Diagram of problem

[From Bill Powers (951124.1105 MST)]

Martin Taylor 951124 11:00 --

As you know, I think best in diagrams. Here is a diagram of what I think
you're talking about:

                      ref
                       >
         ----->---- Comp 2 ----->--------
        > >
        > >
        > >Ref
        > -----> Comp 1 ----->
        > > >
        > > >
        > LOOP 2 | LOOP 1 |
        > > >
      CEV 2 CEV 1 ACTION
        > > >
        > <--- LINK 1 --<-----
        > >
        > > SIDE-EFFECT
         ------<------LINK 2 ---<-------

Imagine that LINK 1 is a complex set of variables that convey effects of
ACTION to the first controlled variable, CEV 1. The side-effect can come
off of any intermediate variable from ACTION itself to CEV 1 itself. The
side-effect is the effect of ACTION that alters CEV 2 and is used by
LOOP 2 to control it.

Imagine that buried inside LINK 1 is the "actually effective variable."
This variable has to be included in LINK 1 for any LOOP 1 that is
chosen, if an effect on CEV 2 is to occur. You can imagine different
LOOP 1's involving different actions and different CEV 1's, but in the
link between ACTION and CEV 1 in each case, the actually effective
variable must be part of the link, somewhere between the action and its
effect on CEV 1. Reorganization based on errors in CEV 2 will stop when
this is the case.

Control of CEV 2 is most certain and direct when a LOOP 1 is chosen in
which the actually effective variable is identical to CEV 1. CEV 1 is
protected against disturbances of ACTION or LINK 1, and will follow the
reference signal given by LOOP 2's output. When the actually effective
variable is identical to the variable by which we measure ACTION,
disturbances of CEV 1 or entering LINK 1 would change the action and
interfere with control of CEV 2.

I think this covers all the cases you have introduced.

As to "error free learning," is there really such a thing? The only
example I can think of is memorizing, which we should probably not call
learning at all.

     You are right that most rituals don't achieve the desired results
     because of anything in the ritual itself (though some do, and I did
     give a real life example that you now ask for--willow-bark tea).

But in that case, you can vary all sorts of things about the ritual and
still have success: leave out the shreds of bark after steeping, get rid
of the tannin, change the temperature, serve it with the other hand or
just ladle in into the mouth, leave out the chants, etc. etc..

     However, I disagree that people have faith in them because they
     have no other way to get what they want. If that were the case,
     one would expect the rituals to be destroyed by reorganization. We
     need some new postulate to explain why that doesn't happen

No, we don't. When you're as near to zero error as you're going to get,
you just keep reorganizing, with the organization doing a random walk in
the vicinity of zero intrinsic error. You may have moments of doubt, but
you'll reorganize back again, because nothing else within the local
minimum reduces intrinsic error any further.

Oh, dear. This game is very seductive, and I know why I'm indulging in
it. The alternative is to sign off and go lug some rocks around in my
yard. I seem to have passed a balance in the conflict, because what I'm
going to do now is lug rocks. It sounds like marginally more fun.

···

-----------------------------------------------------------------------
Best,

Bill P.
----------------------------------------------------------------------
Best, Bill P.

[Martin Taylor 951124 14:25]

Bill Powers (951124.1105 MST)

I tried to make a diagram like yours, but gave up half-way through. Let me
try a couple of variants, which may diagram what I think your text says.

Variant 1.

                      ref2
                       >
         ----->---- Comp 2 ----->--------
        > >
        > output 2
        >perception 2 |Ref 1
        > -----> Comp 1 ----->
        > > >
        > >perception 1 |
        > LOOP 2 | LOOP 1 output 1
        > > >
      CEV 2 CEV 1 ACTION
        > > >
        > <--- LINK 1 --<-----
        > effects only on |
        > CEV1 | SIDE-EFFECT
         ------<------LINK 2 ---<-------------------- (effects on the part
                                                    of the universe orthogonal
                                                      to CEV 1)

In this variant, ref 1 does depend on the output 2, but perception 2 does
not incorporate perception 1 in any way. Perception 2 is orthogonal to
perception 1 (and therefore CEV2 is orthogonal to CEV1). In this diagram,
the "actually effective variable" is in the effects on the part of the
universe that is orthogonal to CEV1. It is irrelevant to the control of
perception 1 whether the "actually effective variable" in LINK 2 is
influenced by output 1. The actual setup of the environment makes that
connection.

Variant 2.

                      ref
                       >
         ----->---- Comp 2 ----->--------
        > >
        > percception 2 |
        > >Ref
        >------<---------------------> Comp 1 ----->
        > > >
        > > >
        > LOOP 2 | LOOP 1 |
        > > >
      CEV 2 -CEV 1 ACTION
        > > > >
        > V <--- LINK 1 --<-----
        > > >
        > > SIDE-EFFECT (?)|
         ------<---LINK 2 -----------<---------------

In this diagram, perception 2 is not orthogonal to perception 1, which
is equivalent to saying that changes in CEV 1 influence CEV2. Now CEV 1
is correlated with the "actually effective variable", and in the limit,
when CEV 1 becomes the AEV and the "SIDE-EFFECT" component of LINK 2
goes to zero, it is control of perception 1 that permits control of
perception 2.

In variant 1, the SIDE-EFFECT could be achieved in many ways not diagrammed,
none of which would affect the control of CEV 1. In variant 2, control
of CEV2 entails control of CEV 1, and there are no necessary effects on
the part of the world orthogonal to CEV 1 (SIDE EFFECTS).

Imagine that buried inside LINK 1 is the "actually effective variable."
This variable has to be included in LINK 1 for any LOOP 1 that is
chosen, if an effect on CEV 2 is to occur. You can imagine different
LOOP 1's involving different actions and different CEV 1's, but in the
link between ACTION and CEV 1 in each case, the actually effective
variable must be part of the link, somewhere between the action and its
effect on CEV 1.

I imagine not a single link, but a fan-out of effects of ACTION 1 on the
universe, some of which compose LINK 1, and the rest of which compose
SIDE-EFFECTS. So I don't buy the verbalism that "the actually effective

variable must be part of the link, somewhere between the action and its
effect on CEV 1."

The word "between" is what I don't buy.

Variant 2 is what you said:

Control of CEV 2 is most certain and direct when a LOOP 1 is chosen in
which the actually effective variable is identical to CEV 1.

The simplistic situations of Jack-the-cat and of the jacket-flipping dancer
both have one well-defined "actually effective variable," so that Variant 2
is unique. Variant 1, however, has a wide range of possibilities, as you
note. Furthermore, in variant 1, changes in the environment affect whether
the SIDE-EFFECT includes the "actually effective variable" whereas variant 2
is robust against changes in the environment (which are disturbances to
CEV1 or to LINK 1, both countered by the control action of LOOP 1).

One question of interest is under what kind of circumstances is Variant 1
likely to be transmuted to Variant 2 (i.e. by what changes in CEV 1 is
it likely that CEV 1 is chosen so that it becomes non-orthogonal to, or
an essential component of, CEV 2). One part of the answer I have suggested
is that the more CEV1's there are for which ACTION 1 affects the "actually
effective variable" the more probable it is that this AEV will be a
component of some controlled perception--a part of some CEV 1 that is
encountered during random reorganization. There are probably other
criteria that are less trivial.

When there are several possible independent AEVs (many means to the same
end), the situation is a bit different, and that is what I tried to
address by giving the cat a stick, a floor switch, and some other way
of getting the door open. Under these conditions, it's perhaps easier
for the cat to find a way to get out of the box, but harder to find what
critical possibilities are robust against environmental variation. It can
test whether escape is possible without touching the stick, and find that
it is--so the stick is clearly not a critical aspect of whatever worked
the last time. Neither is the floor switch, but both successful escapes
(one using the switch and one the stick) involved a particular move against
the side of the box a little earlier. Maybe that was the critical consistency
in the actions. It's quite possible that having many ways to get what you
want might sometimes be even more ritualizing than having only one truly
effective variable, especially if the "effective" variables are unreliable.
(In case you didn't notice, that's pure speculation and an invitation for
someone to look at the maths or to simulate, to see whether such conditions
can exist).

···

----------------

As to "error free learning," is there really such a thing?

Not if you take "error" in the PCT sense, I'm sure. But you can take it
in a different way, using the above diagrams. In "error-full" learning,
we are dealing with Variant 1, and reorganizing by trying different CEV 1s,
some of which have the AEV as part of their side effects, and some of
which do not. When the "learner" is using a non-effective CEV 1, CEV 2
is out of control, and that's called an error. It might better be called
a "mistake", but language usage is what it is, and "error" is the word
people seem to use.

As I understand it, "error-free" learning occurs when the situation is
as in Variant 2, where CEV 1 is not necessarily identical to the AEV, but is
correlated with it. Learning then involves a smooth transition in which
the components of CEV 1 that are orthogonal to the AEV are eliminated,
so that CEV 1 becomes the AEV rather than just entailing it. Control of
perception 2 is never lost, but becomes better and more precise with
learning. In the PCT sense, there is always error in each loop, but not
in the other sense, of "mistake".

    You are right that most rituals don't achieve the desired results
    because of anything in the ritual itself (though some do, and I did
    give a real life example that you now ask for--willow-bark tea).

But in that case, you can vary all sorts of things about the ritual and
still have success: leave out the shreds of bark after steeping, get rid
of the tannin, change the temperature, serve it with the other hand or
just ladle in into the mouth, leave out the chants, etc. etc..

Sure, you can vary them, but how many do? If you vary what works, you are
dealing with experiment, not ritual. These kinds of effective
witch-doctory are usually incorporated into a larger ritual that is not
(much) varied. It's the scientist who tries the variants and discovers
a substance that is easily synthesized into little white pills that give
him the headache of making a lot of money (as a side effect:-).

    However, I disagree that people have faith in them because they
    have no other way to get what they want. If that were the case,
    one would expect the rituals to be destroyed by reorganization. We
    need some new postulate to explain why that doesn't happen

No, we don't. When you're as near to zero error as you're going to get,
you just keep reorganizing, with the organization doing a random walk in
the vicinity of zero intrinsic error.

Whoop-e-do, here! We're talking, in the sentences immediately preceding
what you quote, about cases in which the error is FAR from zero, cases in
which people do NOT get what they want, and yet persist in the rituals.

Now, that may be as near to zero as these people are going to get, but
when "as near as you are going to get" is far away, there are lots of
ways to be just as far, and the question is why one particular way is
stabilized. It is that stability for which I conjure the feedback effects
of dealing with other people who use the same rituals. There are reference
signals, associated with perceptions of those other people, that are better
satisfied by staying with the ritual than by abandoning it, even if the
ritual does nothing in respect of the primary object of desire.
---------------

Oh, dear. This game is very seductive, and I know why I'm indulging in
it.

Don't you like lugging rocks? It can be a very rewarding experience, once
the stiffness has worn off! I hope what you mean is that you like having
fun with ideas AND with rocks. I'd hate to think it was the other way
round, with no fun either way.

The game is seductive, and like most good games, it has a serious point
for life outside the game. Come on, other CSG-L-ers! Join the game!
This one is not a zero-sum game, nor do the rules restrict it to two
persons.

Martin