adjusting schedules

[From Bill Powers (960403.1330 MST)]

Chris Cherpas (960403.0911 PT) --

     A procedure is then described which, as Bill P. points out, can be
     read as containing a crucial contradiction. However, I read this
     as just loosely describing an "adjusting" schedule, in which a
     single response _is_ immediately followed by food delivery (i.e.,
     there is no time delay between a response and a food delivery), but
     that responses also incremented a counter which would subsequently
     have to time out for the next response to be (immediately) followed
     by a food delivery. The ambiguity could be resolved.
     ...
     I didn't think Bruce was doing this experiment, but since he _does_
     run "cleverly designed" experiments, I began to be fooled again --
     this time _hoping_ it was true, since this is a general area in
     which I did my dissertation.
     ...
     After all, this sounded like a "Melioration versus Maximization" (a
     favorite) type of experiment.

Oh-oh. To me, this suggests that

(1) You have done a considerable amount of work with "adjusting"
schedules of this kind and have spent a lot of effort looking for
regularities in the behavior of animals under such schedules. You have
become an expert on behavior under "adjusting" schedules.

(2) You are not actively looking for an explanation of the observed
behavior that will call any of the conclusions of your dissertation into
doubt.

(3) You have an investment in the categories "melioration" and
"maximization" which will not easily be written off.

If I am right about (1), (2), and (3), or any two of them, then it will
be difficult for us to come to any kind of agreement about a PCT
explanation of this behavior, because the PCT explanation is quite
likely to be simpler than the one you have found, the regularities you
observed are likely to prove to be forced interpretations of a much
simpler situation, and the terms melioration and maximization will not
be used (control systems do not maximize, nor, probably, do they
meliorate. They control). If you realize that this is a quite possible
outcome, and still want to see what PCT might have to say about this
behavior, we can go ahead. But if my apprehension about getting into
this is justified, we would be better off leaving the subject alone.

I will assume the best.

···

----------------------------
In trying to imagine the actual arrangement of the adjusting schedule
from your description (if I misinterpret, please correct me), I have
been able to construct a rough picture of the feedback function.

Let c be the current counter value, b be the current behavior rate, and
k be the amount by which c is decremented per second. C is incremented
by m*b counts per second.

I assume that a behavioral event will produce a reinforcer only if the
timer has counted down to zero.

If k is less than m*b, that is, if the amount by which the counter is
decremented per second is less than the amount by which it is
incremented per second, the count will never reach zero and the
reinforcement rate will be zero.

If k is equal to m*b, the counter will be incremented and decremented
equally per unit time, and the outcome is ambiguous. If the timer is not
counting down to zero, it will never get to zero and there will be no
reinforcements.

If k is greater than m*b, the counter will count down faster than it is
incremented by the behavior, and will eventually reach zero. Once it
reaches zero, the next behavior will produce a reinforcement and the
counter will be incremented by m counts. The counter will then begin
counting down by k counts per second. Because the behavior rate is low
enough that k is greater than m*b, the counter will time out after each
behavior and there will be one reinforcement per behavior.

So if the behavior rate is less than or equal to k/m, there will be one
reinforcement per behavior; otherwise there will be zero reinforcements
per behavior.

The transfer function of the schedule can thus be plotted this way:

     >
     >
     >
Reinf>
Rate | | B = k/m
     > >
     > v
     ><------------->*
     > R = B *
     > *
     > *
     > *
     > *
     > * <--- R = 0 --------------------->
     > *
      ----------------* * * * * * * * * * * * * * * * * *>>
                         Behavior rate

Does this agree with your analysis, or are there some details I don't
know about? (Assuming you want to proceed).
-----------------------------------------------------------------------
Best,

Bill P.

[From Chris Cherpas (960403.1545 PT)]
   [re:> Bill Powers (960403.1330 MST)]

Oh-oh. To me, this suggests that

(1) You have done a considerable amount of work with "adjusting"
schedules of this kind and have spent a lot of effort looking for
regularities in the behavior of animals under such schedules. You have
become an expert on behavior under "adjusting" schedules.

Nah. Not true. If anything, I'm more of an "expert" on concurrent schedules,
but I'm no Dick Herrnstein either. Sorry if I gave you the wrong
impression.

(2) You are not actively looking for an explanation of the observed
behavior that will call any of the conclusions of your dissertation into
doubt.

Not true. Explanations I accept do not have to conform to notions I
entertained during my dissertation (nor during elementary school).
They might be easier to accept, though. You give me credit for being
more conscientious than I am. Are you getting enough fiber?

(3) You have an investment in the categories "melioration" and
"maximization" which will not easily be written off.

I have invested in those categories, but there can be multiple ways of
understanding. When I have difficulty with a given view, I try out
a new one for a while and watch for what happens next. If it takes a while
to really appreciate the new view, and I have to make a decision
in the moment, I'll revert back to the default view for a while, defending
it here, foresaking there, but trying out the new view, until such time
that it takes over as the new default position. And maybe it never
does. What do you do?

If I am right about (1), (2), and (3), or any two of them, then it will
be difficult for us to come to any kind of agreement about a PCT
explanation of this behavior, because the PCT explanation is quite
likely to be simpler than the one you have found, the regularities you
observed are likely to prove to be forced interpretations of a much
simpler situation, and the terms melioration and maximization will not
be used (control systems do not maximize, nor, probably, do they
meliorate. They control).

No problem. When I see reasons for agreement I'll eventually agree.
(Meanwhile, if we don't see control, then we must not be looking at a
controlled variable, right? I can recite the same kind of rap in
reinforcement theory: "if doesn't act like reinforcement, it isn't."
It really doesn't bother much me at this point. Does it bother you?)

A reminder: I've had much less invested in my technical categories
than you've had in yours; I'm not proud, just less skilled in
developing, and less invested in, a particular theory -- and a little
younger.

But, by the way, don't control systems minimize error? Can't that
be called optimization?

I will assume the best.

What would that be? By "best" are you implying some sort of
optimization? Is PCT the best of all theories or just consistently
(at least) minimally better than others you've encountered?

Let c be the current counter value, b be the current behavior rate, and
k be the amount by which c is decremented per second. C is incremented
by m*b counts per second.

...

So if the behavior rate is less than or equal to k/m, there will be one
reinforcement per behavior; otherwise there will be zero reinforcements
per behavior.

Does this agree with your analysis, or are there some details I don't
know about? (Assuming you want to proceed).

That version's OK. With only one explicitly defined behavior X, we could
discuss the (albeit delayed) reinforcement of "not X" behavior or of the
differential reinforcement of an IRT (interresponse time). I've studied choice
situations where you can more obviously define what an "optimal" allocation
between alternatives might be (versus some sub-optimal mix that makes the
alternatives appear no different locally). Bruce's "procedure" reminded
me of this kind of local versus global situation. There _are_ people who
spend lots of effort doing explicit response and not-response schedules.

Is your goal to understand what the term "adjusting schedule" means?
There are lots of ways to adjust schedules, often based on some dynamic
property of the performance. But I guess all schedules "adjust" in
some sense if there's any contingency at all (but a so-called Fixed-Time
schedule just dumps food on the subject, independently of behavior).

If your goal is to convince me that I can understand Bruce's
experiment without invoking maximization or melioration, you
need not be concerned. I am gradually learning to see the world
from a control perspective (of course, according to control theory,
I'm already doing that); I may eventually reject it, but I doubt that.

Perhaps one must become a "reborn" control theorist: we a born control
theorists, we get indoctrinated in various misconceptions, and then we
see the light again through discipline and experimentation.

Saved but not healed,
cc