[From Bill Powers (920910.1100)]
Chris Love (920909.1530) --
I read Bill's post on imagination [Bill Powers (920906)] and felt
that >something was either left un- explained or missing. This dealt
specifically with the objective(s) of imagination. To me, an
objective >is to assist ECS' that are having difficulty arranging
their weights >to percieve the desired control variable (I think you
guys call it CEV >now).
The imagination connection isn't in the theory as a way of helping the
system do something. It's there because we imagine. The phenomenon to
be accounted for is basically that of perceiving something that isn't
coming in through the senses. This includes memory, hallucination,
dreaming, and planning (by manipulating imaginary images and
thoughts). Human beings do these things; the model has to provide a
place for them. This is quite aside from asking why imagination exists
-- what an organism with imagination can do that an organism without
it can't do.
Now I like Bill's proposal since it does away with having to *pair
up* >each and every percept (input) with the corresponding *output* of
an >ECS in the imagination loop....
Perhaps what you feel is missing is that pairing-up. In fact, it's
still there, although without the complications that the previous
version of this idea required.
Suppose a system of level n controls a perception that is a function
of 3 lower-level perceptions. Normally, control is accomplished by
adjusting the reference signals in all 3 lower-level systems. Those
systems in turn work through still-lower systems, and eventually
through effects on the environment that alter primary sensory
information. Then the result is passed up the perceptual hierarchy
until it gets back to the original system as three perceptual signals
from the immediately-lower systems that received the three reference
signals. When all is well, those three perceptual signals match the
three reference signals.
Let's say that some higher system decides that our level-n system is
to operate with imagined perceptions rather than real-time ones. To do
this, under the new arrangement, it does not throw the level-n system
into the imagination mode, but the three level n-1 systems.
Now the error signal in each level n-1 system is short-circuited back
to the comparator as a pseudo-perceptual signal. A copy of this
pseudo-perceptual signal heads upward just as it does with real
perceptual signals. This happens in each of the three level n-1
systems, so the level n system gets three perceptual input signals as
usual. It combines them according to the form of its input function to
generate the level n perceptual signal. So for the level n system (and
all higher systems), control proceeds normally.
The great advantage of this method is that we don't have to have
duplicate copies of all the weights that are applied in the level n
perceptual function (or whatever form of computation exists), in order
to get the right imagined signal. The level-n perceptual function
applies its processing in the same way as usual. The level n-1 systems
do no perceptual processing at all, so we don't worry about the
weights there, either. So we get the best of all possible worlds:
switching to imagination is an elementary process, while applying the
correct perceptual function is done by the same means that is used in
real-time perception.
I'll be interest in your implementation of an imagination loop that
can adjust perceptual weights. This would definitely be part of
planning or rehearsing a control process.
ยทยทยท
-----------------------------------------------------
So now the question arises as to how I will propose to test out these
theories? I now want to make a pseudo control system, where most
everything that can be preprocessed is, excluding a single ECS layer.
Here I can isolate the effects of the various learning/reorg.
algorithms that the group here at DCIEM (Martin Taylor's group) will
develop. I hope this will provide a nice testbed for any proposals
that come up on the net for percept learning and/or output learning.
Sounds wonderful. I will NOT ask stupid questions like "when?"
----------------------------------------------------------------------
Greg Williams (920910)--
Back to precision (Talmudicity?): Some kinds of manipulation,
including >examples of ALL FOUR types I enumerated in my last post,
can SOMETIMES >MAKE USE OF deception. They are NOT deception, just
MAKE USE OF IT >sometimes.
(from previous post)
There are four kinds of manipulation (I claim): those which, if
successful ... result in the manipulee's ... judging them ... harmful
["exploitation"], neutral ["?"], good ["facilitation"], or mixed/who
knows/can't say ["?"].
These aren't kinds of manipulation; they're judgements about the
outcomes of manipulations. Any single manipulative act of any kind
could be judged in all four of these ways by different people.
What I'm trying to sort out is the NATURE of manipulation, not its
situation-dependent effects. Judging the effects gets us into the
subject of "importance," which I still want to postpone for a while.
I'm still concerned with establishing what kinds of manipulation are
POSSIBLE.
We've established one kind of manipulation, which is controlling a
person's action by applying a disturbing influence directly to the
variable that the action is being used to control. For short, this is
"rubber-banding."
Another kind involves the manipulator having control of part of the
environment on which the manipulee's controlled perception depends, in
a case where that part of the perception is NOT under direct control
by the manipulee, but only contributes one variable to a multi-input-
variable perception. For short, this is "manipulation of action
through disturbance of uncontrolled perceptions" or just
"manipulation through uncontrolled perceptions" (OK?).
In the final analysis, both of these forms of manipulation are
manipulation through disturbance of controlled variables. In the
"rubber-banding" case, the disturbance is applied directly to the
controlled variable. In the "manipulation through uncontrolled
perceptions" case, it is applied to a higher-level variable through
disturbing an uncontrolled component of that variable. Both cases
result in the manipulator being able to exert predictable and
quantitative control over the actions of the manipulee, with the usual
provisos about not violating other goals of the manipulee. I assume,
of course, that the manipulator has correctly identified all pertinent
controlled variables.
Can we agree, then, that we have established one method of
manipulation, which we can call "manipulation of action through
disturbance of controlled variables"? The nature of this method
(although not its success) is independent of what the manipulee thinks
of the results.
There is another type of manipulation that you've brought up
previously. We could call it "manipulation by altering the properties
of the environment." This is what Skinner referred to as altering a
"schedule of reinforcement." To do this, the manipulator requires the
ability to alter the part of the environment that lies between the
manipulee's actions and the manipulee's perceived consequences of
those actions. As the manipulee is trying to keep the perceptual
consequences matching an inner reference state, the result will be a
change in the actions that quantitatively compensates for the change
of environmental properties. If the change in the environment is not
too large, the manipulee can keep control without any change of
internal organization. If the change is large, higher-order control
may be needed, or reorganization may commence. In any case, successful
manipulation requires that the manipulee keep or regain control by the
time the process has ended.
We can call this "manipulation by altering properties of the
environment."
Purposeful manipulation of actions is possible only when
reorganization does not result from the manipulation. If there is
reorganization, the outcome can't be predicted. There is no
possibility of purposeful manipulation unless there is a systematic
relationship between the manipulative action and the behavioral
outcome.
I can think of one exception: the case when the manipulator is using
the E. coli method of acting. If the manipulator chooses a
manipulative act randomly, but changes the action immediately if the
observed behavior of the manipulee departs further from the desired
action, eventually the manipulator will arrive at an effective
manipulation (or die of old age). This, however, would not seem to be
a markedly effective method, particularly if the manipulee is also
reorganizing.
If we stick to cases in which systematic manipulative action has a
systematic and reliable effect on the actions of the manipulee, we are
left with
(a) Manipulation by disturbance of controlled perceptions, and
(b) Manipulation by altering properties of the environment.
Another type of manipulation can be called "giving new information."
To give NEW information is not the same as disturbing an already-
existing and already-controlled perception by either method (a) or
method (b). New information is initially uncontrolled, and in fact may
initially be unperceivable by the manipulee. Let's assume, however,
that the manipulee can understand the new perception.
New information can alter the way behavior looks, in a predictable
way. You see someone trying to pull a nail with a hammer, but holding
the hammer-handle very close to the head. You say "The farther out on
the handle you hold it, the more force you can apply to the nail."
Assuming that the manipulee is not getting anywhere with pulling the
nail out, and continues to want it out, you can predict that after a
moment's thought, the manipulee will make the connection between the
relationship you have just described and the task at hand. You can
predict that the manipulee will change the grip toward the end of the
handle. Of course you can't predict whether this will pull the nail,
or whether the manipulee will make the connection and adopt your new
information, but that's another matter.
This form of manipulation becomes far more reliable if the manipulee
REQUESTS the new information. "Why can't I get this thing out?" Such a
request for information practically guarantees that new information
will at least be heard. Of course then giving new information is not
an act of manipulation, but a response to a disturbance: it is the
"manipulee" who is controlling for getting the information. This still
leaves the manipulator, however, in the position of being able to
choose the information in terms of the effect the manipulator wants.
The manipulator could say "Because you're not holding the handle close
enough to the head of the hammer," if the objective is to see the
manipulee continue to fail, or even just to see if the manipulee will
do any stupid thing you suggest.
Asking a question has already been covered under "manipulation by
disturbance of controlled perceptions." Most people believe that
hearing a question should be followed by hearing oneself or someone
else emit an answer. It's like answering a telephone -- and the use of
the term "answering" in that context is interesting. What do they say
in other languages?
My last entry is "manipulation to see what the other person will do."
This is not strictly purposeful manipulation, but only a statement of
the purpose for which manipulations might be carried out. It can,
however, lead to the other person doing things that satisfy the
manipulator's purpose of understanding. Like the other methods, it
depends for its success on doing nothing that prevents the manipulee
from maintaining control of all variables for which the manipulee has
reference levels.
Now we have three methods of purposeful manipulation, excluding the
one above that really entails the first three:
(a) Manipulation by disturbance of controlled perceptions.
(b) Manipulation by altering properties of the environment.
(c) Manipulation by giving new information.
Note that nothing has been said about whether the manipulee later
judges the results to be good, bad, or neutral, or whether the
application of a method of manipulation succeeds.
Can you think of any more?
----------------------------------------------------------------------
David Goldstein (920909) --
Thanks for the post on Mahrer (and for showing me how to spell his
name).
----------------------------------------------------------------------
-
Best to all,
Bill P.