Causation

[William_T_Powers2]

[From Bill Powers (920615.0900)]

Mark Olson (920614.1700) --

1) Outcomes are influenced by the environment and reference signals.
   Outcomes are determined by reference signals.
   Actions are influenced by the environment and reference signals.
   Actions are determined by the environment.
   Actions control outcomes.

2) The cause of an outcome is the reference signal with the
   environment as the enabling condition.
   The cause of an action is the environment, with the reference
   signal as the enabling condition.
   Actions control outcomes.

This kind of summing-up is a good idea, I think. Let me throw in some more
considerations for you to sort through.

I think "the environment" is perhaps too general, because properties of the
environment enter differently from independent variables in the
environment. A lever has the property that pushing down on one end makes
the other end go up. So it determines that IF you want the far end to go
up, THEN you push down on the near end. Skinner used the term "contingency"
for this: movement of the far end is contingent on movement of the near
end. Not a bad term.

But the lever neither causes nor influences anything by just sitting there.
The variable position of its far end, given the lever's properties, is
influenced in a particular way by the action of moving the near end. It is
also influenced by independent forces acting anywhere along the lever. The
lever doesn't determine whether any such actions or forces will occur.

Maybe this could be summed up by saying that the KIND of outcome of an
action is determined by properties of the environment, while the STATE of
the outcome is influenced by action and independent variables: this comes
down to JOINT DETERMINATION of the state of the outcome by action and
disturbance, given the properties of the feedback connection.

Of course this doesn't bring purpose into it yet. There are always joint
outcomes of action and disturbance, but INTENDED outcomes are a subset:
those maintained near a given state by variations in action.

Shoot, now I'm getting into details that seem to go on forever. For
example, what is an "outcome?" Isn't it defined by perception? When you act
on an environment you cause a never-ending series of shifts in
relationships among thousands of detailed variables. It's perception that
puts order into these variables, represents them as abstract things like
"objects." Outcomes are defined in terms of things we perceive. While
perception doesn't CAUSE the physical variables to be in certain states, it
SELECTS certain functions of those variables as significant aspects of the
world, and intended outcomes are among those aspects.

Oh, blah, blah, blah. Suddenly I'm wondering if there really is any way to
reduce the relationships to a sequence of simple statements of influence,
determination, or causation. Those concepts are all really two-element
ideas: the influence or determiner or causer, and the influenced,
determined, or caused. Maybe all this boils down to trying to express
circular causation in terms of lineal causation. I'm beginning to see why
Aristotle ended up with four kinds of causes.

I guess the problem with your summing-up is that it makes perfect sense if
you already understand the principles of control, but that each statement
requires endless explanation for someone who doesn't. In a real control
system, there are all kinds of determinations and causations and influences
going on at the same time. Perception defines what amounts to an outcome.
The environment, given that definition, determines what actions would be
needed to affect that outcome in any given direction, and also determines
how any remote independent variables that may exist can affect that same
outcome. Reference signals specify which state of a perceived outcome is to
be maintained; disturbances that cause errors relative to that reference
state then determine what amount and direction of action, applied through
properties of the environment, will occur.

I feel the same way you do: there has to be a way to express all this as a
series of easily-understood statements. But after all the above, I'm
beginning to wonder whether ordinary language doesn't contain too much SR
theory for this to be done using existing words. Just talking about causes
and effects is a problem, and influences and determinations really don't
help that much. Maybe in a subtle way we're trying to make control theory
palatable to people who still think in straight lines.

After all that, I guess I just should have said "Very interesting." But
what you're trying to do deserves more than that.

Best,

Bill P.

[_Martin_Taylor]

[Martin Taylor 920626 17:45] -- meetings finished for today!
(Bill Powers 920626.0830 to Mark Olson)

Bill wrote such a lucid (as usual) comment, the I feel a little bit
presumptuous trying to add to it. But I did think it would have been
helped by another little diagram. And I'd like to use this diagram later
in my follow-up on the statistics and experimental methods discussion.

                    To (from) higher levels
                   > >
                   > > Reference(s)
            --------\-------|-----------------
           > \ comparator |
           > / \ |
           > perceptual output | One of many ECSs at level n
           > combiner gain and |
           > function distribution |
           > / / | | \ \ / / | | \ \ |
            --/-/--|-|--\-\----/-/--|-|--\-\--
            many control systems at many levels
   Person |||||||||||||| ||||||||||||||
             ^^^^^^^^^^^^^^ vvvvvvvvvvvvvv Level 0 effects
   ========== many sensors ==== many effectors ==============
               \\\\\\\\\\ ////////// <--Physical observables
   World \\\\\\ ////// <--Low-level interpretation
                     \\ // <--more abstract interpretation
                      ^ V <--A "solid" testable interpretation
                    controlled complex
                  environmental variable
                    // \\ // || \\
                   ^^ ^^ ^^ ^^ ^^
                 All sorts of disturbances

Note: The upgoing (sensor-based) flow and the downgoing flow (to the
effectors) should be mingled, since reference signals tend to go down to
low-level ECSs whose perceptual functions provide input upwards in a
receprocal manner. But it is hard to draw that in ASCII.

···

-------------------
The ECS controls one variable. To it, the variable is simple, mirrored in the
single=valued result of its perceptual control function. But between that ECS
and the variable it controls are many paths, both down through the effectors
and back through the sensors. At each level down to the "world interface",
there are other ECSs, each of which perceives a variable that is very simple
from its own point of view. All these intermediate controlled variables and
the ECSs that control them are part of the cause AND effect of the variable
controlled by the ECS we are focussing on. All the percepts in the ECSs in the
control path mirror variables in the (excuse me) "cause-effect" circuit
relating to the controlled variable in focus. And all are subject to
disturbances that might or might not affect the focus variable.

In this diagram, it is hard to identify anything like causes, effects,
influences, or properties. The nearest thing to a cause is the fact that
changes in the reference at the top of the diagram cause mirroring changes in
the environmental variable, if the properties and disturbances of the world
permit it. But, if the reference stays stable, then the world disturbances
can be said to "cause" restoring behaviour at the level marked with a single
"V" marked "a solid testable interpretation" (which is both solid and testable,
and more abstract than those above it).

The world disturbances "influence" behaviour that restores variables controlled
by intermediate-level ECSs, since the reference levels that apply to those
intermediate ECSs vary with the corrective outputs of the ECS in focus; hence
there is no fully predictable effect of the disturbance on the intermediate-
level behaviours. Consider the recent VOT discussion. If the talker was
controlling for the perception that /p/ was being spoken, there is no need for
VOT to be controlled except for being in the /p/ range, and it can therefore
be partly controlled according to other references not determined in the
experiment.

If we look very close to the world interface, the disturbances become "stimuli"
and the controlled variables become "responses", but one can seldom talk about
cause and effect at this level, firstly because the relevant references are
unlikely to stay stable, and secondly because it is hard to observe either
all the "stimuli" that combine in the perceptual function of any ECS or all
the "responses" that might affect the variable that ECS controls.

The words "cause" "effect" and "influence" are very hard to give a precise
meaning in a dynamic world that contains feedback. My description differs
from Bill's in manner, but I think (hope) it carries the same import.

Martin