[From Bruce Gregory (970409.1015 EST)]
1. The only way we learn is by trial and error-elimination.
2. The only reason we learn is to extend the domain in which we
can exercise control.
Corollary: Educational efforts that ignore (1) and (2) will be
unsuccessful.
Bruce Gregory
[From Rick Marken (970410.0800 PDT)]
Bruce Gregory (970409.1015 EST) --
1. The only way we learn is by trial and error-elimination.
Trial and error, yes. But I don't think learning _necessarily_ involves
_elimination_ of possibilities. For example, in the
"E. coli" model of learning (illustrated in the "Selection of
Consequences" demo at
http://www.leonardo.net/Marken/ControlDemo/Select.html) the set of
options (directions of movement) from which the bug randomly selects is
the same after every "tumble". Options
are not _eliminated_ if the are not successful; they are just not tried
as long as a particular option is successfully moving the
bug to the target.
I think the degree to which possibilities are actually eliminated
during learning is an empirical question. The fact that learning is
"trial and error" is probably true by definition; if a person
is not trying AND failing to control a variable then that person already
_can_ control that variable.
2. The only reason we learn is to extend the domain in which we
can exercise control.
I think this is correct. though I would say that "extending the
domain" is a side effect rather than an aim of learning. I think
we learn because we are not perceiving what we want to be perceiving.
The perception that is not "right" (at the reference level) might be
a "simple" one, like the perception of the crispness of the ping when we
hit a tennis ball, or a very complex one, like the meaningfulness
of our existence. We try various ways (by trial and error) to get these
perceptions under control (maintained in their reference state).
When (and if) we succeed, we have effectively extended the domain
of perceptions over which we have control.
By the way, the correct answer to the Hans Blom question is:
(7) Not worth worrying about
Best
Rick
[Martin Taylor 970410 12:00]
Bruce Gregory (970409.1015 EST)]
1. The only way we learn is by trial and error-elimination.
I think this is a misconception about reorganization. It refers only to the
class of learning methods that involve random relinkage of elementary control
units, random changes in the weight structure of the hierarchy or random
generation of new control units. It does not refer to gradient (e-coli)
learning, in which there may be a small shift in a random direction followed
by progressive continuation in the same direction (refinement or tuning might
be words for this). And it does not (seem to) cover the kind of learning that
we clearly do, based on observing experts or asking their advice. This latter
is not well covered in my understanding of PCT, but it seems to occur in
humans and other primates--the dissemination of invention within social groups
seems most unlikely to be primarily trial and error-elimination, though I
guess the success of book-learning and imitation could be founded on that.
2. The only reason we learn is to extend the domain in which we
can exercise control.
Or to improve control within the current domain. But you have that word
"reason" in there. I don't think I want to open that can of worms at this
point, but I would like to suggest that worms may be the content of the can.
Corollary: Educational efforts that ignore (1) and (2) will be
unsuccessful.
Put it the other way: educational efforts that accept (1) and (2) have
a better chance of being successful than efforts based on ones that ignore
(2) (I'm not sure about 1).
Martin
[From Rick Marken (970410.1100)]
Me:
I would say that "extending the domain" is a side effect rather than
an aim of learning.
Bruce Gregory (970410.1330 EST)
I take you point. I do think, however, that we can accept invitations to
extend our control to new domains. In this sense, we acquire new things
that we want to perceive. This is what happens whenever education is
exciting and stimulating.
I agree. But I think it is important to remember that (from a HPCT
perspective) we accept these invitations only if controlling these new
perceptions seems to help us control other, higher level perceptions.
Some people do seem to enjoy learning new things just for the sake of
extending their control over the world. And this may be why some
people, who do _not_ seem to have control problems, learn to control new
perceptions (develop new skills): one of their goals may be to perceive
themselves as having many different skills. But even in this case,
the invitation to extent control is accepted because it helps the person
control a perception: the perception of themselves as "having many skills".
I think it's important to keep this in mind because not everyone may have
a goal olike "having many skills". People who don't have such a goal may
find invitations to extend their control a disturbance rather than an
opportunity. I have met people like this. I find it hard to believe that
anyone wouldn't want to extent his or her control. But that's one thing
I've learned from PCT: people really do want the things they seem to want
(Rap music being one example) and they don't necessarily want to learn the
stuff (real music, literature, science, art, craft, etc ) that could
actually extend their control over the world.
People control for what they control for and there we jolly well are,
aren't we;-)
Best
Rick
[From Bruce Gregory (970410.1330 EST)]
Rick Marken (970410.0800 PDT)
Options
are not _eliminated_ if the are not successful; they are just not tried
as long as a particular option is successfully moving the
bug to the target.
Excellent point. Actions remain in our repertoire to be
exercised in other situations or the same situation at other
times.
I think the degree to which possibilities are actually eliminated
during learning is an empirical question. The fact that learning is
"trial and error" is probably true by definition; if a person
is not trying AND failing to control a variable then that person already
_can_ control that variable.
Yes. I think the important educational point is that in order to
learn you must act. Education involves invitations to act in
order to control. I _think_ we respond positively to such
invitations so long as we have not been frustrated too often in
our efforts to exercise control.
> 2. The only reason we learn is to extend the domain in which we
> can exercise control.I think this is correct. though I would say that "extending the
domain" is a side effect rather than an aim of learning. I think
we learn because we are not perceiving what we want to be perceiving.
The perception that is not "right" (at the reference level) might be
a "simple" one, like the perception of the crispness of the ping when we
hit a tennis ball, or a very complex one, like the meaningfulness
of our existence. We try various ways (by trial and error) to get these
perceptions under control (maintained in their reference state).
When (and if) we succeed, we have effectively extended the domain
of perceptions over which we have control.
I take you point. I do think, however, that we can accept
invitations to extend our control to new domains. In this
sense, we acquire new things that we want to perceive. This is
what happens whenever education is exciting and stimulating.
By the way, the correct answer to the Hans Blom question is:
(7) Not worth worrying about
Duly noted 
Bruce Gregory
[From Bruce Gregory (970410.2030 EST)]
Martin Taylor 970410 12:00
> Bruce Gregory (970409.1015 EST)]
>
> 1. The only way we learn is by trial and error-elimination.I think this is a misconception about reorganization. It refers only to
the
class of learning methods that involve random relinkage of elementary
control
units, random changes in the weight structure of the hierarchy or random
generation of new control units. It does not refer to gradient (e-coli)
learning, in which there may be a small shift in a random direction
followed
by progressive continuation in the same direction (refinement or tuning
might
be words for this).
It seems to me that you are assuming that the trials must involve random
relinkage, but that is not necessarily true of all trail and
error-elimination learning. In particular, I think the e-coli learning is
an example of trial and error-elimination.
And it does not (seem to) cover the kind of learning that
we clearly do, based on observing experts or asking their advice. This
latter
is not well covered in my understanding of PCT, but it seems to occur in
humans and other primates--the dissemination of invention within social
groups
seems most unlikely to be primarily trial and error-elimination, though I
guess the success of book-learning and imitation could be founded on
that.
I think even these examples fit the trial and error-elimination pattern.
Imagine that I ask you for advice. How do I know that I understand what you
tell me? My understanding is based on what I think I already know. I have
to cast what you tell me into my own words. This restatement is a
conjecture, whose truth I must test. One way to test the conjecture is to
ask you if this is indeed what you meant. To the extent that I have not
captured your meaning completely, I will eliminate some of my errors and
test again. My point is that learning is an active process and the learner
must act without knowing the results of that action (at least initially).
> 2. The only reason we learn is to extend the domain in which we
> can exercise control.Or to improve control within the current domain. But you have that word
"reason" in there. I don't think I want to open that can of worms at this
point, but I would like to suggest that worms may be the content of the
can.
I introduced this can of worms because I needed to deal with situations in
which the learner does not perceive that his or her limited understanding
needs improvement because he or she is not trying to extend his or her
domain of control.
Bruce
[From Bruce Gregory (970410.2035 EST)]
Rick Marken (970410.1100)
I agree. But I think it is important to remember that (from a HPCT
perspective) we accept these invitations only if controlling these new
perceptions seems to help us control other, higher level perceptions.
Yes, I quite agree.
Some people do seem to enjoy learning new things just for the sake of
extending their control over the world. And this may be why some
people, who do _not_ seem to have control problems, learn to control new
perceptions (develop new skills): one of their goals may be to perceive
themselves as having many different skills. But even in this case,
the invitation to extent control is accepted because it helps the person
control a perception: the perception of themselves as "having many
skills".
I think it's important to keep this in mind because not everyone may have
a goal olike "having many skills". People who don't have such a goal may
find invitations to extend their control a disturbance rather than an
opportunity. I have met people like this. I find it hard to believe that
anyone wouldn't want to extent his or her control. But that's one thing
I've learned from PCT: people really do want the things they seem to want
(Rap music being one example) and they don't necessarily want to learn
the
stuff (real music, literature, science, art, craft, etc ) that could
actually extend their control over the world.
Yes, again I agree. My frustration comes when people _want_ to extend their
control and I can't figure out how to coach them in productive ways. This
is the problem I am wrestling with in the course I will teach again in the
Fall.
People control for what they control for and there we jolly well are,
aren't we;-)
Thou hast said it, Sir Richard.
Bruce
Bruce (direct):
I think even these examples fit the trial and error-elimination pattern.
Imagine that I ask you for advice. How do I know that I understand what you
tell me? My understanding is based on what I think I already know. I have
to cast what you tell me into my own words. This restatement is a
conjecture, whose truth I must test. One way to test the conjecture is to
ask you if this is indeed what you meant. To the extent that I have not
captured your meaning completely, I will eliminate some of my errors and
test again. My point is that learning is an active process and the learner
must act without knowing the results of that action (at least initially).
Spoken like a true Popperian (are you?).
Have you seen Henry Perkinson's _Learning from our mistakes_ (1984,
Greenwood Press)? If not, I think you would enjoy it. It helped me to
understand education and indeed all knowledge acquisition as an active,
selectionist process.
You can see his influence in Chapter 12 of my _Without Miracles_ (now in
paperback).
--Gary
--Gary
[Martin taylor 970411 10:11]
Bruce Gregory (970410.2030 EST)]
Martin Taylor 970410 12:00
> Bruce Gregory (970409.1015 EST)]
>
> 1. The only way we learn is by trial and error-elimination.I think this is a misconception about reorganization. It refers only to
the
class of learning methods that involve random relinkage of elementary
control
units, random changes in the weight structure of the hierarchy or random
generation of new control units. It does not refer to gradient (e-coli)
learning, in which there may be a small shift in a random directionfollowed
by progressive continuation in the same direction (refinement or tuning
might
be words for this).
It seems to me that you are assuming that the trials must involve random
relinkage, but that is not necessarily true of all trail and
error-elimination learning. In particular, I think the e-coli learning is
an example of trial and error-elimination.
Then we understand e-coli learning quite differently. I used it as the
prime example of _non_ trial and error-elimination, and as the method
that is most commonly used. Random relinkage probably happens much less.
Here's how I think e-coli learning works. I'm sure Bill or Rick will
correct me if I am wrong. And I imagine you must disagree with me, so it's
worth setting out, anyway.
(1) We have to assume that there is some kind of a criterion of good
control. That may be error in some intrinsic variable, for example. Or
it may be something like the local RMS error in a single control system,
or it may be something else. It doesn't matter, so long as changes in the
weight structure of whatever is doing the learning (either inputs to
a Perceptual Input Function or the distribution to lower levels of the
output) change the criterion smoothly over most of their range of variation.
And there exists some optimum point (call it a minimum squared error if
you want an example).
(2) There is a parameter set, called "weights" in the preceding paragraph.
Variations in this parameter set affect the criterion. The parameters can
be considered as basis vectors in a high-dimensional space, and the values
of the parameters determine the location of a point in this space. The
criterion is a single-valued smooth function of location in this space.
(3) Parameter values can be varied by small or large amounts and the effect
on the criterion of the variation evaluated.
(4) Procedure:
(a) Evaluate the criterion at the current location.
(b) Make a small change by moving the location of the point in a random
direction.
(c) If the criterion gets worse, return to (b).
(d) If the criterion gets better, make another change in the same direction
again (some variants increase the size of the change, some don't).
That's it. I see no trial and error-elimination in that, unless there is
some meaning to the term as you use it that covers this rather precise
tuning procedure. Or, do you see the e-coli procedure as different from
this?
Notice that this procedure never ends. When the criterion reaches its
optimum value, the e-coli still makes random small changes. If external
circumstances alter the location of the optimum, our e-coli will move
to the new optimum, provided that the old one is not a local optimum
in the new configuration. E-coli can get caught in local optima, but
the higher-dimensioned the space, the less likely it is that a point
is an optimum in all directions if it isn't a true optimum. There's
usually a valley somewhere that leads downhill, though in high-dimensional
spaces narrow valleys can be exceedingly hard to find, so it may take
a while for the e-coli to get out of a local constriction. It will
eventually, though.
I just thought of a way you might see this as "error elimination." Perhaps
you think that e-coli won't again try a direction that didn't work on an
earlier random try. That's not true. It won't try a direction that proved
to be bad once it finds a good direction, because once it finds a good
direction it just keeps going in that direction until it's gone too far.
But until then, and again at the next "turning point", it may well retry
a direction that was bad. The procedure (in its pure form) has no memory
beyond keeping going so long as the criterion keeps improving.
Your turn.
My point is that learning is an active process and the learner
must act without knowing the results of that action (at least initially).
If that's your point, I couldn't agree more. I didn't see it in your
language. But that's a hazard we all face, don't we:-(
Martin
[From Bruce Gregory (970411.1040 EST)
Gary Cziko 970410
Spoken like a true Popperian (are you?).
More like a true Powersian. But I have read a good deal of
Popper. He is not enough of a pragmatist for my taste. (He of
course would be delighted with this assessment!)
Have you seen Henry Perkinson's _Learning from our mistakes_ (1984,
Greenwood Press)? If not, I think you would enjoy it. It helped me to
understand education and indeed all knowledge acquisition as an active,
selectionist process.
No, but I have just finished reading _Teachers Without Goals;
Students Without Purposes_. I very persuasive little book.
You can see his influence in Chapter 12 of my _Without Miracles_ (now in
paperback).
_Without Miracles_ was my introduction to bother Powers and
Perkinson. Thanks!
Bruce
[From Bruce Gregory (970411.1125 EST)]
Martin taylor 970411 10:11
I just thought of a way you might see this as "error elimination." Perhaps
you think that e-coli won't again try a direction that didn't work on an
earlier random try. That's not true. It won't try a direction that proved
to be bad once it finds a good direction, because once it finds a good
direction it just keeps going in that direction until it's gone too far.
But until then, and again at the next "turning point", it may well retry
a direction that was bad. The procedure (in its pure form) has no memory
beyond keeping going so long as the criterion keeps improving.
Yes, Rick made this point to me, too. The word elimination is
misleading in this case (and it may be misleading as well where
natural selection is involved). It is probably clearer to talk
in terms of "error rejection". Memory is not required. In fact,
we see this return to earlier rejected solutions in many forms
of learning. E-coli rejects errors, but does not eliminate the
actions that led to them. It's hard to believe that we do
either. This mechanism, which does _not_ require permanent
changes to the hierarchy, may be a lot more prevalent than I
imagined. Hmmmm.....
> My point is that learning is an active process and the learner
>must act without knowing the results of that action (at least initially).If that's your point, I couldn't agree more. I didn't see it in your
language. But that's a hazard we all face, don't we:-(
Too true...
Bruce