[Eetu Pikkarainen 2017-10-09 2]
Martin, Bruce
Thanks again for clarifying my immature expressions about immature thoughts. I try to reply to you both is this same message
[From Bruce Abbott (2017.10.09.0945 EDT)]
BA: The discussion is about the “behavioral illusion,� in which a researcher manipulates a variable that acts as a disturbance to some control system’s controlled variable.
EP: Yes, sorry, I tried to turn it more to the role of causality (by changing the title). It is important question for me (like some others), because I have used to think that human action is not determined causally.
I have earlier made a strict conceptual difference between causal effects and meaning effects. Now I try to get used to the idea that also meaning effects are causal effects.
BA: For the output changes to form a mirror image of the disturbance changes, the reference signal must have been steady and thus not contributing variation to the output. The output is driven by the error,
which of course is a joint function of reference and perceptual signals
EP: This is a nice expression: Human action is error driven. But isn’t it so that if also the reference changes then you cannot say that output is determined by environmental feedback function? (And that is why TCV can be difficult if
the subject’s reference varies.) I just meant that “joint function of reference and perceptual signalsâ€? with the expression “double causalityâ€?: The output is caused and determined always by two causes: external and internal – as a joint ffunction, like you
say.
[Martin Taylor 2017.10.09.09.13]
MT:
More correctly, the history of the mismatch between reference and perception, if the control loop’s output function has any time-binding process such as integration.
Just the mismatch, if not.
EP: Yes, certainly. There must be a mismatch, otherwise they would nullify each other. The integrator output function is really interesting because it in a way maintains the history of past effects like some kind of short
time memory. This is also feature which seems strange when compared with our usual view of causal relationships.
···
EP:
And reference is a part of the internal workings of the organism. Of course it is the characteristics of the environmental feedback function which determines how strong the output must be to cancel
the disturbance. But still it is the internal output function which produces the output.
MT:
Yes, that is approximately true if control is good and if you start with the assumption that the internal processes include an “internal output function”.
If you start with the assumption that the PCT structure is correct, you have a lot less to explain.
EP: Not sure I understand. What alternatives there are? What else could produce the output of the control system than the internal output function? I was really just trying to taste the concept of causality in the PCT
context.
MT:
But even if you do, you still have the question associated with any observation, as Rick keeps reminding use: “Just what perception(s) is (are) being controlled.” For that, the TCV is often useful, especially in the constrained conditions of an experiment.
MT: Once you use your observational data to approximate the perception being controlled, you then have a problem of how to account for the variation in the output magnitude as the disturbance varies.
That part is the issue raised by the behavioural illusion. You can partition variance, as is done in an ANOVA. When we are talking about the output variance, a large part is accounted for by control. what remains is in
the part of the variation of the CEV that is uncorrelated with the output (technically the projection of the CEV variation onto a subspace orthogonal to the output vector). That remaining variance is what you have available for your analysis of the internal
processes.
MT: Each time you add a parameter in your model of what goes on internally, you use it to account for a certain amount of that remaining variance.
The better the control, the less remaining variance there is, so the less you can say about the parameters of your model, or about the relative likelihoods of different models. A significant problem here is that some
of the remaining variance is specifically not accounted for in any of the models. We call it “noise”, though it does have sources such as momentary inattention, slight stickiness in the mouse or joystick in a tracking study, internal neural firing rate noise,
and so forth. Indeed, Bill suggested that the very act of combining neural firing rates into a “neural current” would be fine if it gave results within 10%, because of the stochastic variation of random firings. The actual result is probably much better than
that, but it is always going to contribute part of the variance that you are accounting for with your model.
EP: Good, now I think I understand this better. If you try to study behavior empirically from what you can perceive in the environment (without going inside the subjects nervous system) then the better the control the
less you can say about the internal workings of the subject. Did I get it right?
EP: It is very clear that the relationship between disturbance (“stimulus�) and output (“response�) is a causal one. But it is not a straight and immediate one. There are a series of mediating smaller
causal mechanisms between them like Bruce describes. Secondly there is the circular reciprocal interaction so that the output affects the disturbance as much as disturbance affected the output.
BA: The disturbance is not affected by the output, the disturbance and output both act on the environmental correlate of the perceptual signal, the CV.
MT:
How does the output affect the disturbance? Let’s use the traditional example to make this question concrete. You are driving and are controlling for keeping the car in the centre of the lane. A gust of wind occurs and
you produce an output in the form of a force on the steering wheel. How does this affect the speed or direction of the wind?
EP: Yeah, I was laughing at me when I read these comments! Of course this relationship is also mediated so that disturbance affects the CEV and CEV affects the subject and subject’s output then (later) affects the CEV
and not (necessarily) the original disturbance. I really did not think that turning the wheel would turn the wind!
MT:
Anyway, that kind of reciprocity wouldn’t be “circular causation”. The “circular causation” in a control loop is that a change in any of the variables causes later changes in that same variable. The disturbance is an
external influence that induces change in one of the loop variables (the CEV), and that change causes later change in the output and therefore in the CEV.
EP: True. The output can affect the disturbance, but it would happen outside the loop as a side effect.
(Actually I mistakenly thought disturbance as a disturbance to the subject, i.e. the change in CEV caused by disturbance as an external cause drawn outside the feedback loop in a PCT diagram.)
EP:
And this is a strange thing for the normal causal relationships: normally it goes so that if the cause can be thought of as a directed force then the effect has the same direction as the cause. But
here the
effect is inverse, like you wrote.
MT:
What did I or Bruce write that suggested the force on the steering wheel could affect the wind?
EP: Nothing, and I did mean it neither. What I did mean was this: If we look at an ordinary causal relationship – physical and immediiate - like that between the wind and the car then the effect (the change of the position
of the car from the right side to the middle of the road) takes place to the same direction as the cause (the direction of the wind from right to left). But if the causal chain goes through the control loop then the effect as an output of the driver will be
against to the direction of the wind. And this is a strange thing that we seldom or never see in the physical relations and it explains why we sometimes think that human action (behavior) cannot be causally caused because it often seems to be directed against
possible causes. Like Rick just wrote it seems like closed loop “erases” the causal connection between disturbance and output ([From Rick Marken (2017.10.09.1225)]). But I don’t think it is not erased but rather turned around by
another effect namely that of the reference.
EP:
This is easily explained by the effect of the reference to the causal chain in the comparator. Causal chains can famously branch like the Wikipedia article says: one effect can have many causes and
one cause can have many effects. So the output is caused by two causes: the disturbance and the reference. So perhaps a better slogan for control phenomenon could be double causality than circular causality?
MT: At this point I have lost your chain of reasoning, so I won’t comment further.
BA: Causality within the loop is circular and continuous, so the term “circular causationâ€? seems appropriate there. […]
EP: Certainly. I would not suggest abandoning “circular causation�. I am just thinking from the point of view of a person who has used to think that human action is not causally caused at
all. There it could help to think that the disturbance or any external factor does NOT cause the output ALONE, but only as combined with the reference as an internal cause. So “double causation�: from outside and from inside at the same time.
EP: Yes, the continuity is important and seemingly difficult idea for me. In as practical research (or any) situation when there is some disturbance (caused by the experimenter or by chance)
the perception probably is already in a reference level (or tolerance zone), so we don’t have to (or we cannot) take the reference into account when explaining the change of output?
EP: Sorry if this thinking (too) aloud sound hopelessly messed but I feel that this way my own thoughts become clearer…
Eetu
–
Please, regard all my statements as questions,
no matter how they are formulated.