[From Bruce Abbott (950306.1505 EST)]
Rick Marken (950306.0900)
Bruce Abbott (950304.1925 EST) --
Yes, I was trying to at least hint at that by stating that "subroutine calls
are just outputs of higher-level control systems." It's easy to get trapped
by grabbing hold of some analogy and then taking it too far. Still, the
subroutine notion does capture rather neatly the idea that control passes to
one pre-organized subsystem or another depending on the state of a
perception, such as whether the pigeon's response key is green or red.
But control does NOT pass "to one pre-organized subsystem or another
depending on the state of a perception"; that's why Bill suggested that you
look at the spreadsheet model.
Perhaps "pass" was a poor choice of words; what I meant by this is that the
higher-level system, by setting lower-level references, determines which
lower-level subsystem will be actively controlling the perceptions required
to correct the error in the higher-level system. Of course, the
higher-level system is still controlling its percpetual input, and the
lower-level systems are still controlling theirs. I can visualize that a
system organized to perceive keypecks occurring at a given rate could be set
to a zero reference while the reference to a system organized to perceive
counterclockwise turning is set to a positive rate--the pigeon would stop
keypecking and start turning.
But aren't the keypecking and turning perceptual control systems
preorganized lower-level systems? And by manipulating the references in
this way, isn't one system being, in a sense, "activated" in the service of
the food-magazine perceptual control system while the other one is put on
hold (although, again, both systems are active all the time in that both are
constantly perceiving and controlling their respective perceptions)? But
you are right, my description stayed too close to the appearances. Now that
I've clarified my intent, are we saying essentially the same thing?
A potential obstacle to clear thinking is the fact that discriminative
stimuli as studied in the laboratory represent discrete (usually mutually
exclusive) and clearly differentiated states of some perceptual variable,
such as key color.
I don't think this is the problem; control systems can maintain discrete
variables at discrete reference levels (as is done at the "logic" level of
the spreadsheet hierarchy) just as well as they maintain continuous
variables at possibly continuous reference levels.
A pigeon has been trained to peck while a vertical line is projected on the
key and not to peck when a horizontal line is projected there. What happens
at the logic level when a line at 45 degrees is projected? What the pigeon
actually DOES (at least initially) is peck at an intermediate rate. Does
the logic-level system try to set positive references to TWO lower-level
systems (keypeck rate and, say "exploration"? Or should one conceive of the
logic-level system as setting a single reference (to keypeck rate) to
positive or zero depending on the S-D? But then, how does a "logic-level"
system end up generating an intermediate output to an intermediate input?
Not very logic-like...
Given these possibilities (and they are only two of many that could be
proposed), I think we would do well not to assume, simply because the
discriminative stimuli are presented at discrete values, that we are
necessarily dealing with a logic-level system. I'm not disputing the fact
that a logic-level system CAN handle control of a discrete, logical
perceptual variable. I'm suggesting that we should not automatically
conclude that, in this instance, one actually does, simply because the S-D
I think that the real obstacle to clear thinking about control systems is the
fact that perceptual control _look like_ S-R behavior to the observer; the
S-R illusion is compelling. In "stimulus control", for example, it looks like
light color acts as a _behavioral switch_ turning on pecking when the
light is red and turning on circling when the light is green.
It certainly is ONE obstacle, I agree.
Bill Powers (950306.0930 MST)
As it stands, the SD is not totally necessary; I'm picking up the switch as
much from tracking error as from the color change. I'm still not very good
at it, but it may be worth taking some data and seeing what it looks like.
Is it usually the case that an SD is redundant with some other indication
that the situation has changed?
In the usual operant studies, rewards are usually scheduled intermittently,
making it difficult for the pigeon to discriminate the change in contingency
from the consequences of responding alone. With the addition of the SD, the
observed pattern of output changes immediately with the change in SD, once
there has been sufficient training.
I suggest trying another task. Suppose the two halves of the target were
independently disturbed. The task is to track the half that is red. Above
the position of the tracking area, a score is shown which is a leaky
cumulative total of the error (average of the error on whichever tracking
task is active). Since there is no change in the relation of the handle to
the cursor to signal a change in conditions, the person (with no SD
present) would have to notice the increasing score and switch to tracking
the other target. So now the SD would be the only immediate signal that the
situation has changed. The person would learn to start tracking the other
target before the score had changed appreciably.
How about modifying 3CV1 to include SDs to identify which of the three
cursors should be kept on target? It would be just a matter of adding the
SDs and the scoring system, and modifying the data collection/saving to
include the state of the SD.
By the way, after receiving Rick's post (950305.2000) suggesting the
reversal task, and before receiving this one from Bill, I modified
SDTEST1.PAS to create SDTEST2.PAS, which does exactly as Rick proposes. If
anyone's interested, I could post it, but it suffers from the same
redundancy problem as SDTEST1.