[From Bill Powers (2009.04.24.0144 MDT)]
Rick Marken (2009.04.23.1745) --
> BP:But you're misreading and misquoting what he says.
Ah, this is more like the Bill Powers I know and love; it was a little
weird when you said you agreed with me earlier. Now I feel much
better;-)
It's a vexing situation for me, but I feel that when you put words into Martin's mouth, it greatly weakens your arguments even when I agree with them.
Yes, I know that Martin was talking about ONE COMPONENT of the loop.
It just seemed irrelevant.
But he thought it was relevant and you chose to talk about something else as if he had meant what you talked about. That's not playing fair. That's like attacking Iraq because Al Queda attacked us. If you're going to criticize what he says, at least criticize what he says, not something else for which you happen to have a ready-made argument.
He thought he could use a model to measure just part of the forward path through the system. Don't we do something similar when we deduce/assume a perceptual gain and comparator gain of one, a perceptual delay of 150 msec, and an integrating output function with a certain gain and time constant? We have no way to verify each of those properties by itself; what we get is the best-fit values under the assumption that the whole model's architecture is right.
But maybe I'm wrong. What we observe in a
conventional experiment is DV = f(IV). Martin is claiming that this
relationship is a valid measure of a characteristic of ONE open loop
component of the link between IV and DV:specifically, the perceptual
input function.
What the Schouten experiment does is the equivalent of assuming a delay in the input function, and arranging to force the output to occur at a specific time (one of the independent variables, another being when the light turns on). In my diagrams of a couple of days ago, I showed how the observed uncertainty data could work out even if the identification of the light was always correct. My version of the model puts all the uncertainty into the detection of the signal identifying the light after the decision has been made. This is done, I now propose, by making the reference signal turn on at a specific delay after the onset of the light intensity. The Schouten experiment gives us another degree of freedom (or several more) to put in the model. The other parameters would still remain to be determined. And of course Martin introduced the idea of the imagination connection at the second level, which is still another independent variable to worry about. And there's the problem of timing the output relative to the beeps -- it's a mess.
The worst problem is that the experiment is set up with discrete variables which make it impossible (as far as I can see right now) to deduce things like the gain and time constant of the output function in the assumed standard model. When the output is sampled at only one instant during every trial (the moment that the contact in the button closes), the dynamic characteristics remain unknown. Perhaps if a continuous-variable experiment had been done we might have been able to get more out of it. I don't know -- model-wise, it's a horribly complex experiment, and there just aren't enough degrees of freedom in the data to get a working model out of the numbers. We have to imagine a lot more than we can observe.
I have to point out that with discrete variables in the model, the calculations used for demonstrating the behavioral illusion don't work. What is the inverse of the effect of a contact closure on a relationship? You seem to get irritated with me when I point out such problems, but what should I do? Keep quiet and hope nobody else notices them?
But we know from your analysis (in the 1978 Psych
Review article) that, when the system under study is closed loop, the
function, f(), relating IV to DV, is the inverse of the environmental
feedback function, g(), relating DV the CV: CV = g(DV). In other
words (as you often emphasize) the apparent system function, f(), is
actually the inverse of the feedback function, g() -- f() = 1/g() --
which measures a property of the environment, not of the system
itself.
That's approximately true, assuming very high loop gain. The actual solution of the equations is more complex since both forward and feedback functions appear in it, and with a varying reference signal the simple inverse can't be computed unless you have some way of estimating the changes in the reference signal. And as I just said, those equations are hard to apply when the functions involve discrete on-off variables.
While in specific experiments you can demonstrate the behavioral illusion quite nicely, this is done because knowing control theory, you can make sure the reference signal stays constant, you can also be sure the loop gain is high enough to make the approximations valid, and you can use continuous variables. I think that a much more general approach is simply to point out that the controlled variable is a function of the action and of disturbances, so the perception actually being controlled is not a perception of the disturbance alone, or of the action alone, but of some combination of the two. That means that the stimulus you think you're applying to the organism is not the stimulus that the organism experiences. That's enough to sink S-R theory.
This, of course, is the behavioral illusion (and ,as you have
also mentioned, the way a lot of measurement is done using analog
computers). My response to Martin was based on the assumption that f()
-- the relationship between IV and DV observed in a conventional
experiment -- reflects only properties of the system's environment and
that, therefore, there is no way that it could be a valid measure of
any property of a closed-loop system; and the perceptual function is
one component property of such a system.
The observed relationship between IV and DV does not, in general, reflect ONLY properties of the system's environment. In fact it reflects NEITHER the forward nor the feedback function exactly. When the loop gain is low, it reflects mainly the forward function. When the gain is high, the feedback function. Between "low" and "high," both. The point of most importance is that the input experienced by the organism is not a simple reflection of the disturbance, or what is known as the stimulus in SR theory. That generalization remains true even when the extreme one you use isn't true.
Since no test for the CV was done in the Schouten experiment
-- nor is such a test done in any other conventional experiment -- I
think it's fair to say that it is impossible to learn about the open
loop characteristics of the system or of open loop COMPONENTS of the
system using conventional methodology since the hallmark of
conventional methodology is the failure to test for a CV.
I agree with the need to test for a CV. Martin is assuming he knows what the subject is perceiving. In normal psychology, all you really have to say is that it would be absurd to assume anything else. You mean the subject acts as if he's seeing the light but he isn't, even though he responds to it? What are you, some kind of nut? That passes for scientific method.
In PCT, we're trying to be more rigorous when we can. If we see a way to test the idea that any assertion is true, we put it into the experiment. That makes the results much more solid, more believable. Ideally, verifying the variables would amount to solving N equations in N unknowns and getting the one solution that is possible. In the tracking model we have just enough variables and just enough relationships to get a unique solution, which we do by running an analog simulation. The simulation is really an analog computer that solves the simultaneous equations. The results are still model-dependent, but at least we have demonstrated internal consistency. That is NEVER done in normal psychology, as far as I have seen.
I'm not carrying on a vendetta against old psychology. It's just that I have never seen an old experiment that I could accept at face value and use as data with any confidence. Every time I have looked into the details, I have seem problems, OBVIOUS problems, with the observations and with the conclusions. That's not true of every experiment in every discipline. But it is true of every experiment I've seen in psychology.
The reason isn't that old-time psychologists were unobservant or stupid. They noticed lots of interesting things; they just didn't have any way to understand them that would hold water. PCT opens our eyes to basic important relationships that they couldn't have imagined. That's why we can see defects where they couldn't. They didn't know what questions to ask. And that's why I don't see much to gain in trying to use old experiments. They all need to be done over again, and if the phenomena were important, they'll come up again as we investigate living control systems.
Best,
Bill P.