NetNews Usenet Archive 1993 #1/NN_1993_1.iso/spool/bit/listserv/csgl/2052

Text File | 1993-01-05 | 7.5 KB | 150 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!zaphod.mps.ohio-state.edu!sdd.hp.com!wupost!psuvax1!psuvm!auvm!MCIMAIL.COM!0004972767
Return-Path: <@VMD.CSO.UIUC.EDU:[e-mail]>
Message-ID: <20930105183202/[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Tue, 5 Jan 1993 18:32:00 GMT
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         Hortideas Publishing <[e-mail]>
Subject:      The Devil Influenced Me to Do It
Lines: 138

From Dr. Diabolo, channeled by Greg Williams (920105)

Note to Joe Lubin: MIND READINGS is $18 postpaid most anywhere.

>Bill Powers (930104.0900)

>Even trying to play Devil's Advocate pushes one into the
>qualitative mode. Greg said that it is possible to detect the
>disturbance through its effect on the cursor. That is true,
>qualitativly. Everyone can tell immediately THAT there is a
>disturbance, because the cursor doesn't behave as they expect it
>to under the hypothesis that they have the sole means of
>affecting it. But nobody can tell, on that qualitative basis
>alone, WHAT THE DISTURBANCE IS.

For the tracker to "respond" to the "discriminative stimuli," all that is
necessary is for him/her to be able to see the cursor movement, NOT to
"tell... WHAT THE DISTURBANCE IS." If the cursor is seen to be moving away
from the target position -- due to the net COMBINATION of handle position and
net disturbance, of course -- then the tracker responds by moving the handle
in the direction (determined previously in practice, via "reinforced"
learning) which moves the cursor in the direction toward the target position.

>I should also point out that no matter what method a person uses,
>there is no way of telling HOW MANY disturbing variables are
>acting at the same time, or what their individual magnitudes and
>directions are.

Agreed. Only the net result of all disturbances AND handle position upon
cursor movement is visible. If the net disturbance is zero and an
inexperienced tracker "noisily" jerks the handle so that the cursor heads
upward, the stimulus of solely-tracker-induced cursor movement will result in
a response of change in handle velocity increasing in the direction which
moves the cursor back toward the target. There might be overshoot, in which
case the tracker responds again, appropriately. If the tracker is really
"jumpy," there might be large continuing oscillations in the absence of any
net disturbance. The degree of "jumpiness" or "smoothness" manifested by a
tracker would be reflected in his/her personal stimulus-response functions.

>Greg caught on to this in showing how, by looking at the
>relationship between the handle and the cursor, the participant
>could get more information about the disturbance.

I didn't mean that the person is perceiving the handle position and using that
to compute something about the disturbance. The person doesn't care where the
handle is. He/she simply moves the handle away (in the "corrective" direction)
from WHEREVER it is at a given time IF he/she sees the cursor move away from
the target.

If there are "noisy" functions relating "stimuli" and "responses" for
practiced trackers, so be it. We demons admit that we need to add in random
terms to our models of the PURELY EXTERNAL (TO THE ORGANISMS) FUNCTIONAL
RELATIONSHIPS BETWEEN STIMULI AND RESPONSES. You PCTers boast about a high
correlation between handle position and disturbance, but that glosses over the
fact of irreducible noise in the stimulus-response functions. You simply
integrate and the noise becomes less apparent. We, on the other hand, choose
to tell it like it is: the tracker sees the cursor moving away from the target
(under the influence of net disturbance AND handle position), and the tracker
responds by changing -- with a pseudorandom component -- the velocity of the
handle. If you like correlations, the integral of cursor velocity over time
equals 0 equals the integral of handle velocity over time, but mere
correlations don't identify "stimuli" and "responses." The record of handle
position vs. time does not describe the true "responses" over the course of
the experiment, it is only the RESULT of those responses. There is no
necessity that the responses correlate well with the stimuli when noise is
significant, since the responses tend to correct the "mistakes" of previous
responses.

-----

Rick Marken (930104.1800)

>The experiment does NOT rule out the possibility that there may be
>a relationship between derivatives or integrals of c1(t) and c2(t)
>and, indeed, if you compute the indefinite integral of the cursor
>traces  -- call it int(c(t)) -- you DO find a correlation between
>the integrals -- and there IS a high correlation between
>int(c1(t)) and o1(t), for example. So now the clever nonPCTer
>can get excited and say -- AH HA!! int(c1(t)) IS THE STIMULUS
>that guides responses in a tracking task -- the INPUT-OUTPUT
>MODEL IS SAVED!!!

Again, high correlations don't necessarily identify the real "stimuli" and
"responses" when there is noise. The integration simply washes out the noise
and leads the clever PCTer to get excited and say -- AH HA!! the INPUT-OUTPUT
MODEL IS LOST!!! Of course, the inputs and outputs that got lost were those
chosen by PCTers, not by us devils (and devilettes).

>This is where quantitative modelling is needed again (one little
>demo can't shut the non-PCTer up forever -- if at all). If int(c(t))
>is the stimulus for tracking then we should be able to build a
>model using int(c(t)) as the stimulus.

One last time: don't take high correlations as THE sign of stimulus-response
relationships. Low correlations between stimuli and responses can result in
"good-enough" performance of organisms over the long-run. (I.e., if there is a
two-second constant lag between a light coming on and your throwing a switch
that turns it off, and you want the light off, if the light comes on once
every few minutes, over 24 hours it will be almost always off (nearly perfect
correlation with the desired outcome), yet there will be zero correlation
between the stimulus and the response each time the light comes on.)

>>Where is the person who is claiming that there is ONE function which maps
>>all of the different i's to the same o? They will say that each i has a
>>different function mapping to the same o, which is perfectly possible,
>>mathematically.

>Then I hope they will also say HOW the system knows which function to
>pick each time in order to map the different c(t)'s into the same o(t).

You should know us demons well enough by now to realize that we think trying
to answer such questions is not important.

-----

Gary Cziko 920105.1523 GMT

>Greg, have you gotten the Koza book on _Genetic Programming_ yet?

No, not yet. I expect it this week.

>It seems to me that this would be a nice problem for symbolic regression
>(i.e., function identification) using genetic programming.  You could
>include all kinds of functions that you think could possibly relate the
>output to the input and then let evolution do its stuff to find the magic
>combination.  If there is anyway to express the output as a function of the
>input, it seems to me that genetic programming should be able to snuff it
>out.  These functions could involve time lags as well.

And pseudorandom components, too, presumably. So the task for Koza is to come
up with the function relating cursor position and velocity to handle velocity.

Pat says that it might be appropriate to stress that velocity is a vector,
including a specification of direction as well as speed.

As ever,

Greg