[This mail was originally sent on Thursday 10 Dec but returned because
of the breakdown of the internet connection]
[Martin Taylor 921210 17:10]
(Bill Powers 921209.0900)
Bill--Little Man 2 arrived today. It runs fine, so if anyone else has
a bad copy, it at least isn't a general flaw. We have been talking
about whether we can get it to run on a Mac, and what would be involved.
Is anyone else doing that (Is that what Rick wanted to do? I've forgotten).
···
------------------------
Bill to Bob Clark (921205)
I'm glad to see you opening up the discussion. "The" hierarchy is a
figment of my imagination, building on OUR imagination. For most of the
levels I've proposed, the only backing for the specific definitions is
anecdotal and subjective. As far as I'm concerned, these or any other
levels won't be "facts" until we have put them to experimental test.
...
I agree with Rick Marken about decisions: they represent conflicts.
Unless there were at least two competing goals to satisfy, there would
be no need to make a decision. You would just do whatever is required to
achieve the single goal.
I have been spouting off around here at any opportunity about the power
of PCT, and so a couple of colleagues have asked me for a PCT approach to,
and predictions about planning/decision making. The situation is one in
which imagination must be used. The real world cannot be addressed to look
for error. Here's the task they are using:
The planner is given a sheet of squared paper on which is a map of the aisle
of a supermarket, with goodies such as "yoghurt", "cookies", "Coke," "Oranges"
marked on the shelves, one item to a square. There is an "In" door and
an "Out" door, and two robots with shopping baskets, a blue one with a
basket that can hold 6 items and a red one with a basket that can hold
11 items. The blue robot can move (must move) 2 squares in one clock tick,
while the red robot moves one square. Between the two robots, they have
to pick up 15 items on a shopping list in a time less than T ticks, where
T is specified by the experimenter as 20% longer than the best time the
experimenter or a few selected people have managed to achieve. The robots
can transfer goods between their baskets if they stand on the same square.
The baskets are actually stacks, one item wide, so that every item is
explicitly on top of or under every other item in the basket. Drinks
cannot go on top of anything, and bakery goods cannot go under anything.
Dairy products must be picked up "as late as possible before finishing."
The expedition is finished when both robots stand at the OUT door with the
15 items in their baskets.
The subject in this experiment is required to specify every move of each robot,
without retracting a move, so the plan must be pretty well completed in detail
in imagination, without violating any of the constraints. The experimenters
ask the subject to vocalize as much as possible the thoughts that go on in
making the plan, and to gesture and mark up a sheet showing the map of the
supermarket. They think they have identified four levels of abstraction
in the planning process, but these levels do not seem to correspond with
any PCT levels (as one might expect, given Rick's PCT motto # 1: "you
can't tell what anyone is doing by looking at what they are doing," as
modified in LP for communication: "you can't tell what anyone is saying by
listening to what they are saying.")
When I thought about this within the PCT structure, I came up to a point
where it seems to me that there is a need for modelling that is not
accomplished by the canonical "imagination loop." As I understand it,
the imagination loop asserts to the perceptual input function that a
reference level (the output signal of the ECS as seen by a lower ECS) has
been satisfied. But in this planning situation, one has to imagine that
the world has been unkind, and that the desired perception has been blocked.
(The robots can't transfer goods because the red robot hasn't got to the
rendezvous, for example. Or perhaps the transfer won't work because it
would put fruit on top of bread.) One can't see the problem in the real
world, because it hasn't happened. The same sort of thing happens when
planning in the face of danger. Making a wrong move may kill you, so you
can't control the error in real life. You have to do it in imagination.
It seems to me that there must be somewhere a facility that isn't memory
that allows one to put a simulation of real world impedances into the
imagination loop. I don't see where it comes from within standard PCT.
Bill says:
At the level I call "programs," symbol-handling processes occur which I
characterize as a network of choice points. These are like the TOTE
unit, in that there are tests for conditions, with the choice of a
branch being determined by a rule applied to the results of the tests.
The term "choice" seems to imply a decision, but in fact there are no
decisions at this level either. The conditions encountered at each
choice-point, plus the rules, completely determine the path to be
followed next. Only when there is ambiguity or when the rule is self-
contradictory (calling for more than one mutually-exclusive path to be
followed) is anything like a decision required. If you have an algorithm
for making decisions, you don't have to decide anything!
To me, this works only after the "plan" has been produced. The plan is
the program that determines the choice given the perception. It is in the
making of the plan that the imagined perception needs to provide the problem,
assess the conflict possibility, detect the barrier that the world will
put in the way of achieving the desired perception. The decisions are
in the plan making, I think. But where is the world model that allows the
imagination to know that decisions are required?
========================
Related, but different.
I detect now and in some discussions of last year two quite distinct uses of
the term "reorganization." One works on-line and affects what to do right
now, the other affects how to do things for ever afterwards. The latter
is what I have always addressed when I have talked about reorganization--
changes in the link structure among ECSs, changes in the perceptual input
functions, changes in reference input functions, and so forth. It's a
long-term process, on-line in a way, in that it occurs when what one is
doing is not working. But its effects are permanent until once again
reorganized away.
The other form of reorganization seems to be what has been talked about from
time to time, as in Bill's posting:
If we eliminate programs -- the execution of algorithms for choosing
paths -- from decision-making, what is left? As far as I can see, only
the cases in which for some reason we wish to do two contradictory
things at once. At that point we must reorganize or simply suffer the
paralysis of conflict.
A while ago, we had a little discussion about choosing to take a car or a bike
or to walk to a destination. Bill argued for a decision-making process
within the ECS that was trying to satisfy the reference to perceive its
organism as being at the destination. I argued for the impedance of the
world, as demonstrated in the world's inhibition of taking any two modes
of transport simultaneously. Now I'm asking about the imaginary process
of deciding NOW which to take (making the plan), when I will travel
tomorrow. Where is the information about the world's constraints held,
and how does it come into play? I don't see where reorganization comes
into the act, but quite a bit of Bill's posting to Bob Clark (921205) seems
to suggest that he does.
In puzzlement.
Martin
(PS. The internet connection still seems flakey, so I don't know when this
will get out. Maybe right away, maybe not.)