises being true. What we need is a demonstration with a
live human subject in which a phenomenon occurs that can only be
explained by this organization in a model. It should be something
simple enough to do on a PC, so that the modelers on the net can do
the experiments, program the model, and test the match. We can
stipulate (as Avery pointed out) the existence of perceptual devices
as required to provide perceptual signals representing whatever
aspects of the environment are needed. The actions required should be
simple motor behaviors, to accomodate mouses and joysticks, but these
behaviors can be translated into any special effects on the screen
desired, simulating other kinds of behaviors (such as cranking,
pressing buttons, uttering words, and so on).

My proposal that a given ECS can be applied to different lower-level
perceptions at different times is somewhat similar to yours. Mine is
no better founded in experiment than yours. I believe we have
approached the limits of what can be accomplished by passing words
back and forth. We need to devise critical experiments and carry them
out. Otherwise we might as well be arguing about angels and pinheads.
---------------------------------------------------------------------
Ed Ford (920826) --

Congrats on your new computer. And most particularly, congrats doubled
for having not only one backup disc, but two.

I have now read "Outcomes Drive the System," and I can see why you
think that ODDM (Outcomes Driven Development Model) is ripe for a
merger with control theory.

To all and sundry. ODDM is a method for organizing educational systems
according to some basic assumptions. The author of this booklet
(unnamed but possibly John Champlin) lists them this way:

1. What I want determines what I do.

2. What I do must reconcile and align with the most current, valid
knowledge.

3. All actions and all decisions are aligned and intentional.

.... [skipping a few rather vague generalities]

6. There is no condition [or] feature of behavior which cannot be
intenionally altered in such a way as to insure success.

7. Change is a holistic activity. Single parts cannot be altered
without significant impact in all of the organization.

8. Any change must be in response to an established and accepted need.

...[and another generality].

Later:

   A need is created by
    (a) a discrepancy between what we want and the condition presently
        existing.
    ....
    (c) a difference between what we claim to be doing versus what one
        actually observes.

Rick, I'm a Republican....more on your convention comments later.

No, Ed, you are not a Republican. You are a control theorist. Being a
control theorist, you can find aspects of the Republican point of view
that are consistent with human nature, and others that are not. You
can do the same with the Democrat point of view, or Sun Yung Moon's. A
theory of human nature is contained in every political point of view
-- but is in important ways unfounded, because political points of
view are not based on any underlying principles that can be tested. I
put it to you that PCT and HPCT transcend political points of view, as
well as many others. Except for Rick. Yukayukayuk.

Belief: Each of us has criteria for what constitutes enough evidence
to justify belief (i.e., acting and thinking as if something is true).
If the true foundation of belief is evidence, then new evidence or a
reconsideration of old evidence can easily change a belief, because in
that case a belief is the result of evidence, not the reason for
accepting evidence. I define faith the other way around: faith exists
when the belief comes first, and the evidence is selected in any way
required to prove the belief true. So faith is preserved in spite of
the facts, while belief is generated because of the facts.

Of course there are several layers to this concept: what do you
believe to be the facts? And why?
---------------------------------------------------------------------
Best to all,

Bill P.

[Martin Taylor 920827 0020]
(Bill Powers 920826.2000)

Talking about my claim to have demonstrated a novel and important fact about
PCT through consideration of dynamics and the associated degrees of freedom...

What we need is a demonstration with a
live human subject in which a phenomenon occurs that can only be
explained by this organization in a model. It should be something
simple enough to do on a PC, so that the modelers on the net can do
the experiments, program the model, and test the match. We can
stipulate (as Avery pointed out) the existence of perceptual devices
as required to provide perceptual signals representing whatever
aspects of the environment are needed.

I doubt you could ever find a phenomenon that can be described by only one
organization. If you want an anecdotal set of facts that fit the proposed
organization, then you need only look at someone's head turn to face the
source of an unexpected noise, or the flick of the eyes toward a distant
light flash.

The argument goes the other way: if there are more sensory df than motor df
(either instantaneously or over time) and if a system is a living one (i.e.
has been built by Darwinian evolution), and if PCT is valid for living systems,
THEN there must be passive alerting systems that can direct control among
the possible percepts permitted by the sensory systems.

I make no claim for any specific organization of how these alerting systems
might work. I only claim that different percepts must come under control at
different times, and that there must be some mechanism for shifting control
to those percepts it is most important to control at any particular moment.
The instigator of that change is what I call the alerting system, which might
well not be a separate system from the control hierarchy, but which must work
passively and effectively in parallel across all the percepts that might be
controlled (or at least across those not currently being controlled). Your
proposed organization might fulfil these criteria. I don't know.

There really isn't any experimental demonstration that I can think of, because
you can't unbuild an evolved living system, and if you deliberately build a
simulation to have alerting systems that can affect what is controlled, then
you are testing a tautology. I guess you might set up (or attempt to set up)
a negative demonstration--a system that had far more sensory inputs than
motor outputs, that could control appropriately many percepts at several
levels, and that had no passive parallel alerting system. Success at such
a demonstration would show that such an alerting system is not necessary,
but I think it would be necessary to provide an equivalent (e.g. an autonomous
scanning system checking on the errors of currently non-controlling ECSs).

This basic condition for any such simulation is that the sensory system of
the simulated controller be much broader-band than the motor system, and that
the environment be rich enough to justify the sensory bandwidth.
What this means is that the experimenter-simulator must set up lots of CEVs
that the simulated organism could learn or be constructed to control (depending
on whether we wanted to bring an experiment on reorganization into this).
It may be that it would be easiest to do this in the natural world. But
perhaps one could work with a very restricted world, using a 2 df motor
system and, say a 6df sensory system of bandwidth higher than the motor
system by an order of magnitude. That would give 30 times as many sensory
df per second as motor df per second. Even this, I think, would be hard
to mate with a simulated world, especially when the objective is to show that
the organism can control a considerable variety of percepts, for each of
which the environment must provide the data, at several justifiable levels.

Maybe more experienced modellers than me could see a way to try such a
negative demonstration. Maybe someone else can think of a way to do a positive
demo or experiment. I think of it as a fact on the same level as that you
can't keep a cursor on a moving target unless you can see at least the
deviation of the cursor from the target.

Martin

(ps 920901.1200)

   [From Bill Powers (920826.2000)]

···

----------------------------------------------------------------------
   Penni Sibun (previous posts) --

   >From all the comments by others, I suspect that I misunderstood your
   last post in most of the possible ways. My apologies. As Avery said,
   we are ALL arguing against somebody else's point of view without
   really understanding it. You keep your temper remarkably well, which
   should be a lesson to all of us.

well, thanks. i'm not so sure about that though....

   [From Bill Powers (920827.0800)]

   ----------------------------------------------------------------------
   --
   Penni Sibun (920826) --

   I'm being cautious here to avoid more podiphagia.

   You say
   >the interactionist argument is in general deemphasizes individuals
   and >emphasises their interactions.

   Does this mean that interactionists believe that interactions can be
   EXPLAINED without reference to the properties of individuals, or that
   they can be DESCRIBED without such reference?

i don't think i mean either. we fall into our language problems here
though. maybe we can say that interactionists think that you can't
describe or explain the individual w/o also the interaction.

   Does the interactionist
   viewpoint differ from, say, that of conventional sociology, in which
   it is held that there are laws of social systems that exist apart
   from, and indeed unconnected to, laws of individual behavior?

yes, indeed. a primary interactionist field, ethnomethodology, came
about explicitly in reaction to sociology. rather than asking ``how
can we describe there here institutions,'' it asks, ``are there
institutions, and if so what, and if so how constituted.''

   ------------------------------------------------------------------
   I have a copy of a '92 paper for AAAI (whatever that is)

american association for artificial intelligence.

   by Agre and
   Horswill called "Cultural support for improvisation." In it he

``they''!!

the theory is certainly primarily agre's, but horswill is the one who
wrote toast.

   Before this section, the authors discuss the philosophy of their
   approach, which depends on using the properties of objects in the
   environment to replace a great deal of computation. Here's my
   interpretation of what they say; make allowances for the fact that I'm
   trying to swim through a sea of unfamiliar terminology. Gratuitous
   comments by me are in square brackets.

   Objects in the world are characterized by the roles they play in
   customary routine activities. [The playing of roles is attributed to
   the objects. A lot of the language that follows could be cleaned up by
   inserting "are perceived to be" or some such proviso at frequent
   intervals].

well, ``playing a role'' is a well-understood way of talking about
these things. the advantage of talking about roles instead of
perceptions, is that you don't have to ask perceived by whom.

   A given object type is characterized by a state graph: nodes are
   states and arcs are labeled as operations that change the state from
   one node to another. [These are apparently thought of as properties of
   the objects, but operations are defined in terms of their outcomes,
   not in terms of the actions that bring them about. In HPCT terms, the
   operations could be re-phrased as the setting of goals specifying a
   new state which lower systems then bring about, thus moving the object
   state to the new node].

   An operation can be a label that applies to several arcs leaving a
   given node, thus being "nondeterministic." [This baffles me. Do they
   mean "stochastic?" Do they mean that under various circumstances, a
   given act can have different outcomes?

they defined it: ``in the sense that it labels several arcs leaving
from the same state.'' the terminology is straight from computer
science theory: a machine is deterministic if there is only one way
to get from one state to the next; it's nondeterministic otherwise.

the point is that there's more than one way to get from one state to
another.

   A world state is a set of instances and a mapping from each of these
   instances to a state in its type's graph. A history follows a set of
   instances through a series of changes. [This would seem to amount to a
   model of the environment's behavior]

no, a history is what happens between one point in time and another
point in time.

   An action type is a sequence of pairs made of an object type and an
   operation from that type's state graph.[ So an action type is a set
   sequence of operations that produces a set sequence of states.

it's not a set sequence of states if the operations are
nondeterministic. it's a set of possible sequences.
   A goal for an object is satisfied if the world state contains some
   instance of the object type that is in the indicated state.(So one
   cannot assert that two instances of a given type must be in a given
   state). [I don't see the advantage of this rule. It says, apparently,
   that the goal of breaking a given egg is satisfied if the egg being
   operated on remains intact, but another egg breaks.

no, it says that the goal of breaking an egg is satisfied if an egg
(any old egg) is broken. it doesn't have to be egg-47 or egg-13.

   So the only egg that can
   break is the one being operated on by the break operation and there's
   no need to keep track of which egg is broken.]

that's true. why do you find this contradicts the above?

   Summing up, the authors give this description of the algorithm:

   While there are unsatisfied goals,

   Choose aribtrarily an unsatisfied goal.

   Let G be the specified goal state.

   Find an instance M of the specified material type.

   Determine its current state S.

   Let A be the entry for the S and G in M's action table.
   Choose tool instances to fill each of A's participant slots,
   preferring instances which are already in their normal state.

   If some chosen tool T is not in its normal state, then perform the
   action specified in T's action table for its current state. Otherwise
   perform A using M and the chosen tools.
   ----------------------------------------------------------------------
   -
   This last summary is ever so much simpler than all the verbiage that
   led up to it, because the attempt to present a general case has been
   reduced to a specific case.

in what way is it specific?

   One change I would suggest is the explicit addition of an error graph
   for each operation. The main weakeness of this algorithm as it stands
   is that every operation must produce a predictable state. By using an
   error graph this limitation could be removed.

yes, i agree. it would at least be interesting.

it would certainly be interesting if you (or someone wanted to try to
incorporate your suggestion into toast. it would be instructive for
everyone, i'm sure. i think you can get the code from horswill,
though i would guess it's all in lisp.

cheers.

        --penni