Allen's model; Challenge 2

[From Bill Powers (930330.2000 MST)]

Rick Marken (930330) --

I sort of get it. Actually, it's a mistake to assume that o = d
instead of using the actual value of o. With an integrator as an
output function, the pattern of o will differ from d by an amount
depending on how fast d is changing.

Anyway, it's beginning to seem to me that the point is once again
being lost. If you know what o is, you can exactly reconstruct
the waveform of d from knowing o and p (assuming p is equivalent
to the CEV). You don't need to know anything else about the
control system. We've never said you couldn't do that, or that
some higher-level system capable of perceiving both o and p
couldn't do it. Allan is using information about o in his method,
as I vaguely understand it now, with your help. It shouldn't be
surprising if he can also reconstruct d.

But what if he were limited to the same circumstances as the
control system, which has perceptual information only in the form
of p and has no information about o? Suppose I gave him JUST the
record of the behavior of the CEV, and asked him to reconstruct d
(or for that matter, o) on that basis alone?

By the way, I think this notation H(p|d) is not just an ordinary
function, but represents some sort of probability calculation
with base-2 logs and all that. Allen?


Greg Williams (930330 -3) --

My understanding what you are supposed to do to meet the
challenge I originally issued is to come up with an
"understanding" of a particular sample of my behavior which is
satisfactory to me.

Explaining a sample of behavior doesn't make much sense in terms
of PCT. According to PCT, the challenge was an output intended to
bring about some desired perception. Under somewhat different
external circumstances, some other output might have been used
toward the same end. The same output, for that matter, might have
been used for a different end. The output in itself is of little
interest. The task under PCT is not just to explain the output,
but to find out what it was being used to control for -- what
input was wanted. When that is known, all behaviors that bear on
keeping that perception under control will be understood. Control
of that particular perception may never entail just that behavior
again, and recurrance of that behavior may mean nothing regarding
a particular goal. Trying to explain a "sample of behavior" is a
waste of time. You might say that this is the essence of my
objections to conventional psychology.

If you are unaware of what you expected to get out of issuing the
challenge (or are suffering from amnesia), it's a little
unreasonable to expect me to deduce (from what?) the nature of
your goal as it existed at a time that you no longer remember and
when you were in a state of mind which is apparently beyond
recapture. And I might add, that is beyond my ability to
experiment with. All I can do is to try to discern what some
current goal is, by seeing how your outputs vary with
disturbances now.

Perhaps there is a simple question I should have asked
originally, if I failed to do so: what kind of "understanding"
would be satisfactory to you? Am I aiming at a moving target, or
any target at all? Is there any substance in your challenge?

It has crossed my mind that you are mainly defending against
being understood. That would certainly explain a great deal. I
seem to recall a time at one of our miniconferences when you
announced that nobody could understand you if you didn't want
them to. I recall, I hope correctly, your description of a
strategy that would always make the Test fail: simply change your
mind continually about the nature of the goal. Of course if this
strategy were detected, it would reveal the aim of preventing the
Test from working, in which case the Test would succeed.

I don't know if such a guess is near the truth, and I'm
uncomfortably aware that even to make such a guess might sound
like looking for excuses. However, it's a rather obvious guess,
so it might as well be put on the table. A scientist who is
embarrassed to ask an obvious question won't get anywhere.

This leads to an interesting impasse. Suppose I were to announce
that I have discovered your purpose, which is to show that the
Test can't work on you. To demonstrate this, all I have to do is
use the Test to search for signs of a controlled variable, and
show that every time the Test appears to succeed for a moment, it
ceases to work. Is this evidence that the Test isn't working, or
that it is? Since you are the judge of success, you have both
complete control over the success of the Test in any specific
situation (other than the overall strategy) and the ability to
say that it has failed even when you can see that it has
succeeded -- because that would be honestly consistent with the
same strategy.

Another consideration is whether the Test would always seem to
fail with any person to whom it was applied in similar
circumstances. If the Test were to succeed with everyone but you,
that, too, would be evidence in favor of my hypothesis -- but you
would still have to be the judge, so that kind of "objective"
evidence wouldn't wash.

This raises some other interesting questions in my mind, but I'll
leave them for another time.
I'll put BURN and UNBURN on Silvert's Server, but didn't you and
Pat have a version that could be transmitted in ASCII and then
converted using a debug script, also in ASCII? People who can't
use ftp won't have much use for your programs if they can be
accessed only through ftp, and if they have access to ftp they
can probably download binaries. It would make the most sense to
publish the ASCII-accessible version on the net. Then I could
send SIMCONZ1.ASH directly to them, on request.
Best to all,

Bill P.