The Test for the Controlled Variable

RM: My understanding is that if you find that a mouse is controlling for receiving pellets at a particular rate

Why the emphasis on controlling?

RM: But before concluding that rate of pressing is the controlled variable in this situation you should try testing other possibilities. For example, you might test to see whether grams of food/ minute is a better definition of the controlled variable than pellets/minute. Grams of food/ minute might be what the mouse was actually controlling but you couldn’t tell because all the pellets were the same weight.

One operational definition of the CV can be better than another operational definition of the CV. But, because the CV is a part of the observer’s environment, The TCV can never find the controlled variable from the subject’s perspective. There is no finding the “actual CV”, mice don’t care or know about grams or pellets or what we call protein or whatever.

Using the TCV you can never find the controlled variable that the subject is controlling, only a correlate of the controlled variable; or a slightly better correlate.

The Test for the controlled variable is called the test for the controlled quantity in B:CP. I think that is a slightly better name. It emphasizes that the hypothesis of the CV is a part of observer’s environment (quantity).

RM: Once you have convinced yourself that the mouse is controlling

Technically, you should only convince yourself that you cannot think of a better approximation of the controlled variable.

RM: The perceptual signal in PCT is always an analog of (and, in that sense, is equivalent to) the controlled variable.

No. In PCT, the perceptual signal is the controlled variable. Sometimes, Bill calls the input quantity or the controlled quantity, or the CV - controlled variables, but also emphasizes that the perceptual signal is the one actually being controlled by the subject or by the control system, not the “objective” CV.

There is a naming conflict here. The CV in TCV is not actually finding the controlled variable, but just an approximation, a correlate.

RM: How that perceptual signal is experienced by the controller – man of mouse – is not something that PCT deals with. In the mouse, maybe it’s experienced as variations in the feeling of “satiety” or (more likely) “the feeling of getting full”. But no matter how it is experienced, in PCT, variations in perceptual signals are theoretical analogs of variations in controlled variables.

When we talk about our own controlled variables, then the experience is all we ever talk about, and all we ever can talk about. PCT very much does deal with experiences of men, and the hypothesis is that perceptual signals are our experience. It is a bit difficult to translate those terms to other people or organisms, so that is why the naming conflicts.

This nicely dices the question

 O where is that which is controlled 
 Or in the head or in the world? 

We control a perception of the sun rising in the east by orienting organs of perception so as to receive inputs from which our brains can construct that perception. If the light of the sun is too bright for comfort, we don’t somehow turn down the rate of fusion in the sun, we shade our eyes.

A cat is intently staring at a mousehole. There is no mouse there. What sounds like gnawing behind the baseboard is actually an old toilet-paper roll stirred back and forth by air from a leaky heating duct. The cat obviously cannot capture a mouse that is not there. Nonetheless, the cat watching the mousehole is controlling a perception of capturing a mouse, irrespective of whether a mouse is there to be captured. The capture is the culmination of sequences of controlled perceptions–whether the sequence of sequences constitutes a program is immaterial here. The reference value for that culmination is the motivation for the first step of it, the intent watch over the mousehole. The visual image of a mouse would provide a reference value for the next step, reducing the distance between paws and mouse.

This has come up before, for example in this 1998 exchange with Fred:

And in 2017 in our exchange about some of Vancouver’s work:

I agree with what you said in 2017, in which you agreed with Bill. I would throw away the verbal fog word “for”. To paraphrase Yoda, there is no “control for”, there is only control. (Tom Bourbon and I had a brief conversation about this unfortunate locution at the meeting in Vancouver.) And the word “implicit” only means that the observer is inferring control as an “implication” of other observations, something which is true even when control has been strongly confirmed by the Test, so throw that bit of verbal fluff away too. From the cat’s point of view, its control of the perception of a mouse is not implicit.

Your present assertion is that there can there be no control if the loop has perceptual input, inhibitory p synapsed with excitatory r yielding error signal e, and diverse means of behavioral output, but no effective environmental feedback path by which those outputs change p.

But we cannot say there is no control, only that control fails–in this case, due to outputs not addressing any effective environmental feedback path. Faced with less cosmic derangements, the usual response often starts with reorganizing in the perceived environment. Go get the needed tool, say. Faced with the appearance of the sunrise in the northwest, an obdurate emperor could have his servants arrange mirrors so that the current manifestation of sunrise appears to him through an east-facing window frame. But if the earth were actually to shift on its axis, failure to regain control by problem-solving strategies arranging an effective environmental feedback function would lead next to internal reorganization. This might result in establishing a new reference value for the location of sunrise (there are less sane possibilities). But if the location of sunrise is not controlled such reorganization will not occur. No control, no error. No error, no reorganization.

Bill’s examples are of sequences: the occurrence of perception A sets a reference for perceiving B. In 1968 I set a reference value of getting a PhD in Linguistics from Penn by means of writing a dissertation describing a language very different from English. I started work on the genocidally extinguished Yana language. In 1970, after giving a paper on this in La Jolla I was invited to work on a related language that was still spoken, and started work with the Achumawi language. In 1975 economic and (academic) political limitations presented overwhelming disturbances for a decade. In 1986 I resumed (while raising a family with unrelated work). In 1998 I turned in a partial description and received the PhD. I am still working on the description of the language. During that 52-year period control outputs were often not evident and elements needful for environmental feedback often were not available, but control has never ceased. It has involved the trial-and-error creation of many sequence perceptions, most of them adapted, diverted, or merged into another sequence after being only partially executed, many interrupted and taken up again, some abandoned.

It also explains how we can conclude that a variable V (which we perceive) is controlled by the subject. Having first verified that the subject can and does perceive V, we perceive that concurrent environmental influences d should ‘adjust’ the state of V but are opposed by the subject so as to maintain V at what we can conclude is its reference value. It is perfectly possible to control a perception with low gain in one loop and at the same time observe it with no control output in another loop. In the Test, we control a perception of the expected state of V under influence of d with the reference value provided from memory in the canonical way (“this is the way physics works”, or whatever). Concurrently, we observe a perception of the actual state of V under the combined influence of d and the subject’s observed outputs q.o. If we are ‘objective’ we are careful not to control this perception of the actual state of V, simply observe. A higher-level system compares the expected state (which we are controlling by means of d) with the observed state (which we are careful only to observe).

Hi Bruce

BN: Somewhere in the csgnet archive Bill wrote that yes indeed we control the sun rising over the eastern horizon, as witness our consternation should it rise anywhere else.

RM: Consternation is not really evidence of control…

RM: I think I found the place where you might have gotten the impression that Bill was saying that consternation is evidence of control:

BP: There’s a television ad showing a sunrise in speeded-up motion. I felt uncomfortable about it several times before realizing what was wrong. It simply shows the disk of the sun moving toward the upper left, away from the horizon. The problem is, I believe the sun always moves up and to the right away from the horizon (in my northern hemisphere). Seeing it move “wrongly” creates a definite sense of error, even though there’s nothing I can do to change it but not look.

RM: Bill gives this as a demonstration that the consternation felt when experience fails to match expectation is evidence that expectation is a reference signal. Since there is nothing that can be done about the error I guess the consternation could be seen as evidence that the perception would be controlled if there were something that the organism could do to control it.

RM: So displays of consternation can be taken as evidence that the system has a reference for some perception even if that perception can’t be controlled. I had an interesting experience of this when we took our kids to see a “mystery house” at Knott’s Berry Farm years ago. A mystery house is one which has been designed so that visual vertical deviates from gravitational vertical. So when you walk in it feels like gravity is pulling at about a 20 degree angle relative to straight down.

RM: My daughter Lise was a babe in arms at the time – about 2 months old – so she was being carried into the house. As soon as we entered the point where the illusion took hold – where gravity felt like it was off kilter – Lise started to cry hysterically. Clearly, the apparent reorientation of gravity was not what she “expected” and this created considerable consternation. I took it as evidence that humans (and probably other animals) have built in references for the perception of the pull of gravity being aligned with the perception of visual vertical. I think having a built in reference for this perception has obvious adaptive value for organisms that will have to learn to balance on four and then, possibly, two legs.

BN: This nicely dices the question

O where is that which is controlled
Or in the head or in the world?

RM: That question is already answered in PCT: both. What is controlled is a perceptual analog, in the head, of a variable aspect of the world. To paraphrase one of the great movie lines, from Night of the Living Dead: Control the perception and you control the world.

BN: Your present assertion is that there can be no control if the loop has perceptual input, inhibitory p synapsed with excitatory r yielding error signal e, and diverse means of behavioral output, but no effective environmental feedback path by which those outputs change p.

BN: But we cannot say there is no control, only that control fails–in this case,

RM: I see failed control as the same as no control.

BN:…In the Test, we control a perception of the expected state of V under influence of d with the reference value provided from memory in the canonical way (“this is the way physics works”, or whatever). Concurrently, we observe a perception of the actual state of V under the combined influence of d and the subject’s observed outputs q.o. If we are ‘objective’ we are careful not to control this perception of the actual state of V, simply observe. A higher-level system compares the expected state (which we are controlling by means of d) with the observed state (which we are careful only to observe).

RM: Teh only problem with that is that we are not controlling for the expected state of V when we do the Test. If we were we would simply be in a conflict with the testee, which could make him or her rather testy.

Best

Rick

I agree here with Rick. One cannot say that control failed if there were no possibilities even to try.

That is from the observer’s point of view, not from the controller’s point of view.

No, it wasn’t the post about the TV ad. Might have been something Bill said in person at a conference, or it might be that I couldn’t find it hidden under the … expansion widget at the bottom of a post in the search results.

Ah, I assumed you would get the reference.

Absence of observable output is not evidence of absence of control. Inadequacy of output (e.g. in a tug of war) is not evidence of absence of control. If your daughter’s distress was evidence for a reference for agreement between visual vertical and inertial vertical it was also her means of control ("Daddy, fix it!). I bet it worked.

There’s a value for “V as influenced by d” if V is not controlled. This value is provided by controlling principles of physics (or whatever) and operations of mathematics in combination with observed values of V and d, and by these means controlling a perception of the expected value. The means of controlling this perception of the expected value of V are not of the same kind as the subject’s means for controlling her perception of V. The controlled perception “expected state of V as influenced by d” is one input to a higher-level system which compares it with its other perceptual input, the “observed state V as influenced by d”.

Concurrently, the investigator is controlling a perception of d. This control of d does indeed conflict with the subject’s control of V. However, the investigator purposely constrains the magnitude and duration of d so that the subject easily overcomes the disturbance and thereby demonstrates the fact of control and its reference value. This constraint is presumably imposed by higher levels of control, perhaps at the same level as the system that is comparing expected and actual values of V. So far as I know, no one has modeled this purposeful limitation of a deliberately introduced conflict, though I have sketched a diagram:

The little green rectangle in the environment is V as hallucinated by both parties, that is, they experience their respective perceptions as realities in the environment. The perceptual input for the experimenter is not labeled p, it is labeled q.i because that is our name for the particular perception that the experimenter is controlling: a value which by inference corresponds to the value of the perceptual signal p inside the subject. That inference is supported (or not) by the experiment, the Test. The disturbance values shown in the environment are also perceptions in side the experimenter, but they are shown as the experimenter projects them to be in the environment. They are perceptions in control loops which are not shown: controlling a bit of conflict with the subject, experimentally, and managing extraneous variables in the experimental setup. Also not shown are a loop controlling the value predicted by physical sciences sans control, and a loop comparing the expected vs. actual value of V, as discussed above.

And thus, the experimenter is controlling a perceptual signal q.i, a perceptual analog in her head of a variable aspect of the world located inside the subject’s head, a perceptual signal p. The signal q.i is controlled by means of devices and procedures for measuring and recording influences on the subject’s sensory organs. Concurrently, in the other “real” world, the experimenter experiences a perception V as existing in the environment, and with it the perception that she has confirmed that it is the same as a perception V that the subject is experiencing as existing in the environment.

As good William Shakespeare put it, all of this, and we as well,

… are such stuff
As dreams are made on, and our little life
Is rounded with a sleep.

Yeah, I found that post too. Not the right one. It may be that in looking through the search results I haven’t opened all the collapsed text (by clicking the […] at the bottom), or it may have been a remark in person at a conference.

full-moon90

What are we looking at here? Is this an image sent back by Voyager in its travels past the moons of Jupiter and Saturn? The relevance will become evident in a moment.

In general, when control degrades or fails in a given loop because its output function has become ineffective, a living system attempts to regain control, either at that level, or by higher-level systems attempting to recruit other means of influencing and then controlling the desired perception, or by a higher-level system using concurrent input as an alternative, or other strategies that SFAIK have not been systematically investigated (the “Learning Working Group” have been silent), and if all this fails then by random reorganization.

Every non-intrinsic control loop began as an incomplete loop. Playing with outputs and perceiving consequent changes to input, and that controllable input becoming subject to a reference signal. Or a comparator receiving an input signal and reference signal and outputting an error signal that has no effect on the input–a situation similar to an output function becoming ineffective, with like remedies. As Bill observed, we’ve never had to develop an output function to control the sunrise. It always just happens as expected. (Ignoring exceptions like the great London fog, Krakatoa, travel beyond the arctic or antarctic circle, etc.)

“We know” that the hypothetical subject can never develop an output function that moves the location of sunrise, but the biological systems most directly involved do not know this, and it is they which through the somatic branch of the hierarchy cause the feelings which give rise to emotions of consternation and associated activities in the behavioral branch of the hierarchy. When eventually systems controlling principles of physics regain control (if they do), there will still be a desire for things to look familiar, and somatic feelings when they don’t.

The image above will look familiar to John Kirkland for the same reason that the TV commercial with the sun rising on a slant toward the left will look normal to him. My daughter reported that on a trip to Australia it was still a bit unsettling even when she had the rational explanation. Here’s that satellite as we see it at our latitude, rotated about 90º from John’s point of view.

full-moon

There, now–doesn’t that feel better?

Hi Bruce

RM: That question is already answered in PCT: both.

BN: Ah, I assumed you would get the reference.

RM: Still don’t.

RM: I see failed control as the same as no control.

BN: Absence of observable output is not evidence of absence of control.

RM: It’s absence of a controlled variable that is evidence of absence of control.

BN: There’s a value for “V as influenced by d” if V is not controlled…

BN: Concurrently, the investigator is controlling a perception of d. This control of d does indeed conflict with the subject’s control of V. …So far as I know, no one has modeled this purposeful limitation of a deliberately introduced conflict, though I have sketched a diagram:

RM: This diagram is not a correct description of what the observer/experimenter does in the Test. Think about it in terms of my mind-reading demo (Mindreading). In that demo the computer is the observer/experimenter simultaneously testing to determine which of three avatars is being moved intentionally around the screen. The computer is not controlling for anything; it is simply measuring the average correlation between the disturbance to each avatar and the position of that avatar. It is testing for the controlled variable without controlling the hypothetical controlled variables.

RM: What is controlled is a perceptual analog, in the head, of a variable aspect of the world.

BN: And thus, the experimenter is controlling a perceptual signal q.i, a perceptual analog in her head of a variable aspect of the world located inside the subject’s head, a perceptual signal p.

RM: The observer/experimenter is observing the correlation between disturbance and hypothetical controlled variable. If the correlation is low then the observer/experimenter is observing – not controlling – the controlled variable, which is a perceptual variable in both observer and controller.

BN: As good William Shakespeare put it, all of this, and we as well,
… are such stuff
As dreams are made on, and our little life
Is rounded with a sleep.

RM: Let’s hope so;-)

Best

Rick

Thatʽs an empty statement. The perceptual variable is the sun rising over the eastern horizon. The presence of the variable is not in question. What is in question is whether or not that variable is controlled. So what you are saying is “absence of control is evidence of absence of control.”

What you said before is more plausible; if the system has no output function available that can affect the variable so as to maintain it near the reference value, it cannot control the variable. But there are many situations where a variable is controlled, but no effective output function is presently available. As means of control, the system might recruit, free up, find, learn to use, or develop an effective output function. “Sorry, can’t help you until I put this package down.” Many a resentment has been sustained for years, a reference value for control manqué–control that becomes manifest when effective means become available. “As soon as I put a new point on this arrow and climb that mountain path after him, that Ötzi is a dead man!” Were the poles of the earth to shift, survivors looking toward the formerly eastern landmarks would remember the old days with the visceral intensity of frustrated intentions associated with the sun rising and getting on with another routine day. To distinguish deferred gratification from those cases where there can never possibly be an effective output function is not an easy matter.

This is a novel definition of the Test. The standard definition is quoted for reference in the first post at the start of this topic. Scroll up to the top, or go to p. 77 of Phil Runkel’s big book.

Maybe you’ve got a shortcut here, a simplified methodology. The computer program generates a set of identified variables specifying the locations of pixels which we perceive as the avatars, applies the same disturbance to all of them, calculates the correlation between location and disturbance for each, and identifies the one for which the correlation is low. I don’t think this can be expressed as an experimental procedure for the Test. For example, in the coin game I apply the same disturbance to all possible arrangements of the coins at once and one of them stands out because of its low correlation to the disturbance?

The claim that your demo is demonstrating the Test asserts that the program is an analog of the experimenter conducting the Test. Its program outputs specify the locations of the avatars. If the program were structured as a control system (as the experimenter is), those specifications would be its reference values for variables controlled by the experimenter, the locations of the avatars, and the user input would be a disturbance to one of those variables. The program already ‘knows’ that the user-controlled variable is the x-y location of one of the avatars, so the experimental situation must be likewise constrained. The program identifies the avatar whose location differs from that specified. It does this by comparing the program-specified location with the actual location and calculating their correlation. What is the analogous Test operation for the experimenter? Does the experimenter actually measure and calculate the correlation for each avatar, or is the term ‘correlation’ a bit distracting?

Let’s take the analogy of program to experimenter at face value. The experimenter employs this methodological shortcut in a particularly constrained experimental setup. The experimenter is controlling a number of replications of an identified variable (or instances of the same ‘kind’ of variable) simultaneously, and identifies which one the subject is controlling by observing which one is disturbed from its intended value. As the experimenter, I don’t need to calculate any correlations because I am a control system controlling each of those variables, and there’s error output causing control output in respect to just one of them.

OK, but the program is not resisting the user’s control because the program is not controlling the locations of the avatars, it is only determining them, and that is why there is no conflict. The program presumably has an analog of perceptual input for the actual current locations, but instead of subtracting that from the specified location and transforming the delta into output that changes the actual location it calculates the correlation between the actual location and the user’s x-y specification for the location, which you call the disturbance as though the computer were a control system controlling the location.

In the experimental analog, the experimenter is in fact a control system, the locations of all the avatars are controlled by the experimenter, and the user’s control of the location of one avatar is a disturbance to the experimenter’s control of it because the user and the experimenter are in conflict over the value of that variable. In the demo, there is no conflict because the program is not structured as a control system.

The computer program does not implement a control system acting in a way that is analogous to an experimenter conducting the Test, so the demo is not a demonstration of the Test.

The term disturbance is often equivocated in a less obvious way. The disturbance is an environmental influence upon the variable under consideration, and a measure of the effect of that influence is properly called the disturbance quantity. Equivocation using the word “disturbance” both for the environmental influence and for its effect probably arose in the description and discussion of computer programs which do not simulate the physical properties of the environment. In early tracking experiments, the disturbance quantities are not measurements or specifications of any disturbing influence in the environment. They are specifications of changes in x-y location values, quantities that would be transforms of physical pressures upon another joystick (or mouse) affecting the cursor, perhaps, if such a disturbing influence existed in the environment. But because there is no influence in the environment to be represented by the disturbance quantity–indeed, there is no simulation of the physics of the environment at all in these programs–there can be no distinction between the disturbance and the disturbance quantity. They are conflated. This is the case also in the demo in which the program identifies which avatar has its location controlled by the user. The demo has no simulation of the physics of the environment and cannot distinguish between the disturbance and the disturbance quantity. For this reason, too, it is not an analog of an experimenter conducting the Test.

In our diagrams, the disturbance d is represented as an influence that affects the value of the controlled input variable. If the objects whose locations are measured were billiard balls or shuffleboard discs then d would be a measurement of a physical influence that changes their locations. Computer code just knows to paint pixels in specified locations. It knows nothing of forces that move painted pixels from one location to another. When you’re just counting pixels it’s all location and there is no choice, the word disturbance can only refer to the effect of the disturbance upon the value of the variable. This goes unnoticed when there is no physical environmental feedback function. If your demo included code to simulate physical relationships and influences, then the calculation of a correlation of d to V would be rather different.

The reference was to the place where I had already said “it’s both”. Since you didn’t get the reference I gave you a link to it, and in addition I quoted it. I don’t have an effective output function at present, but it’s because the effect of the output depends upon your perceptual input functions, a part of the environment over which I have no control.

Hi Bruce

A friend of mine reminded me of this nice quote from Winston Churchill: “You will never reach your destination if you stop and throw stones at every dog that barks”. There are a lot of dogs barking in these posts of yours – most of them the same old mutts that have been barking at me since I got into PCT. So I will heed Winston’s advice and throw stones only at a couple of them and then continue on to my destination.

RM: It’s absence of a controlled variable that is evidence of absence of control.

BN: Thatʽs an empty statement.

RM: To you, perhaps. To me the statement is filled with the essence of perceptual control theory.

BN: But there are many situations where a variable is controlled, but no effective output function is presently available.

RM: I think it is interesting that organisms do apparently come into the world with some built in references for the state of perceptual variables – such as a reference for alignment of the direction of gravitational pull with visual vertical. But I would prefer to call that a potentially controlled variable and reserve the term “control” for describing behavior that actually involves control – maintaining variables in constant or variable reference states.

RM: In that demo the computer is the observer/experimenter simultaneously testing to determine which of three avatars is being moved intentionally around the screen. The computer is not controlling … anything; it is simply measuring the average correlation between the disturbance to each avatar and the position of that avatar. It is testing for the controlled variable without controlling the hypothetical controlled variables.

BN: This is a novel definition of the Test.

RM: There are several ways to do the Test. But they are all the same inasmuch as they are ways of determining the same thing: the best definition of the variable(s) around which the observed behavior is organized.

BN: The claim that your demo is demonstrating the Test asserts that the program is an analog of the experimenter conducting the Test. Its program outputs specify the locations of the avatars.

RM: The subject specifies the location of one of the avatars; the computer determines which avatar that is using the Test: looking for lack of expected effect of a disturbance to the variable.

BN: As the experimenter, I don’t need to calculate any correlations because I am a control system controlling each of those variables, and there’s error output causing control output in respect to just one of them.

RM: A human experimenter could not do (unassisted) what the computer experimenter is doing because a human cannot apply three different time varying disturbances simultaneously to the positions of three different objects and at the same time determine which of these disturbances is being resisted – and therefore which object’s position is being controlled.

BN: The computer program does not implement a control system acting in a way that is analogous to an experimenter conducting the Test, so the demo is not a demonstration of the Test.

RM: False. But I think you have to believe that this is true for some reason so I’m not going to take the time to try to explain why it’s false.

BN: The demo has no simulation of the physics of the environment and cannot distinguish between the disturbance and the disturbance quantity. For this reason, too, it is not an analog of an experimenter conducting the Test.

RM: This is also false. If you don’t think my Mind Reading program (Mindreading) is a good example of the Test then you are living in a completely different PCT world than I am.

Best

Rick

This topic begins with a description of how to test for the controlled variable, as specified by Bill Powers and Phil Runkel. The purpose of recapitulating it there is to provide a convenient reference for people interested in doing PCT research, especially in naturalistic research (Phil’s particular interest).

Your demo shows that the program can identify which avatar the user is moving. The user has no direct insight into what the computer is doing—it could be just reading off user input. But as the user assumes good faith, this is a demonstration that it is possible to identify which of several like variables is controlled. This might encourage further investigation and learning, but it does not show or teach how to do it. It is not analogous to a human experimenter for several reasons, one of which you stated:

The rest speaks for itself. I scarcely recognize what I wrote, snippets ripped out, forced into dog collars, dragged from their contexts, and leashed to rhetorical posts to have stones thrown at them. Readers should refer to the first post of this topic to understand how to apply the Test in their research.

Here is a diagram intended to help think through what the Experimenter and the Subject are controlling when the Experimenter has succeeded in identifying the perceptual variable that the Subject is controlling. Control is represented as a single loop, omitting the complexity of lower levels (unless the CV is an Intensity perception). Higher levels setting the reference Rs in the subject and Re in the Experimenter are also not shown but can readily be imagined. Systems setting Re are especially essential for understanding the TCV how the Experimenter introduces and varies the disturbance D.

Elsewhere, I have said that the disturbance (usually represented by d, but D in this diagram) is a perceptual variable controlled by the Experimenter. This is false with respect to this TCV loop, because the output quantity of a control loop, Q.o, is not what is controlled in that loop. But the Experimenter has other control loops (not shown), among them those which measure and record the experimental variables, and in respect to those systems the Experimenter’s Qo=D is a controlled perceptual variable.

There is no contradiction here. What is measured as Qo is not the aggregate of behavioral output, but only that aspect which affects the state of the controlled input Q.i. Other effects, if any, are extraneous side effects (not shown).

The Environment function may be thought of as a quantitative representation of the relevant (affected and perceived) aspects of whatever is Really Real in the observed time and space of the experiment. For experiments with motor control, laws of physics are assumed, but it is a radical simplification to reduce those mathematical formulae (F=MA, etc.) to a couple of ad hoc constants Kd, Kf, the products KdD and KfQo, and their sum KdD + KfQo.

For higher-level perceptual variables such quantification becomes quite indirectly related to the physical effects of behavioral outputs, and the Environment function represents each person’s ‘projection’ of perceptions as though they were realities in the environment. In the Roberts et al. experiments with self image, words like “You’re a liar!” certainly do have quantifiable physical attributes, but varying those attributes slightly does not make them less or more a disturbance. There are problems of experimental design here which have not been addressed.

Hi Bruce

BN: This topic begins with a description of how to test for the controlled variable, as specified by Bill Powers and Phil Runkel. The purpose of recapitulating it there is to provide a convenient reference for people interested in doing PCT research, especially in naturalistic research (Phil’s particular interest).

RM: There are also excellent examples of how to do the Test in Powers (1971),
Powers (1978), and Marken (2013).

BN: Your demo shows that the program can identify which avatar the user is moving.

RM: Yes, it does that using the Test to determine which avatar is being moved intentionally.

BN: The user has no direct insight into what the computer is doing—it could be just reading off user input.

RM: Any user can learn how the Mind Reading program works. But whether or not they know how it works, it works – it successfully reads their mind, determining which avatar is currently being moved around on purpose – as long as the user is controlling the intentionally moved avatar.

BN: But as the user assumes good faith, this is a demonstration that it is possible to identify which of several like variables is controlled.

RM: Actually, the quality of the user’s faith doesn’t affect the possibility of identifying which variable (position of an avatar) is being controlled. The Test that is carried out in the demo works as long as the user is controlling the position of one the avatars. It’s all about detection of control.

Best

Rick

RM: Thanks for all this information about what is involved in doing the Test. But I do find your discussion a bit abstract. Since this level of understanding must be based on a considerable amount of experience with Testing for Controlled Variables I wonder if you could give some nice, concrete examples of how you did the Test.

Merry Xmas

Rick

No, not observer’s point of view. Rather it is analyst/theorist’s point of view.

An observer, say an anthropologist, could say that ancient Egyptians were controlling for sun rise from the east every future morning by building a pyramid as a tomb for their pharaoh so that he could help the sun god to fight against the powers of darkness during the night. Or that the Aztecs were controlling for sun rise every morning by offering human sacrifices to the gods. That is a point of view of an attentive observer. Instead, a more scientifically minded analyst or theorist would say that they just – deeply but erroneously – think and believe that they are controlling for the sun rise, but in truth, they cannot be controlling for the sun rise because the sun rise is not something that humans could be controlling.

I agree that someone can want and try to control for sun rise and that someone can also think and believe that she is controlling, but that does not mean that they really control.

If the environmental feedback path is missing, then the loop is not closed. and it cannot be a case of perceptual control.

The full control loop must contain both p-control (of the perceptual variable) and e-control (of the corresponding environmental

variable), I think.

Every CS block diagram is abstract in the sense that you are invoking.

It’s OK, we all get forgetful.

The Test is typically performed in linguistics by repeating an experimental utterance while making a substitution for some identified part of it. An early exposition:

  • http://www.iapct.org/festschrift/nevin.html
    A more recent exposition begins on p. 378 of my chapter in LCS IV.
    On pp. 383-386 of that chapter is a diagram for a PCT account of experimental work done by Katseff et al. in the phonology lab at UC Berkeley. I did not perform the test in that instance, not having the specialized hardware on loan from the Otolarangeology Department at USF as they did, but I did propose it on CSGnet in 1991-1992.

I have provided you data on populations of people controlling self-concept by differentiating their pronunciations of words from the way “those other people” pronounce them. While such interactions are not instances of the Test, they provide naturalistic observational data in which consistent ‘correction’ of disturbing pronunciations demonstrates control of certain thereby identified variables of speech (which may be reported quantitatively in terms of the center frequencies of vowel formants and are experienced subjectively as different ‘vowel qualities’). Bill Labov provided references for much more such data, and the associated literature is quite large.

RM: This is an edited version of the reply I sent earlier:

BN: The Test is typically performed in linguistics by repeating an experimental utterance while making a substitution for some identified part of it. An early exposition:

BN:A more recent exposition begins on p. 378 of my chapter in LCS IV.

RM: I would appreciate it if you would explain how you carried out these examples of testing for controlled variables.

BN: On pp. 383-386 of that chapter is a diagram for a PCT account of experimental work done by Katseff et al. in the phonology lab at UC Berkeley. I did not perform the test in that instance, not having the specialized hardware on loan from the Otolaryngology Department at USF as they did, but I did propose it on CSGnet in 1991-1992.

RM: The Katseff et al study is very close to being a good example of a test for the controlled variable. What’s missing is any effort to test an alternative definition of the controlled variable. The fact that the subject’s output didn’t return the formant to the undisturbed value suggests that the formant frequency is not the controlled variable. If the vowel sounded ok nevertheless then that’s evidence that they were controlling for the desired perception ok, it just wasn’t the absolute formant frequency that they were controlling for.

RM: But Katseff et al’s method of doing the Test contradicts your model of how the Test is done. Your model has E in conflict with S over the state of the hypothetical controlled variable. In the Katseff et al study there was no conflict; the computer (E) did not increase its disturbance to the formant when S varied her output to compensate for it. The disturbance (which was a change in the feedback connection between S’s output and input) was applied completely passively.

BN: I have provided you data on populations of people controlling self-concept by differentiating their pronunciations of words from the way “those other people” pronounce them.

RM: That was a great study but I didn’t see it as a Test of “self concept”. Indeed, I didn’t see anything in that paper that looked like the Test. Rather, I saw some excellent data collection. I created a simple model that predicted this data by assuming that individuals are controlling for imitating the pronunciation of those with whom they interact. So the model implicitly tests to see if one controlled variable in this situation is pronouncing diphthongs the way others pronounce them. The model worked pretty well so it is evidence that people do control for this variable. This was a model-based version of the Test for the Controlled Variable. Inasmuch as this was the case, my approach to testing for the controlled variable is nothing like the way you say it must be done – with E and S in conflict. There really should be no conflict between E and S in any properly done version of the TCV.

Best regards

Rick

Discussion of linguistics and the variables that constitute languages is off topic here. You can reply to the topic I have placed in the Science/Language subcategory. You might want to reconsider your conclusion that the Katseff et al. work shows that formants are not controlled.

What’s relevant to the present topic is that while testing for controlled variables one must be alert to the possibility of hidden or non-obvious conflict as the reason for the appearance that a variable is not controlled. As Katseff & Houde put it (Lab. Phon. 11):

These results suggest that both acoustic and sensorimotor feedback are part of one’s lexical expectation. Because auditory feedback is altered while motor feedback is not, feedback from these two sources can conflict. For small shifts in auditory feedback, the amount of potential conflict is small and the normal motor feedback does not affect compensation. But for large shifts in auditory feedback, the amount of conflict is large. Abnormal acoustic feedback pushes the articulatory system to compensate, and normal motor feedback pushes the articulatory system to remain in its current configuration, damping the compensatory response.

On another note, I agree that by varying the disturbance we can demonstrate that control outputs are quantitatively equal but of opposite sign (or vector). Neither Bill nor Phil mentions this as essential to the definition of the Test, but it is important. You are incorrect to say, however, that Katseff et al. did not vary the disturbance. After drawing a parallel to work on reaching with and without prismatic lenses, they say (op cit.):

In speech adaptation, subjects wear a headset. They speak into the microphone and hear their speech played back to them through the earphones. The auditory version of the task used here involves four stages: baseline, ramp, plateau, and adaptation. Subjects repeat a single word, in this case, ‘head’, over a large number of trials. In “reaching” sensorimotor adaptation experiments, subjects initially see the object on the table in its true position. In speech adaptation experiments, subjects hear their voices unaltered during the baseline stage. During each trial in the ramp stage, auditory feedback is altered a small amount until it reaches a maximum value. Feedback alteration is held at that maximum value during the plateau stage. In this experiment, there are 5 sets of ramps and plateaus, after which feedback drops suddenly back to normal for the adaptation stage.

… subjects generally change their speech to oppose the auditory feedback change. For example, when F1 in auditory feedback is raised, making their /ɛ/ sound more like an /a/, subjects compensate by speaking with a lower F1; the vowels they produce sound more like /ı/. Similar experiments show that subjects will compensate for alterations in F0, F1, and F2 feedback, indicating that all three of these formants are important to a speaker’s representation of the target utterance… .

(I don’t know why the publication has /a/ as in “haha” where it should have /æ/ as in “had”. To a phonetician or phonologist it’s an obvious typographical error.)

RM: OK, I’ll go over there when I get the chance. And I didn’t say that the Katseff et al. work shows that formants are not controlled. I said it “suggests that formant frequency is not the controlled variable”. Other possibilities include the value of the disturbed formant frequency relative to the frequencies of the other relevant formants.

BN: What’s relevant to the present topic is that while testing for controlled variables one must be alert to the possibility of hidden or non-obvious conflict as the reason for the appearance that a variable is not controlled.

RM: I think the idea that the observed failure to compensate for the disturbance results from a conflict between auditory and motor feedback simply reflects Katseff et al’s lack of understanding of how hierarchical control works. There can be no conflict involving different types of variables…

BN: On another note, I agree that by varying the disturbance we can demonstrate that control outputs are quantitatively equal but of opposite sign (or vector). Neither Bill nor Phil mentions this as essential to the definition of the Test, but it is important.

RM: They don’t mention it because the only variables you really need to monitor in the Test are the disturbance and controlled variable. The output variable is often difficult to specify let alone measure.

BN: You are incorrect to say, however, that Katseff et al. did not vary the disturbance.

RM: No, I said that the disturbance in their study was a change in the feedback connection between S’s output and input. This was a fineway to introduce a disturbance to the hypothetical controlled variable (formant frequency). I was just noting that it wasn’t equivalent to what is called the “disturbance variable”, d, in PCT. In PCT, d is a variable that has an independent effect on the hypothetical controlled variable – “independent” in the sense that the controller has nothing to do with its value. So a PCT type disturbance in the Katseff et al study might have been an externally generated noise waveform added to the sound produced by the speaker.

RM: In the Katseff et al study the disturbance was a kind of “smart filter” inserted between speaker output (what was said) and input (what was heard). Again, there was nothing wrong with this approach to disturbing the hypothetical controlled variable. Indeed, it was a very clever study and I should have put it in my methods book. But the problem was that the results were not clear because the speaker was no able to completely resist the disturbance (the change in the feedback function); the speaker was apparently never able to vary their output in a way that allowed them (and the tester) to hear the word they were meant to say. This could have been because the disturbance was too great or because there was some other variable being controlled. But the study was certainly a good start at the Test.

Best

Rick

The “it” that they don’t mention is (continuously) varying the disturbance. (Step 3 in Phil’s statement is “Apply various amounts and directions of disturbance directly to the variable.”) I was imagining what perception you were controlling by your emphasis on the importance of varying the disturbance when you said the following:

I was imagining that the reason you emphasized a need to vary the disturbance was to gather data for a quantitative demonstration (output opposed to disturbance). But of course, as I see now, you were inveighing against my statement in other posts that, in the Test, the experimenter is in conflict with the subject. You were saying that if the computer were in conflict with the subject it would increase the disturbance diametrically to the subject’s output so as to maintain the CV (formant frequency) at the computer’s reference value.

I agree that the computer is not a control system in conflict with the subject. The computer altered the formant frequency produced by the subject in the course of the speech signal passing from microphone input to headphone output. Your point seems to be that this was a disturbance in the manner that a crosswind is a disturbance to the driver of a car. The wind is not a control system controlling the relationship of the car to the road, so the disturbance is not a conflict. This is true, unless there is some control system using some extension of its control outputs to determine the influence of the wind on the car. There is a control system using the computer as an extension of her motor control outputs to determine the influence of the disturbance on the formant frequency heard by the subject.

Much as the car is not a control system but its steering linkage, etc. extend the motor control outputs of the driver, who is a living control system, the computer extends the control output of the experimenter. By means of the computer, the experimenter controls disturbances to the relationships of formants as heard by the subject. The computer is not a control system, but the experimenter employing the computer as an instrumental extension of her motor output functions is a control system. By using the computer to introduce measured disturbances the experimenter is in conflict with the subject, a conflict which is tempered by higher level systems which are controlling perceptions of testing to identify and verify controlled variables.

I believe that your unstated assumption is that when two control systems have conflicting reference values for the same variable, the output of each is a disturbance to the other, and each increases its output until one or both reach maximum output capacity. As discussed many times elsewhere, this is true unless, as in a living control system, a higher-level control loop intervenes in some way. In the Test, higher levels of control constrain the extent to which the experimenter’s output d conflicts with the subject’s control because at those higher levels the experimenter is using the disturbance as means to verify what variable the subject is controlling, and an explicit condition of the Test is that the disturbance should be “gentle” and should not overwhelm the subject’s control. In a prior post, earlier in this topic, I have presented a partial block diagram that includes both subject and experimenter to illustrate this interdependency. Here it is again:

Bingo. But to control a relationship R between A and B you must control A as an input to R and you must control B as an input to R. Formants are controlled as means of controlling the relationship between formants which is heard as a given vowel. (There are also temporal differences, see the discussion on pp. 14-15 of Katseff’s dissertation. Page references in this post are to that work.) Note that subjects met disturbances to one formant by changing more than one formant (p. 47).

The relationship is called phonemic contrast. Consider an array of relationship comparators, one for each of the phonemically distinct vowels that one must control in order to speak and understand English. Because of differences in vocal tracts (especially length), dialect, social register (e.g. formality), stressed and unstressed syllables and phrases, and other kinds of variation, there is overlap. To visualize this, consult figs. 2.4 and 2.5 in Katseff’s dissertation. Presumably by innervation from different parts of the cochlea, current input about frequency concentrations during the highest-amplitude parts of utterances (the vowels) goes to comparators for vowels which are adjacent in acoustic space, as exemplified by hid, head, and had, and also hayed, heard, and the hud of “Hudson”. Control of contrast in heard speech is relative to other vowels heard from the same speaker. The experimental situation is simpler because the subject knows what her own speech sounds like and is accustomed to controlling it, but the general case of social variation is consistent with numerous demonstrations that the subject perceives “the same” vowels over considerable ranges of variation, with the boundaries intersecting and discrimination very context dependent. If that seems too abstract to make sense, hang on, it will get more specific shortly.

Katseff (pp. 19f) reviews some of the numerous studies of what speakers do when disturbances of various kinds are introduced, demonstrating that speakers control both auditory and ‘somatosensory’ perceptual inputs concurrently. (One must disregard disturbances to our PCT sensibilities when the authors of these studies use prevalent vocabulary such as ‘plan’, 'reaction, and ‘compensation’, reflecting ignorance of principles of control.) An example:

When a paddle is briefly applied to a speaker’s upper lip during the production of a sequence like [aba], the most straightforward response would be to increase the force pulling the lower lip up, opposing the force from the lip paddle. Speakers do not do this. Instead, they lower their upper lip, maintaining the bilabial closure [references elided]. There are two ways to account for compensation by the upper lip. One explanation is that subjects are trying to maintain the acoustics of the intended syllable, and the best way to make the /b/ in /aba/ sound like a stop is to lower the upper lip. Another explanation is that subjects have somatosensory expectations for intended syllables; that is, they know what those syllables are supposed to feel like. In the case of /aba/, they know that their jaws should feel open during the vowels and that their lips should come in contact during the /b/. Maintaining somatosensory information about this syllable would likewise require compensation for the lip paddle in /aba/.

In some experiments the shape of the oral cavity was deformed by a kind of bladder against the palate which in random trials was either flat or inflated, with changes smooth or abrupt, with or without masking of auditory feedback.

The fact that the reaction times were very short, and that substantial compensation occurred with masked feedback, suggests that somatosensory feedback initiates initial adjustments to articulation in the presence of the inflatable palate. That intelligibility improves after the first token suggests that auditory feedback is also used to adjust articulatory plans.

The references for tactile and kinesthetic perceptions in the oral cavity are adjusted so as to improve the sound that is produced.

The categorial nature of phonemic contrast is evident:

For small feedback shifts, the light gray shape (formants heard as a result of the feedback shift) falls entirely within the dashed shape (the subject’s baseline range), indicating that the vowels that the subject heard were all within his baseline region and that compensation was complete. As the amount of feedback shift increases (the dark solid shape), compensation is less and less complete. (p. 54)
[…]
for large feedback shifts, the re-synthesized vowel fell outside of the subject’s natural vowel region. There is some evidence that different brain regions are recruited to deal with feedback perturbations that fall outside of the intended vowel region. (p. 57)

However,

Because the relationship between articulation and acoustics is nonlinear, subjects may be compensating more or less steadily than they appear to be from their vowel formants alone.

You objected that subjects didn’t hear the disturbance.

But the vowel didn’t sound OK to the subject, as evidenced by their resistance to the disturbance. I suppose you’re getting this idea from e.g. p. 43:

Post-session interviews indicated that subjects did not notice either formant shifts or delays.

As you know, awareness is not necessary for control, and control outputs reduce the ‘noticeability’ of the effects of disturbances.

Your point that subjects were not controlling the absolute frequency of the disturbed formant (assuming that’s your meaning) is better illustrated when disturbance to one formant was resisted by changing two formants so as to approximate the target relationship of formants. But we already know that speakers do not control absolute formant frequencies, because we know they (we) control phonemic contrasts when the formants are at very different frequencies, e.g. shifted higher or lower by variable length of the vocal tract, for regional and social-class dialects, and so on. Speaking with helium in the lungs is a fun example.

By “different types of variables” I assume you mean different levels. What is involved here is different sensory modalities providing input to control of phonemic contrast. The auditory perceptions and the tactile/kinesthetic perceptions are at the same level, they are just in different sensory modalities. Phonemic contrast keeps words and syllables that are partially alike distinct from each other so that the hearer knows which the speaker intends. (If other context makes that obvious, gain on pronunciation is lower and auditory ambiguity results.) Perceptual inputs for controlling phonemic contrast are in two sensory modalities: kinesthetic/tactile perceptions involved in articulation of speech, auditory perceptions of the sounds thereby produced. We are concerned here only with the first two. (The McGurk effect and other evidence shows that visual input from the speaker’s face is also important, but that is not a variable in consideration here.)

As noted above the range of inputs that are perceived as one vowel overlap the range of inputs that are perceived as an adjacent vowel. This is so both for the sounds and for the ‘feel’ of articulation. In the diagram repeated below, auditory sound is labeled S and tactile/kinesthetic articulation input ls labeled A. It should be obvious that we our ability to control the sounds of speech in real time is very limited, because once a syllable has become audible it is impossible not to have already produced it. We correct a mispronunciation by saying the intended utterance again. The tactile and kinesthetic perceptions are controlled in real time. The former are controlled by setting references for the latter over time.

To resist disturbance to the sound, the subject changes references for articulation. When disturbance is relatively small, resistance to it (“compensation”) is complete. The articulatory perceptions move farther into the intersection of two ranges but are still close enough to ‘canonical’ values so that the intended vowel is perceived.

There are three adjacent vowels, call them V1, V2, and V3. The subject’s intended vowel is V2 and V1 is the target toward which the disturbance is moving the sound perception S. As the disturbance is increased, so that the sound is more like V1, and the articulation is changed so as to maintain the perception of the sound of V2, at some point the articulatory perceptions A are farther out of range to be perceived as the intended vowel V2 and begin to be perceived as the opposite adjacent vowel V3. (Compare the mathematical notion of cusp catastrophe which non-control theorists invoke to explain shifts in categorial perception.) Perceiving V3 increases error output in the system which has a reference (from the word-perception system at the top of the diagram) for perceiving V2.

Control of articulatory perceptions A is the means of controlling the sound perceptions S. Each accepts input that departs from the canonical value for V1 into the intersection of its range with the range of values for V2. What is measured is the formant frequency which is not completely “compensated”. What is not measured are the positions of the articulators which also are not completely “compensated” to where they would be for producing the sound V1, absent the disturbance. This is one of the ways in which the experimental data are incomplete.

Articulatory perception of the ‘feel’ of pronouncing V3 creates error in control of the intended perception of the sound of V2. By means of the disturbance, the experimenter is in conflict with the subject’s control of the word which contains V2. When the disturbance is small, the subject maintains good control. As the disturbance increases, the subject’s control is compromised in each sensory modality, sound and ‘feel’. So long as the perception S and the perception A are both within their respective ranges (as one modality extends into the intersection with V3 and the other into the intersection with V1) the subject does not notice.