[From Rick Marken (941204.1720)]
Bill Leach (941204.12:14 EST(EDT)--
I suppose that I am remiss by even proposing this without having
tried it myself but I also assert that Bruce's models were control
system models.
Yes. The "law of effect" model was a control model; the parameters of
this model, however, were selected by the consequences of previous
behavior (tumbles) so the model could end up not controlling.
Your own "demonstation of failure" [ of the law of effect model] was
not a demonstration of the failure of the model but rather a
demonstration that the perception under control was not
the one that was being assumed.
I don't think this is quite right.My (and Bill's) demonstration of the
failure of the law of effect model had nothing to do with differences in
the perception under control. We showed that control could not be based
on "selection by consequences"; the law of effect model controls only if
the consequences of behavior are very cooperative -- that is, if they
"select" values for the control parameters that result in control.
Bruce Abbott (941203.1930 EST) --
I have nowhere claimed that the consequences are selecting the
behavior in the sense that you take it.
Here is the sense in which I mean "selecting":
if LNut>0 then
if Diff >= 0 then
put PlusAfterPlus+1 into PlusAfterPlus
else
put MinusAfterPlus+1 into MinusAfterPlus
end if
end if
if LNut<=0 then
if Diff >= 0 then
put PlusAfterMinus+1 into PlusAfterMinus
else
put MinusAfterMinus+1 into MinusAfterMinus
end if
end if
put n+1 into n
put MinusAfterPlus/(MinusAfterPlus+ PlusAfterPlus) into
pTumbleMinus
put MinusAfterMinus/(PlusAfterMinus+ MinusAfterMinus) into
pTumblePlus
LNut is the gradient before the tumble and Diff is the "reinforcing"
consequence -- the difference between the gradient before and after the
tumble. Diff (the consequences of a tumble) directly determines the
probabilities of tumbling in positive and negative gradients
(pTumblePlus and pTumbleMinus, respectively). The effect of Diff is
cumulative, so the current values of pTumblePlus and pTumbleMinus
always depend on the organism's history of experience with its
consequences. This is what I mean (and what I think Skinner meant)
by "selection by consequences"; the probabilities of response --
pTumblePlus and pTumbleMinus -- are selected (determined by)
previous consequences of the organism's behavior.
The whole point of my effort at e. coli modelling was different: to
demonstrate how a set of appropriate behaviors could be acquired and
maintained as a function of an organism's experience with their
consequences.
Well, then we were working on two very different verbal goals. If,
indeed, our goal was simpy to show that "a set of appropriate behaviors
could be acquired and maintained as a function of an organism's
experience", then I take it that you understood all along that the
behaviorist notion that operant behaviors are "selected by their
consequences" is just a description of one of the "behavioral illusions"
discussed in Bill Powers' 1973 Science article. In fact, it was just this
behavioral illustion that Baum didn't see (or understand) in his reply
to Bill's paper.
How else can you represent learning?
In PCT, learning is represented as a change in a structural characteristic
of the model: a change, for example, in the _nature_ of the perceptual,
comparator and/or output functions. Read the chapter in BCP on
"reorganization" for a more detailed discussion. Changes in
parameters of control, such as gain or slowing, can typically be handled
as a part of other control processes.
If a response that led to a mild shock on the last three occasions now
produces a bit of food, will this not tend to increase responding on the
lever?
If your point was only that the law of effect APPEARS to be a correct
description of the situation, then I would have had no argument and
the modelling would have been unnecessary. PCT theorists already
know that control can look like stimulus-response, selection by
consequences (law of effect) and generated output (see my "Blind men
and the elephant" paper). When you build a model of a law of effect
process, however, I presume that you are doing so to show that some
observed characteristic of behavior can be accounted for by the model. I
thought you built the law of effect model to show that control can be
modelled using "selection by consequences" (law of effect). Bill and I
showed that, in general, it cannot.
the "law of effect" model will learn such control, given that the
consequences its learning system monitors are such as to permit it.
The result is a lower-level control system whose parameters have
been determined by experience.
The result is NOT a lower-level control system; the system is already
structured as a control system. The law of effect process will bring the
parameters of this control system (pTumblePlus and pTumbleMinus)
to the "right" values (the ones that result in control) ONLY IF (as you
say) the consequences "permit it". The law of effect model is a control
system with control parameters that are manipulated (selected) by the con-
sequences of behavior. The best that can be said about it is that it's
a lousy control system.
I seem to recall some very strong (and satirical) statements on your
part which came down to the idea that only fools and idiots believe
that there is any value in the law of effect.
Well, I thought I'd sworn off using the word "idiot" but, now that you
mention it...
Actually, I can't imagine having said such a thing. People who believe
in the law of effect are neither stupid nor stubborn; they are just like
the second blind man in my "Bind men" paper. A person who is
blindfolded (that is, a person who is not familiar with the nature of the
phenomenon of control) can only "feel" the "elephant" of control
(using conventional psychological methodology) so this person is being
perfectly reasonable when he describes one aspect of the elephant
(operant conditioning) as being "like a snake" (law of effect) ; after
all, all he can feel is the tail.
Me:
It is my experience that one has a better chance of learning PCT by
taking classes in it than by giving them -- awake or not;-)
Bruce:
Ouch. I think my classes have been quite instructive. However, since
I am evidently not, in your opinion, qualified to teach, I'll be happy to
let you take over.
Sorry, but I got pretty upset about your "sleeping in class" remark to
Tom. Even though that comment was made in jest, it hit a nerve.
Tom Bourbon happens to be one of the very few people who has
actually done extensive, high quality PCT research and modelling.
Tom REALLY understands PCT because he has worked with it --
watching control happen in real live people and in models that match
the behavior of these people to near perfection. Tom hasn't just kicked
the PCT tires; he's gone under the hood and seen how it works.
Because Tom understands PCT, he understands what it means for
conventional psychology and he (like me) has paid the price of this
understanding -- dropping out of an idyllic, tenured position in a
psychology department because he knew that the conventional
psychology curriculum is, well, hogwash. And he has gone through
the ultimate trial by fire; he has tried to publish the results of his PCT
research and modelling efforts in conventional psychology journals.
Tom had the courage to learn PCT and he paid the price (the biggest
price is the fact that there is only one other scientific psychologist in the
world who knows that he's NOT crazy and, worse still, that psychologist
is ME).
So it gets me more than a little ticked off when people who have done
no PCT research and modelling, let alone tried to publish it, start
teaching us about PCT and its relationship to conventional theories of
behavior. We may be stupid as hell but we've been in the scientific
trenches ALONE for quite a while; some of us (like me) are just a tad
shell-shoked.
PCT is not your garden variety "new theory" of psychology; there is a
reason why VERY few people are doing PCT; there's a reason why
Baum reacted strongly to Bill Powers' Science article. There's a reason
why PCT is ignored in all psych texts. Read Bill's 1978 Psych Review
paper; read the 1973 Science article. It's not easy to present PCT
honestly and not seem like you are intentionally "dis-ing" someone's
most cherished idea about what constitutes a fact or theory of
psychology. If you want to learn PCT, you have to at least be WILLING
to CHANGE the structure of your most basic ideas about behavior.
By the way, I'm very well aware of the fact that Tom and Bill and I are
not always right; and we welcome challenges to anything we say about
PCT. In fact, it is important to challenge PCT ideas -- that's how you
learn it. That's how I learned it. But it would sure be nice if this
challenging were done in the spirit of trying to _learn_ PCT. You can't
learn PCT if you already know it.
You ask the following question about Thorkdike's law of effect. My
answers are included:
.
Imagine Thorndike's cat in the puzzle box. It would like to get out of
the box and have a bit of that fish that's in the plate just outside.
1. What brought about the changes in the cat's behavior?
Reorganization; random changes in the structure of and connections
between control systems in the cat; the rate of such changes varies
depending on the deviation of the intrinsic variable (say, nutrient
level) from its reference. The changes are the visible output of the
reorganizing system -- a control system using variations in the rate of
random changes in existing control systems to control intrinsic
variables. Tom Bourbon and Bill Powers have built models of this
process.
2. Do these changes have anything to do with the consequences of
pulling the string? Explain.
Yes. Since pulling the string affects a physical variable (food) that
affects an intrinsic variable (nutrient level), control systems that
control perceptual variables such as "closeness to the string", "force
exerted on the string", etc., will tend to be built and/or connected to
existing control systems in the perceptual control hierarchy.
3. What would happen if you disabled the cat's sensory apparatus so
that it could no longer tell that its string-pulling had in fact unlatched
the door?
Possibly nothing. If the cat can control variables (like the sensed
position of its paw) that get the string pulled, then it might not need to
control the perceptual variables directly related to string pulling.
Deafferented monkeys, for example, can learn to control sensed
variables (like visible arm position) using now unsensed variables (like
felt arm position), though, of course, they cannot control the unsensed
variables so it looks like they are "thowing" their arms to the intended
position. Their ability to regain control of a visual variable after spinal
deafferentiation is evidence of reorganization.
Now I'll go sit in the back with the rest of the class; you come up front
here and fully answer these questions from a PCT perspective.
Thanks for listening;-)
Bill Powers (941204.1020 MST) --
To get here, everybody must start somewhere else.
True. But in 40 years, only three or four people have gotten here (doing
PCT reseach and modelling). People are starting in different places but
they're not ending up at PCT.
We still have a lot to learn about how to serve as guides on the
journey.
Well, I'd say that that's the understatement of the century!! I'm
beginning to think that there's really nothing to learn about how to
serve as guides because people will not be guided unless that want to be
-- isn't that PCT?
Best
Rick